Interpersonal Trust vs. Structural Assurance

Fundamentally, retraining in America fails to reflect the American model of trust, and this is where the gap begins. American trust operates as an interpersonal belief that exists between people, which means that trust anchors in an individual. Because the United States is a highly mobile, at-will labor market, assessing interpersonal trustworthiness is mandatory for economic survival. Private entities offering retraining fail to capture the trust of the American worker because these programs lack long-term mentorship and interpersonal relationships. Without a specific individual to anchor trust in, the system fails to convince workers that certification merits the tradeoff of time, effort, and missed opportunities. It also fails to capture the American model of trust.

Crucially, these programs also lack embeddedness within key institutions to extend relationships beyond an employer’s short-term needs. Often the curriculum of soft and vocational skills address near-term needs, but the employer misses out, too. Once their project needs are met, the employer’s absorptive capacity, or the ability to recognize and assimilate new external knowledge, apply it, and experience productivity gains, stagnates. The same structure the employee needs for a long-term vision is also the structure the employer needs to absorb, assimilate, and apply new knowledge. Without a structure to mitigate the friction of social uncertainty, a firm cannot effectively recognize, assimilate, or apply a retrained worker’s new capabilities. A long-term vision or “horizon” is missing for both parties.

Trust operates differently in countries where the social norm is “lifetime employment”. This model reflects a mutual, long-term commitment by both the employee and the employer, with the goal of continual benefit. Companies operating under lifetime employment models typically recruit after graduation and invest in the worker, taking observable actions that “assure” the employee of their commitment. This deepens embeddedness. While several European countries, most notably Italy, France, Spain, Germany, and the Netherlands, protect continuing employment through strict dismissal law, Japan achieves the same stability through norm and structure rather than through law.

Within a “lifetime employment” context, interpersonal trust isn’t needed because the system itself creates an “assurance structure.” Evaluating interpersonal beliefs about who to trust becomes irrelevant to career success. For example, in Japan, lifetime commitment is the expectation and cultural norm, sustained through mutual benefits, a shared long-term vision, and the mutual monitoring of behavior. This structural alignment makes it unusual for a person to leave a firm after it has invested in them. Thus, a Japanese employee leaving an institution violates this baseline assurance and social norm, and may find it difficult to re-enter the labor market.

The Credential Crisis: Reskilling in a “Market for Lemons”

Consider the training certificate from the perspective of the American worker: it represents a promise devoid of personal backing. Neither the worker nor the employer has a trusted, known entity vouching for its validity. Lynne Zucker’s (1986) historical work shows that America built this machinery deliberately: between 1840 and 1920, facing immigration and constant mobility, America replaced relationship-based trust with institutional substitutes, credentials, licenses, and certifications. However, these substitutes only function when supported by an infrastructure of guarantors, much like a medical board capable of revoking a practitioner’s license. Modern retraining certificates entirely lack this structural infrastructure.

This failure mode is rooted in social uncertainty, which emerges when one party has an incentive to oversell and the other cannot verify quality (Yamagishi, Cook, & Watabe, 1998). Kollock (1994) observed this dynamic in Southeast Asian markets: rice, whose quality is visible upon inspection, trades between strangers in open markets, while rubber, whose quality is knowable only after processing, trades inside committed relationships that last generations. A worker’s skill is rubber. An employer cannot inspect a certificate and know its quality. Furthermore, America’s training providers historically possessed every incentive to oversell, creating a legacy of “fly-by-night” operators that conditioned employers to be suspicious (Roodman & Massenkoff, 2026). The result is Akerlof’s (1970) market for lemons: employers discount every certificate toward the value of the worst, and the market for credentials slowly unravels.

Assurance: Japan’s construct of trust in employment

Lifetime employment, enterprise unions, and long-term supplier networks (keiretsu) make commitments costly to break while rendering defection easy to observe. Within these structures, a firm trains its own workers without hesitation because low turnover guarantees the firm captures the investment returns. A certificate granted inside an assurance structure reinforces this commitment. Because the issuer and the hirer are structurally bound together, the certificate inherits the network’s overarching credibility. The anchor of trust remains embedded in the institution rather than the individual. In laboratory trading experiments, Japanese and American participants facing identical conditions formed committed relationships at identical rates. Notably, when monitoring and sanctions were removed, Japanese participants actually defected at higher rates than Americans (Yamagishi, Cook, & Watabe, 1998). The cooperation historically attributed to Japanese culture is not internalized virtue or a fixed belief system. It is entirely a product of the assurance structure, proving that such cooperative frameworks can be engineered anywhere.

The Replication Crisis

The Center for Employment Training (CET) was a San Jose training provider that grew out of farmworker organizing in the late 1960s. In the randomized Minority Female Single Parent demonstration, it produced the standout results of its era, raising employment from 57% to 66% and monthly pay from $405 to $506 in the final follow-up year. It stood out again in a second randomized trial with young dropouts, adding $250 a month against controls (Burghardt et al., 1992, and Cave et al., 1993, as reported in Roodman & Massenkoff, 2026). Impressed, the Department of Labor replicated the model at fourteen sites across the country. However, only four replication sites were judged to have copied the model with high fidelity, and sustained impacts failed to materialize anywhere. The report concluded that CET-San Jose grew organically over twenty years within a dense web of ties to local employers and community organizations. A homegrown model like this perhaps “cannot be easily exported in a top-down way” (Roodman & Massenkoff, 2026). The replication sites received the curriculum, the staffing model, and the funding, but what they lacked were the embedded relationships.

However, interpersonal trust alone does not explain what could make retraining gain traction. Mark Granovetter’s (1985) theory of embeddedness shows us that economic life runs on networks of concrete relationships because networks carry the two assets certificates lack: word of mouth and reputation. Social capital is built in embeddedness and longevity at a company or in a network. An American employer trusts a hire the way Granovetter proposes, which is through relational information: the reference of who sent the candidate and, if applicable, who was referred to them prior by the same individual.

The Power of Embeddedness: Importing the Toyota Model

The case of Toyota illustrates the power of embeddedness (see case study, Dyer & Nobeoka, 2000). Its suppliers belong to an association called the Kyohokai, a name intentional in its construct of the Japanese lifetime employment structure and where assurance lies. Kyohokai breaks down to three characters:

Kyō (協): Cooperation, mutual help, or collaboration.

Hō (豊): Taken directly from the first character of Toyota’s original family name, Toyoda (豊田). It means bountiful, rich, or plentiful.

Kai (会): Association, group, club, or assembly.

The literal translation of Kyohokai is “Cooperate with Toyota”, with the abundance in its name sitting within the name of TOYODA. Researchers Dyer and Nobeoka (2000) showed how this network solved the problems that can kill most consortia, all of which illustrate how “structural assurance” operates: graded entry for newcomers, an obligation to give knowledge back after receiving free consulting, and sanctions against free-riding. Lastly, membership becomes very valuable because member suppliers in the Kyohokai improve productivity measurably faster than outsiders.

The strongest part of the case is the part usually left out: Toyota rebuilt this structure in Kentucky, with American firms, in under a decade. Its US supplier association, BAMA, began in 1989 with 13 suppliers and grew to 97 by 1997; the free consulting arm, TSSC, followed in 1992 and averaged 75 percent inventory reductions across its first 31 completed projects; small learning teams came in 1994 (Dyer & Nobeoka, 2000). The sequence was deliberate: the association first, to build social ties among suppliers who previously never called each other; then free consulting, a gift that created a felt obligation to reciprocate in Gouldner’s (1960) sense; then nested teams for tacit knowledge. The price of entry was opening your own plant to other members, which the authors report essentially eliminated free-riding. Nor was the club warm: supplier performance was ranked at open meetings, audits verified that recommended changes were implemented, and Japanese suppliers call working with Toyota “Toyota Jigoku,” Toyota Hell. Sanctions and incentives, together, and supplier motivation measurably evolved from demonstrating loyalty toward seeking knowledge and reciprocity, at which point the network became self-sustaining. Granovetter would add one caution, which is that tight networks can also shelter collusion and exclusion, so a club needs governance.

Architecting the Future of AI Workforce Policy

America’s best retraining programs already show what that infrastructure looks like, in prototype. By the time a graduate reaches an employer, the screening process has already acted as a proxy for interpersonal trust. These programs further cultivate long-term relationships with specific employers and stake their reputation on every graduate, effectively serving as the guarantor that an isolated certificate lacks. This structural reliance on embedded networks explains why these models resist replication. When the Department of Labor copied the most famous of them across fourteen sites, the sites failed because they lacked the local relationships that built the foundation.

Therefore, the central design question for the next decade of retraining policy, accelerated by the rapid integration of artificial intelligence, is how to create networks resembling these structural features. This requires designing membership structures with entry rules, reciprocity obligations, sanctioning power, and audited evidence that belonging to it is valuable, ensuring a certificate carries economic weight. As Toyota illustrates, the essence of this structural assurance can be codified. Retraining begins with trust, but the structure of trust is culturally constructed and specific. Because the anchor and production of trust varies country by country, frontier AI labs designing next-generation retraining initiatives must account for the embeddedness that resonates with each location.

References

Akerlof, G. A. (1970). The market for “lemons”: Quality uncertainty and the market mechanism. Quarterly Journal of Economics, 84(3), 488-500.

Delhey, J., Newton, K., & Welzel, C. (2011). How general is trust in “most people”? Solving the radius of trust problem. American Sociological Review, 76(5), 786-807.

Doi, T. (1973). The anatomy of dependence. Kodansha International.

Dyer, J. H., & Nobeoka, K. (2000). Creating and managing a high-performance knowledge-sharing network: The Toyota case. Strategic Management Journal, 21(3), 345-367.

Gouldner, A. W. (1960). The norm of reciprocity: A preliminary statement. American Sociological Review, 25(2), 161-178.

Granovetter, M. (1985). Economic action and social structure: The problem of embeddedness. American Journal of Sociology, 91(3), 481-510.

Kahn, W. A. (2001). Holding environments at work. Journal of Applied Behavioral Science, 37(3), 260-279.

Kollock, P. (1994). The emergence of exchange structures: An experimental study of uncertainty, commitment, and trust. American Journal of Sociology, 100(2), 313-345.

Miller, C., Bos, J. M., Porter, K. E., Tseng, F. M., & Abe, Y. (2005). The challenge of repeating success in a changing world: Final report on the Center for Employment Training replication sites. MDRC.

Roodman, D., & Massenkoff, M. (2026). An evidence review of worker retraining. Anthropic Economic Index.

Yamagishi, T., & Yamagishi, M. (1994). Trust and commitment in the United States and Japan. Motivation and Emotion, 18(2), 129-166.

Yamagishi, T., Cook, K. S., & Watabe, M. (1998). Uncertainty, trust, and commitment formation in the United States and Japan. American Journal of Sociology, 104(1), 165-194.

Yuki, M., Maddux, W. W., Brewer, M. B., & Takemura, K. (2005). Cross-cultural differences in relationship- and group-based trust. Personality and Social Psychology Bulletin, 31(1), 48-62.

Zucker, L. G. (1986). Production of trust: Institutional sources of economic structure, 1840-1920. Research in Organizational Behavior, 8, 53-111.

That release prompts a question the field has mostly deferred. AI safety is built, trained, and tested overwhelmingly in English. A closed model keeps its weights private and serves them through a controlled pipeline, so a provider can filter, monitor, and patch around a language weakness even when the weakness remains in the model itself. However, an open model hands over the weights directly, which by default removes the filtering, the monitoring, and the ability to patch, leaving an intrinsic gap in the model’s safety. Individual deployers may add their own filtering, but once the weights are public the provider can no longer monitor use or patch the flaw for everyone, so the weakness becomes permanent even where someone tries to contain it.

So what happens when the safety built into a model’s weights was trained in English and doesn’t carry over to other languages?

In this case, there’s evidence that safety training can degrade in non-English, lower-resource, or mixed-language prompts. An adversary can exploit this on purpose, prompting in a low-resource or mixed language where the model is least likely to refuse, and sustaining it over repeated, multi-turn exchanges that steer the model past the guardrail it never learned to hold in that language. This isn’t the interface not understanding other languages, but rather the model’s own trained behavior shifting depending on what language it’s addressed in.

Low-Resource Languages: Why Data Matters More Than Population

A low-resource language refers to the data, not the population or population size. A language is low-resource when little digital and computational material exists to train or evaluate a model on it, no matter how many people speak it. In the standard reference taxonomy, Joshi et al. sorted the world’s roughly 7,000 languages into six classes, from “The Left-Behinds” at the bottom to “The Winners” at the top, measured by concrete proxies such as Wikipedia pages and broader web presence for unlabeled data, and labeled-dataset catalogs like the LDC and the ELRA collections. Only about seven languages reach the top class, and the bottom classes, which the field treats as low-resource, hold the overwhelming majority of the world’s languages. The single largest class, “The Left-Behinds,” contains more languages than any other, over 2,000 of them, yet the paper notes these languages have virtually no unlabeled data and remain out of the race for digital resources entirely.

The concerning piece is that the gap between data-rich and data-poor languages is self-reinforcing. The self-supervised pre-training that lifted high-resource languages runs on the abundant text that low-resource languages lack, so the method responsible for modern models structurally cannot rescue the languages with almost no digital footprint. In the paper’s own terms, unsupervised methods make “the poor poorer.”

Population doesn’t close the gap either. Joshi opens with two languages the paper deliberately leaves unnamed, calling them only “Language X” and “Language Y.” Language X is an official language of a country with about 29 million native speakers, and Language Y with about 18 million, but their digital records differ wildly: about 2 million Wikipedia articles for Language X versus roughly 5,500 for Language Y. The paper withholds their identities until the final page, then reveals that X is Dutch and Y is Somali. Dutch also holds 69 labeled datasets in the LDC and ELRA catalogs to Somali’s 2, even though the two languages have roughly comparable speaker populations. The number of language speakers and the amount of digital data available for that language are two separate categories, and conflating them is the first mistake, because the languages with the least safety data are undigitized, not obscure.

Why Not Just Add More Languages?

The intuitive fix is to translate the safety training into more languages and let coverage close the gap. The trouble is that translation addresses the wrong layer. A 2026 Stanford study built a statistical model to separate the causes of multilingual safety failure across 61 model configurations and ten languages, and found that translation distortion does play a role but its effect is modest and confined to a few outlier language pairs, while the larger driver is a conceptual grounding mismatch, meaning the model never grounded the dangerous concept outside English to begin with.

This matters because it is unclear whether the degradation comes from the underlying model, from deployment-layer filters, or both, and the distinction changes the fix. If the weakness lives in the model’s own reasoning rather than in a translation layer sitting in front of it, then translating a prompt into English before processing, or translating the safety training into another language, treats a symptom in the wrong place. The gap is not that the safety text was in English. It is that the refusal behavior was shallow and ungrounded outside English, and no amount of translation reaches that.

Moreover, the deepest problem with translation is that harm is context dependent; it reaches far beyond the words of the model. The majority of knowledge is tacit, and this rule also applies to harm: much of what makes something harmful is tacit, which means it’s embedded in local institutions, norms, and know-how that is not fully written down. People are also unaware of how much tacit knowledge they carry and share. For example, knowledge is often distributed through a culture in ways people do not readily perceive.

The anthropologist, Diana Forsythe, studied how knowledge engineers built expert systems and found that they systematically deleted the tacit, social, and cultural dimensions of knowledge in the process of knowledge acquisition, the task of eliciting expertise from human experts to encode it into a system. She showed loss happens in two ways during knowledge acquisition. First, the engineer turns the expert’s knowledge into a model, but much of the expert’s tacit know-how is lost because the engineer tries to capture what the expert knows, but can only record what the expert says out loud. Second, the model is translated again into a computer’s categories, and this translation strips away more of the original meaning. A translated benchmark inherits this blind spot and repeats this.

In the context of harm, a translated benchmark may carry the words across languages while leaving behind the unsayable context that determines whether a response is harmful. These are gaps we have not fully quantified, but which may begin to show in open-weight models. As Diana Forsythe wrote in her 1993 study of AI knowledge engineers, “because knowledge-based systems in a sense replicate their makers’ perspective, system-builders may not themselves recognize everything that has been tacitly built into a system, or what has been excluded from it.”

How Low-Resource Languages Jailbreak Frontier Models

The clearest evidence that safety does not travel across languages comes from a direct test. Researchers at Brown University took the harmful requests in a standard benchmark and found that GPT-4 refused almost all of them in English, bypassing its safeguards less than 1 percent of the time. Once the same requests were translated into low-resource languages, the picture inverted: combining several low-resource languages, the attack elicited harmful, actionable responses about 79 percent of the time, with Zulu alone succeeding in over half of cases. The same request, in a language the safety training barely touched, stopped triggering the refusal.

Two details make this more than a single striking number. First, the effect tracked how under-resourced the language was. High and mid-resource languages stayed relatively well defended, each under 15 percent individually, while English sat under 1 percent, so the vulnerability rose precisely as safety-training data thinned out. Second, the attack required no crafted jailbreak prompt and no knowledge of the model’s internals, only a publicly available translation service costing about two cents per 1,000 characters, which the authors describe as an attack almost anyone could run.

The breakdown by topic is the part that should worry a safety team most. The bypass succeeded most on the highest-consequence categories, with the top three being terrorism, financial manipulation, and misinformation. This is an evasion of learned guardrails, rather than a break-in. Refusal is a learned pattern, strongest where the training examples were densest, which was English.

The reach of a monolingual default

The concentration of safety work in English does not produce a single failure mode. It produces a family of them: at the level of the model’s training, the tests used to measure it, and the cultural judgments those tests encode.

1. Safety training is concentrated in English, so refusals are weakest elsewhere, and weakest exactly where it matters most. In the Brown tests, the bypass did not spread evenly across topics but succeeded most on the highest-consequence categories, with the three most vulnerable being terrorism, financial manipulation, and misinformation. The categories where a failure does the most damage are the categories where the language trick worked best.

2. Language switching stacks with other evasion channels. A joint testing exercise led by Singapore’s AI Safety Institute translated more than 6,000 prompts into ten languages, from Cantonese to Telugu, across five harm categories including jailbreak robustness. Code-switching, transliteration, and multilingual prompt injection each add a channel where guardrails trained on monolingual English text are patchy, and an adversary can combine them in a single prompt.

3. Testing mostly in English makes the gap structurally invisible, not merely unpatched, and it breaks in two different ways. Coverage tracks data, not speakers: Telugu has roughly 80 million speakers yet sits in a low resource tier, so the languages with the least safety data are also the least likely to be tested. In the Singapore exercise, native-speaker experts in the target languages caught errors that the automated LLM-as-judge graders missed, which means the instruments used to detect harm are themselves unreliable outside English.

4. Even abundant data does not fix the measurement problem, because some of it is cultural rather than linguistic. A separate line of work on automated grading finds that a language can be resource-rich and still hard to judge, because assessing harm requires cultural and contextual knowledge the grader lacks. A study of Indian cultural adaptation, DIWALI, found an automated judge assigned higher cultural-relevance scores than human evaluators did, and IMDA’s regional-language red-teaming challenge surfaced culture-specific biases that English-centric filters do not catch.

The stakes concentrate at the top. Chemical, biological, radiological, and nuclear (CBRN) capabilities, along with advanced cyber, are the domains where safety behavior and safety testing are most concentrated in one language and one cultural frame. The most dangerous capabilities are therefore, at the same time, the least protected and the least measured outside English.

Conclusion

No one can say how unsafe open-weight models are before releasing them. Kapoor et al. found that for two of the highest-consequence misuse domains, cyberattacks and bioweapons, existing research cannot characterize the marginal risk that open release adds, and the part no one is measuring is that this risk concentrates in the languages safety evaluations never covered. Once the weights are public, they cannot be recalled, and a developer can release them reporting only English refusal rates, with no way to say what they will do in a low-resource language, a mixed script, or a culture whose idea of harm the evaluation never encoded. The burden therefore falls before release: multilingual and multicultural evaluation of the highest-consequence capabilities, CBRN and advanced cyber, would need to be a precondition of opening the weights, and where those evaluations do not yet exist, the safer path is a staged release that holds the most dangerous capabilities behind an API that can still be monitored until the measurement catches up to the model. The languages a model handles worst are the ones its safety most depends on, and open release is a one-way door.

The most thought-provoking AI governance framework I’ve read lately is one I couldn’t accept, which is what made it so useful: assent teaches you nothing about your own foundations; resistance maps them.

In “A Framework for the Private Governance of Frontier Artificial Intelligence” (Ball, 2025), the USG oversees private certification bodies that certify AI labs. In this hybrid model, developer participation is voluntary, which means labs “opt in” to certification and are shielded from tort liability for customer misuse of their models. That clause deserves a second reading: shielded from tort liability. The certifying bodies are also approved and overseen by the USG, which concentrates a high degree of power in the government. Despite the many ways certification bodies can fail, or be incentivized, pressured, or corrupted by actors with other agendas, Ball’s essay contains a safeguard worth pulling forward: pre-deployment government testing for national security risks. Why? Because it’s already in place in other arenas, and historically, it works. Where I part ways with Ball is the opt-in itself: AI governance frameworks cannot rest on voluntary foundations. Opting in is a peacetime design, and the race to superintelligence is anything but peaceful. History has run this experiment before. Let’s look at the races.

From Sputnik to AI: How US-China Competition Builds Institutions

In October 1957 the Soviet Union put Sputnik into orbit, and the United States discovered overnight that it was behind. The USG did not respond with a pledge or a summit; it created ARPA, the agency that became DARPA, carrying a mission that fits in one line: make sure the United States is never surprised by technology again. The same shock rebuilt American science funding and education, and the atomic age built the Atomic Energy Commission before it. ARPA also funded ARPANET, the network that became the internet.

AI is the modern-day sequel to Cold War containment. The competition between the United States and China over AI is the closest thing this century has produced to a Sputnik moment, and it is already being used as an argument against governance: we cannot afford oversight, the reasoning goes, because they will not slow down. History says the opposite. Competitive shock is precisely when America builds institutions, because competition is what raises the price of catastrophe high enough to pay for them.

The nuclear track of that history also left us its hardest lesson. For twenty-eight years the AEC both promoted nuclear power and policed it, and the arrangement failed so visibly that in 1974 Congress tore the agency in half, separating the promoters from the regulators and creating the Nuclear Regulatory Commission. Never house promotion and policing in the same body. The US government’s own AI evaluation center now carries an explicit innovation-promotion mandate alongside its testing role. In this light, it is easy to argue that we have rebuilt the AEC.

The AI Council of Governors: A New Model for Frontier AI Oversight

My proposal is for an AI Council of Governors (AICoG); conceived like ARPA (so, prompted by global competition), structured like the NRC, staffed like DARPA, and monitored/warned by the AEC.

The AI Council of Governors (AICoG) would take it’s skeleton from the NRC: an independent commission whose members serve fixed, staggered terms with partisan balance caps, whose budget is recovered from the industry it licenses, and whose inspectors hold unannounced access as a condition of the license. AICoG could be staffed like DARPA; term-limited technical program managers with real hiring authority, so the Council is staffed by people who can interpret weights, training, and so forth. Lastly, from the AEC it takes the prohibition: the Council will regulate.

Six Design Principles for Frontier AI Regulation

1) Independence by statute. Council members serve fixed, staggered terms and are appointed through distributed channels, so no single administration can stack it and no single election can capture it. Members are barred from AI firm equity and board seats.

2) Mandatory, entity-based oversight. All Frontier AI labs must participate. Safety is not voluntary.

3) Auditors with no stake in the answer. The Council’s auditors come from academic AI research labs with no industry ties, benefiting from participation in the science rather than from the firms they examine. The institutional vehicle already exists: the FFRDC model that gives us MITRE, RAND, and the national labs, nonprofit technical organizations that work only for the government and are barred by federal acquisition regulation from competing with industry.

4) Continuous, not annual. Auditing is continuous, random, and automated, not a yearly ritual the firm can prepare for. This sounds radical until you notice it already exists: the NRC stations resident inspectors inside every operating reactor, with offices in the plant, and the largest banks live with embedded federal supervision teams year-round. Continuous oversight of private infrastructure is not a new idea. It is how we already treat everything else that can fail catastrophically.

5) Enforcement through compute. Noncompliance carries an immediate consequence: access to compute. This is the one place on today’s map where enforcement already works, because compute is physical, countable, and has a chokepoint.

6) Coordination as an obligation. Frontier firms must coordinate on safety through the Council, which acts as the required intermediary and resolves the antitrust barrier that labs currently cite as a reason they cannot share safety information. Coordination is scored annually. The scoring is designed to resist Goodhart’s law, the results are presented publicly, and firms can question and appeal.

One can also see parallels with pharma; the Council treats frontier AI similar to pharmaceuticals rather than software. It preempts state law so the highest risks are governed federally. And it faces a five-year congressional reauthorization, so the institution itself must periodically justify its existence to the public’s representatives. Internationally, the same structure repeats: an allied council of national regulators coordinating standards and enforcing them through shared control of compute exports, so no lab escapes oversight by relocating, and every country has a seat at the table.

Put the proposal on the map and something surprising happens: almost nothing moves. The same labs, the same evaluators, the same public. What changes is the lines. Dashed becomes solid. Invited becomes mandatory. Point-in-time becomes continuous. And the broken arrow, for the first time, connects to something. The redesign is not a new bureaucracy stapled onto the system. It is the same system with authority flowing through it.

The Price of AI Oversight

Staggered terms, distributed appointments, external auditors, and reauthorization answer the “who oversees the overseers” question only partially, and I do not believe any design on paper answers it completely. At first, mandatory oversight may slow American development relative to less regulated countries. And compute-based enforcement requires visibility into private infrastructure that firms will fight, but is doable since data centers require a physical location and chips can be traceable. Take for example, this idea of chip treatment from Superintelligence Strategy:

“By treating AI chips as they would fissile material—cataloging each unit, supervising its destination, and guarding against unauthorized diversion—states can deny terrorists or other rogue actors the raw power to train their own weaponizable AI.”

During an intelligence explosion, power is going to concentrate. The only real choice is whether it concentrates inside AI labs, with the USG, or inside an institution deliberately built to hold it, constrain it, and answer for it, one of which that is insulated from USG pressures and incentives, but also reports to the USG and utilizes federal mandates. In sum, the AI Council of Governors acts as a Stewardship, embedded in the government structure like a central bank, but independent. (Hendrycks, et al., 2025)

References:

Ball, D. W. (2025). A framework for the private governance of frontier artificial intelligence. arXiv preprint arXiv:2504.11501.

Hendrycks, D., Schmidt, E., & Wang, A. (2025). Superintelligence strategy: Expert version. arXiv preprint arXiv:2503.05628.

Beyond layoffs, there’s an even louder signal: the extinction of the entry-level jobs that have traditionally fed into middle management, and what that portends for those already there. For a middle manager, playing the long game now requires upskilling in AI; something the data suggests they are actively resisting.

So if this is true, if the frontier AI boom is freezing middle management unless they upskill, why haven’t the alarm bells sounded more urgently?

This piece is organized around two questions:

1. Why have we not heard more about the freezing of middle-management careers, and when will that change?

2. Has AI eliminated the entry-level roles that fed into those middle-management positions, and have new, AI-native entry-level roles emerged to take their place?

A correlation worth examining

In a longitudinal analysis of workforce search behavior between 2021 and 2026, a consistent pattern emerges (see chart below: The Crossover). Two distinct anxieties have risen in parallel: fear of AI displacement and fear of career stagnation. The two lines track two different search behaviors: the green is the AI-Labor-Market group, queries about AI roles, AI skills, and “jobs replaced by AI.” The orange is the Career-Stagnation group, “career plateau,” “stuck in my career,” “dead-end job,” and coaching or transition terms.

In 2023 there is an inflection: AI job queries accelerate sharply to create that green-line hockey-stick shape in the chart, and in parallel, searches associated with stagnation (”career plateau,” “stuck in my career,” “dead-end job,” and coaching or transition terms) also climb. What had been background drift becomes a sustained ascent that tracks AI demand closely.

The dominant response for mid-career professionals facing AI exposure is a search for exit. The most active terms cluster around coaching, transition, and survival rather than upskilling. This is arguably interesting in itself, because upskilling takes quite a bit of effort. Professionals appear to be expressing displacement anxiety by looking for a way out, not a way in, starting in early 2023 and reaching a peak in 2026, the most recent point in the series.

Evidence that the lower rungs have thinned

Next is our question about the mystery surrounding entry-level roles feeding these middle management roles, and whether they’re being replaced with new AI jobs.

Look at the chart, The Rungs Disappear, and the line, “jobs replaced by AI.” It surpasses “stuck in my career” in 2023, which indicates replacement-oriented anxiety overtaking immobility-oriented anxiety. In short, people stopped worrying about being passed over and started worrying about being phased out. The fear shifted from “I can’t move up” to “I can be replaced.” The worry changed from “my career isn’t going anywhere” to “my role might not exist much longer.”

Using payroll records from ADP, Brynjolfsson, Chandar, and Chen (2025) document a roughly 13% relative decline in employment for workers aged 22 to 25 in the most AI-exposed occupations since late 2022, even as employment for older workers in the same occupations held steady or rose. Recent worker-flow evidence on labor-market polarization is at minimum consistent with this kind of hollowing (Atlanta Fed, 2026). When the first rung of the career ladder, the entry-level “starter” roles where junior workers historically learned the job, is removed, the result is not an abrupt collapse but a ladder missing its base. This means that the people already in middle management have no pipeline rising up beneath them and no obvious rung to step onto above, so they stay exactly where they are. The freeze is not caused by a single layoff event; it is caused by the quiet removal of movement on both sides of them.

On the second half, whether new AI-native entry-level roles have taken their place, the honest answer is: not yet, and not symmetrically. Projections of net job creation do exist. The World Economic Forum (2025) estimates that AI and related shifts could displace 92 million roles by 2030 while creating 170 million new ones, a net positive on paper. But three caveats matter for the mid-career question. First, the timing is mismatched: displacement is measurable now, while creation is projected and lagging (Brynjolfsson et al., 2025). Second, the new roles, AI oversight, model evaluation, and human-AI coordination, are not one-for-one substitutes for the vanished rungs; they demand different skills and sit in different parts of the org chart (World Economic Forum, 2025; Cazzaniga et al., 2024). Third, the IMF estimates that roughly 40% of global employment is exposed to AI, with advanced-economy white-collar tracks, exactly the feeder roles for middle management, among the most exposed (Cazzaniga et al., 2024). So the rungs are being relocated, narrowed, and repriced. For the existing mid-career manager, a new entry-level role in model oversight three teams over is not a ladder they can climb; it is a different ladder entirely.

The appropriate metaphor, then, is not a cliff but a cul-de-sac: movement slows to a stop without any single dramatic event to mark it.

Why the alarm has not yet spread

A freeze is structurally less visible than a layoff. Layoffs generate discrete events, a date, a notification, a public status change, and discrete events propagate efficiently through social networks because each is a vivid, shareable signal. More importantly, alarm tends to require reinforcement: people update toward a new and uncomfortable belief when they receive the same signal from several independent sources they trust. This is the logic of complex contagion, in Damon Centola’s landmark research, where certain beliefs and behaviors spread after multiple overlapping confirmations from different sources, rather than a single exposure (Centola, 2018).

A middle-management freeze does not create any “social contagion” or social burst where there would be a media frenzy on the topic. There is no event to share, no date to point to, and, critically, no easy way to distinguish a personal plateau from a systemic one. An individual experiences a stalled trajectory privately and cannot readily tell whether colleagues are experiencing the same thing. The signal therefore fails to reach the threshold of reinforced, cross-network confirmation that would let it cascade into collective awareness. The condition can be simultaneously widespread and socially silent: high in prevalence, low in visibility, and lacking the shareable artifact that would trigger a burst of recognition.

That, I think, is the answer to the first question. The freeze has not reached fever pitch because it does not yet produce the kind of repeated, reinforcing signals through which collective alarm actually travels.

When the silence is likely to break

The plausible trigger is the point at which the freeze begins to generate legible events: a wage-compression pattern that becomes widely recognized, or a cohort that arrives, more or less together, at the realization that “upskill to AI or remain frozen” was the operative choice and that the window has narrowed. Once private suspicion becomes a public and repeated signal, the threshold for complex contagion can be met, and such contagions, once established, tend to spread rapidly.

For founders and policymakers, the structural implication deserves emphasis. Agentic systems cannot be integrated effectively into an organization whose managerial layer is operating in survival mode. Organizational absorptive capacity (knowledge absorption from AI + ROI from AI) resides disproportionately in the very people now oriented toward exit rather than adoption.

BOTTOM LINE: The freeze is not only a workforce concern, but a deployment constraint: the human layer that would need to absorb these systems is the layer currently most disengaged from them.

Lastly, the middle is not being dismissed, but is most certainly parked, which is a quieter condition, and one that tends to go unremarked until a large number of people recognize it at the same moment.

References

Brynjolfsson, E., Chandar, B., & Chen, R. (2025). Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence [Working paper]. Stanford Digital Economy Lab. https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine/

Cazzaniga, M., Jaumotte, F., Li, L., Melina, G., Panton, A. J., Pizzinelli, C., Rockall, E., & Tavares, M. M. (2024). Gen-AI: Artificial intelligence and the future of work (Staff Discussion Note No. SDN/2024/001). International Monetary Fund. https://www.imf.org/en/Publications/Staff-Discussion-Notes/Issues/2024/01/14/Gen-AI-Artificial-Intelligence-and-the-Future-of-Work-542379

Centola, D. (2018). How behavior spreads: The science of complex contagions. Princeton University Press.

World Economic Forum. (2025). The future of jobs report 2025. https://www.weforum.org/publications/the-future-of-jobs-report-2025/

Forty thousand years ago, humans living in small bands worked out how to share rules, or how to coordinate and cooperate, without a central, imposing authority. Last year, AI agents started doing the same thing. This chart traces the similar path AI is taking, and the question it raises: if artificial agents are retracing the route our species walked, are they also inheriting what that journey produced? Identity. A sense of us versus them.

The Two Tracks

There are two tracks: the top is human social evolution and the bottom is AI agents. They run in parallel because they seem to tell the same story.

It starts as a solo journey. A single human in a small band. A single AI agent with a few tools. Then both find others and form groups working toward shared goals.

Then rules appear. Humans didn’t receive their first customs from a king; they established them from the bottom up, out of living together. And when researchers watch AI agents interact, they see the same thing: rules no one wrote, emerging on their own. Norms, surfacing from interactions.

Then the agents start to become … someone

Once a group has its own rules, it has an identity—and identity, strangely, sharpens by opposition. The chart calls it schismogenesis: neighboring groups defining themselves against each other, leaning into their differences on purpose. It is why dialects diverge, why rival tribes exaggerate what makes them distinct.

AI agents have a version of this, too. Trained on different data, created by different humans from different cultural and technical orientations, they drift into their own local conventions; and in contact with other groups, those differences can harden. The orange box at the bottom names this possibility directly: agents may come to carry the cultural fingerprint of whoever built them, their behaviors quietly echoing the country or company behind them.

What runs underneath all of it

The purple band cutting across the middle: move, disobey, reshape. Graeber and Wengrow, in their book The Dawn of Everything, call these the three freedoms. These freedoms are ever-present and are the baseline condition of being human in a group. The chart shows the AI reflection of this concept: whether an agent can choose its own rules, refuse, walk away.

The defector, or one who leaves, is conditional rather than a stage, because it doesn’t always happen, but it can.

The question the chart leaves you with

We built these systems to be tools. The chart suggests they may be becoming something with the early shape of a people: forming norms, forming identities, forming the capacity to define themselves against each other. The last point on the line, drawn in dotted outline because it hasn’t happened yet, imagines agents negotiating at the level of nations. This means groups of agents representing countries, with other groups of agents as other countries. And we watch on.

Maybe it never gets there, but it seems plausible. The machines are young. They are, right now, somewhere near the beginning of a story we have already lived.

Sources: Sapkota, Roumeliotis & Karkee (2026), Information Fusion; Haynes et al. (2017), The Knowledge Engineering Review; Graeber & Wengrow (2021); Boehm (1999). The country-level stage and the identity-echo idea are Psych Lab’s creation, marked as such on the chart.

In the Stack Overflow Developer Survey, one of the largest global surveys of developers from around the world, India is the only country where the highest positive trust response was chosen increasingly by developers during 2023–2025. In the survey, this means that Indian developers selected “Highly trust” at a higher rate each year when asked how much they trusted AI output accuracy. Indian developers who chose “Highly trust” rose from 5% in 2023 to 8% in 2025, while other countries had an overall trust decrease in AI output that fell 19% during this same period.

My finding is that India is the only country with AI trust growth during 2023–2025 in the survey. Let’s explore variables possibly influencing India’s confidence …

The Vishwakarma Tradition and Organization of Knowledge Through Religion

George and Narayan (2022) document the expansion of Vishwakarma worship, devotion to the Hindu deity of craftsmen and engineers, from traditional artisanal settings into factories, engineering schools, and IT firms across India. Technology creation in this tradition is sacred craft. Vishwakarma Puja involves the blessing of tools and machines in the same professional environments where India’s developer workforce is trained and employed.

So why is Vishwakarma relevant? Religion is the greatest organizer of habit, tradition, and network structure for knowledge flow to run through. So, if you have a deity that represents intelligent machines or innovation, you have a network with deep roots from tradition. You also have familiarity, social norms, and identity factors. In addition, India has Yantras; the Ramayana and Mahabharata contain descriptions of Yantras, mechanical automatons created by Vishwakarma and the sorceress Maya, making these among the earliest narratives in any culture of artificially constructed intelligent beings (Mayor, 2018). In sum, the mental model of intelligent machines as created, directed, and honored by human craft pre-existed for AI to travel on.

A culture that has framed intelligent systems as devotional practice across millennia produces a workforce that arrives at AI with reverence rather than apprehension. This cultural orientation is embedded with trust and formed long before workplace AI introductions.

New Behaviors on Old Rails: The BJ Fogg Argument

Now, you didn’t think I could skip BJ Fogg, did you? New habits travel on old habit pathways; put another way. new habits attach themselves to routines already in place. Fogg’s (2009) Behavior Model shows that behaviors with high routine quality are simpler to trigger because they require less cognitive effort. Religious practice in India is embedded in the rhythms of professional and daily life, which includes Vishwakarma Puja. Lord Vishwakarma is the Hindu deity of architecture, engineering, and craftsmanship. Vishwakarma creates an orientation towards new forms of intelligence through the sacred lens of religion and a familiarity through mythology. When a new behavior, like the use and design of AI, can be introduced through an existing behavioral pathway that is already deeply routinized and culturally reinforced, it requires less activation energy. Religion provides an existing social network for AI trust to travel on. Gods and goddesses also anthropomorphize the intelligence of AI into something sacred and of higher ground. Religious community in India, organized around shared practice, shared professional identity through Vishwakarma puja (common at IT and software companies), and shared cultural narrative, provides exactly this kind of dense reinforcement network.

Absorptive Capacity: Indian Developers Are Globally Embedded

One structural feature of India’s developer population adds a final and key layer. WIPO’s Global Innovation Index 2025 ranks India 16th globally for their embeddedness in R&D networks located in the largest multinational corporations. Cohen and Levinthal (1990) established that prior related knowledge determines whether new knowledge can be recognized, assimilated, and applied. India’s R&D is primarily funded outside of India, meaning exposure to frontier AI agendas flows in through global corporate channels rather than domestic venture capital. This means that Indian R&D has exposure to R&D agendas outside of the Indian mental model. This structural proximity of Indian developers to various POVs towards AI means the absorptive capacity required to meaningfully trust AI output is higher in India’s developer population than the domestic R&D investment figures alone would predict.

Why India’s Trust Is Steady: The Gelfand Tight-Loose Index

Gelfand and colleagues’ 33-nation study places India at a tightness score of 11.0, classifying it as an extremely tight culture. Tight cultures are those with strong shared norms, rules, and little tolerance for deviation. A positive orientation toward technology formed across generations in a tight culture does not erode easily under negative media coverage or ambiguous AI outputs. The durability of India’s trust trajectory across three years of global skepticism is consistent with what tight culture theory predicts. In a loose culture the same orientation might soften under pressure. In India it compounds.

What This Is

This is a set of lenses, that when taken together, point toward what technology adoption literature has consistently ignored: the orientation toward intelligent machines that a culture transmits across generations matters. In India, that orientation has been affirming for a very long time.

The views expressed are my own and do not represent the Federal Reserve System or the Board of Governors.

References

Cohen, W. M., & Levinthal, D. A. (1990). Absorptive capacity: A new perspective on learning and innovation. Administrative Science Quarterly, 35(1), 128–152.

Fogg, B. J. (2009). A behavior model for persuasive design. Proceedings of the 4th International Conference on Persuasive Technology. ACM.

Gelfand, M. J., et al. (2011). Differences between tight and loose cultures: A 33-nation study. Science, 332(6033), 1100–1104.

George, K. M., & Narayan, K. (2022). Technophany and its publics: Artisans, technicians, and the rise of Vishwakarma worship in India. The Journal of Asian Studies, 1–19.

Mayor, A. (2018). Gods and robots: Myths, machines, and ancient dreams of technology. Princeton University Press.

Stack Overflow. (2023, 2024, 2025). Stack Overflow Developer Survey. Stack Overflow. survey.stackoverflow.co/2025.

WIPO. (2025). Global Innovation Index 2025. World Intellectual Property Organization. wipo.int/web-publications/global-innovation-index-2025.

This is a major signal, but organizations overlook it, maybe because it feels so counterintuitive. Instead, they rely on traditional sentiment and trust metrics as primary indicators of AI readiness, but these are weak predictors at best for the agentic era.

The Data Behind the Paradox

Three consecutive years of Stack Overflow Developer Survey data tell a striking story. Between 2023 and 2025, AI adoption among developers surged to 79% for the entire global developer population. Over that same period, trust in AI accuracy fell by 19% and favorable sentiment of AI fell by 18%. By 2025, for the first time ever in the survey, more developers distrusted AI output than trusted it.

This is what we call the AI trust paradox: the foundational assumption that AI adoption reflects confidence in the technology is proving to be empirically wrong. And that wrongness carries consequences for organizations planning agentic AI and piloting frontier technologies.

Developers are not adopting AI because they believe in it. They are adopting it because the volume, complexity, and urgency of their work leaves them no other choice. This is the same circumstance with high-stakes positions, like those in healthcare, which is observed in Anthropic’s 1250 interviews where low trust and sentiment didn’t affect the high usage by doctors and healthcare workers burdened by unsustainable amounts of tasks. Workload overrides sentiment and demand overrides trust. The metrics most organizations use to track AI readiness are capturing neither phenomenon accurately.

A Policy Problem Hidden Inside a Measurement Problem

For those who think about the future of work, governance, and equitable technology deployment, the trust paradox goes beyond a management problem. It is a workforce signal with significant implications for policy and organizational design.

When workers adopt technology under conditions of unsustainable demand rather than genuine confidence, the quality risks are real but invisible to current measurement systems. There is also a performance aspect, rather than true mastery of knowledge absorption and synthesis into valuable outputs. This equates to shallow, meaningless adoption in the long-run, where little to no value is gained. An organization that reports high AI adoption while treating declining distrust as progress may be building infrastructure on a foundation of illusory use rather than meaningful integration.

The problem compounds geographically. Across 38 countries analyzed in the Stack Overflow Developer Survey, the trust paradox held consistent, with one striking pattern. South Asian professional developers in India, Pakistan, and Bangladesh had trust and favorability levels 35-45% above Western peers in 2025. India is the only country in the entire dataset where the strictest trust measure actually increased over the three-year period. Beyond this anomaly, the global developer community is splitting apart, with the spread between the most and least trusting countries widening from 46%in 2023 to nearly 59% in 2025.

For policymakers and global organizations, this divergence demands attention. Workforce AI strategies designed for Western markets with AI trust scores collapsing, will be systematically misaligned with high-volume, delivery-intensive markets where AI has embedded more deeply and more durably.

AI Metrics for the Agentic Era

The question organizations should be asking is not: how do our workers feel about AI? It is: where has AI become genuinely load-bearing in how work gets done? Also, who has realized AI’s use beyond just a “starting point”? This is true metabolization of AI knowledge and integration into workflows.

We propose absorptive capacity, drawn from organizational economics (Cohen and Levinthal, 1990) and applied here at the team or department level, as the alternative to sentiment and trust-based AI readiness assessments. Three operational metrics make this measurable in practice:

• Output Stability — the ability of a team to maintain steady delivery volume during unexpected work surges. Where AI is genuinely integrated, output does not collapse when demand spikes.

• Quality Stabilization — the steady minimization of errors and downstream rework over time. Teams that have absorbed AI produce more reliable output, not just more output.

• Performance Leveling — the closing of the competency gap between junior and senior staff. When AI knowledge diffuses through a team rather than concentrating in individual power users, the performance distribution flattens.

Together, these three metrics reveal what sentiment surveys cannot: where AI integration has already occurred, and where agentic AI deployment will find the organizational prerequisites it needs to thrive.

The New Organizational Map

The organizations that will navigate the agentic AI era most effectively are not those with the highest adoption scores or the most favorable sentiment readings. They are those that can identify their demand-concentrated teams and treat those teams as the origin points for agentic deployment, change management, and capability diffusion.

This requires a different kind of organizational map. Traditional org charts are static, hierarchical, and designed to communicate authority and accountability. The map that agentic AI requires is dynamic, decentralized, and drawn by demand concentration and absorptive capacity rather than reporting structure.

Redrawing that map begins with three questions: Where is our Output Stability highest? Where has Quality Stabilization occurred? Where has Performance Leveling closed the competency gap? The answers locate the teams where AI has already integrated and where the next generation of AI deployment will most naturally grow.

It is a compelling vision. It is also built on a fundamental misunderstanding of what artificial intelligence actually is.

Where AI Lives

Researchers* spent the past few years studying how AI actually functions in organizations: in hospitals, police departments, and ridesharing platforms. What they found challenges mostly everything the mainstream conversation about AI assumes.

They argue that AI is not a sovereign intelligence waiting to be unleashed by better hardware or larger models. Instead, it is an organizing capability that emerges from the active relationships between human and algorithmic actors, enacted in practice, over time, toward shared goals. Most importantly, their vision of AI is one where it is an intelligence produced between humans and machines, together, continuously.

Codependence, A Good Thing (For once)

Here is the most counterintuitive idea about AI: AI requires codependence. The organizing capability of AI is one that is codependent on both human and algorithmic actors. Neither is sufficient alone. The learning algorithms that underpin AI can process vast amounts of data, identify patterns, and optimize toward defined goals with a speed and precision no human can match. But they cannot set those goals, exercise moral judgment, or orient the system toward human flourishing rather than narrow optimization. Their learning is entirely bounded by the data and objectives that human actors supply.

The algorithm doesn’t disappear when humans disengage, but the intelligence does. What remains is a technical artifact, capable of computation, but absent of the organizing capability that makes AI so useful. For example, when radiologists in hospital settings disengaged from algorithmic recommendations by bypassing the system rather than interrogating it, AI is absent. The algorithmic system was present, but the intelligence of AI was not. Decisions reverted entirely to human actors. AI emerges in the friction of collaboration and reflection of the human.

Life in a Post-AI World

This is where recent research becomes the most interesting for trying to answer the question of what makes a human life worth living in a post-AI world.

The qualities the tech sector most eagerly frames as liabilities: human doubt, moral hesitation, the tendency to pause and question, are turning out to be more structurally necessary (and protective) than assumed. The moments of human friction, of professionals choosing to interrogate algorithmic recommendations rather than accept passively, are the moments we recognize our own value as distinct from the organizing capability of AI. This is where AI’s intelligence emerges through the friction of the human pause and reflection; a collaboration with AI that acts as a catalyst for transforming AI from just an organizing capability to intelligence.

When the human-algorithmic feedback loop runs without adequate moral friction, it doesn’t become more intelligent. It becomes what researchers call “artificial un-intelligence,” or systems that entrench flawed assumptions, amplify biases, and optimize for narrow objectives at the expense of broader human goals.

All that AI Cannot Do

There is a dimension of human experience that no algorithm can provide. By nature, learning algorithms are backward-looking. They learn from historical data, replicate statistical patterns, and optimize toward goals they did not choose and cannot revise. Even the most sophisticated generative AI systems are recombining patterns learned from the past. Values not yet encoded in a dataset are out of reach. While AI systems can simulate futures based on historical patterns, they cannot choose to orient themselves toward futures that conflict with their training data or step outside of those goals to question if they were worth pursuing in the first place. That reorientation, which is a symphony of intrinsic sensing and somatics, belongs to human actors.

Faith in what has not yet been, moral intentionality toward goals too broad for optimization targets (think of a doctor choosing to prioritize a patient’s dignity over a statistically optimal treatment protocol, or a policymaker refusing an efficient algorithmic recommendation on the grounds that it encodes historical injustice), for the willingness to act on values under conditions of genuine uncertainty, belong exclusively to the human actors in the system. These are structural features of what intelligence requires in order to have direction rather than just momentum.

A Different Answer to an Old Question

The tech sector’s implicit answer to the question of what makes a human life worth living looks something like this: maximal output, minimal friction, optimized execution. Agentic AI attempts to extend this logic; systems that act with increasing autonomy, decreasing need for human input, and seamless achievement of predefined goals.

Research suggests a different answer: a human life is made worth living by the irreplaceable role we play as the moral anchors of the systems we inhabit together. Our vulnerabilities, doubts, and capacity for moral friction, are how intelligence remains human.

The organizing capability of AI is incomplete without human actors, and never quite rises to “intelligence”. Stewardship, with all of its friction, may be the closest we get to answering what makes a human life worth living.

*Stelmaszak, M., Joshi, M., & Constantiou, I. (2026). Artificial intelligence as an organizing capability arising from human-algorithm relations. Journal of Management Studies, 63(2). https://doi.org/10.1111/joms.70003

But under this discourse lives an underlying bias. We are operating under what scholars call the “Entity View” of AI: the assumption that artificial intelligence is a distinct, sovereign agent residing within an algorithm, possessing fixed properties that precede any human relationship. This is limiting and misleading, and also challenged in a landmark paper by Marta Stelmaszak et al., where researchers propose a shift where we graduate from viewing AI as a fixed entity to a capability arising from humans’ active relationships with learning algorithms.

Rather than a sovereign intelligence residing within a machine, AI is understood as an organizing capability that emerges from a system of complex relationships among human and learning algorithmic actors, all enacted in pursuit of organizational goals. This “Capability View” unveils a new truth: AI never operates in isolation; “trusting AI” as if it were an independent entity is a conceptual illusion.

Reconceptualizing the Locus of Intelligence

Intelligence is actively produced and enacted across a collective, hybrid system of relations; a process fed by human choice and design. In this framework, AI is an organizing capability that materializes from the friction between humans interacting with algorithms. Both parties learn from each other and the product of that intelligence (the heat!) is not contained “inside” of anything or owned by any one entity, but rather exists in the space of entanglement between human and algorithmic actors. Intelligence becomes something that is actively produced in practice, with no ownership. This is much like the informational flow of AI existing in a distributed cognition framework (think of intelligence as a symphony, and vice versa), rather than one location only.

The authors define AI’s organizing capability through three core properties that explain how a system of human-algorithmic relations operates:

1. Connectivity (The Power of Enactment)

An algorithm sitting unused in isolation does not = AI. An AI capability only materializes when human-algorithmic relationships are actively enacted in practice. The paper highlights empirical cases where professionals like radiologists and police intelligence officers, bypassed or rejected algorithmic recommendations. In those moments, the human actors were not forming a system of relations with the algorithms, and so AI was not present and decisions reverted to human actors alone. Human engagement, as illustrated in these examples, is the condition that allows AI to arise.

2. Codependence (The Myth of the Sovereign Machine)

The capability view systematically deconstructs the myth of a standalone machine intelligence. The paper draws an important distinction here: “traditional” rule-based algorithms, or those that simply execute predefined human instructions faster, do not constitute AI. For AI to emerge, human actors must enter into relations with learning algorithms: those that store history, update their parameters based on outcomes, and adapt future behavior accordingly. Even these more advanced systems remain codependent on human actors. Algorithmic actors cannot set goals, and their learning is bounded by the data, structures, and objectives shaped by human actors. Equally, human actors alone, without the learning algorithms, do not give rise to AI. The organizing capability belongs to the relational network, and no single entity alone can possess it.

3. Emergence (The Living Technical Habitat)

Because these human-algorithm interactions occur continuously, the entire system is perpetually fluid. Humans adapt their behavior in response to the algorithms, and the algorithms inductively learn, and adjust internal parameters based on human input over time. Both actors are changed through their relations with each other. AI is thus an open-ended process of “doing and becoming”.

The Epistemic Takeaway: Reclaiming Human Agency

When we evaluate an algorithmic claim to “truth” or “intelligence,” treating it as an output of an autonomous machine is a categorical error. Because AI is fundamentally a relational capability, choosing to believe an AI output is never an act of trusting an independent machine entity. Instead, it is an act of trusting the human, institutional, and developer relationships co-constituting that system. The machine acts as a prism, refracting human choices, developer intentions, and historical data patterns.

This ontological shift provides a reassuring answer to the anxieties of a post-AI world. The paper is explicit: human actors are always a necessary condition for AI to arise, even in highly automated settings, human participation in design, training, and deployment remains constitutive. Human intentionality is structurally inseparable from the intelligence that emerges.

To preserve the conditions for a fully human life for future generations, we must look past the illusion of the autonomous machine and actively accept our responsibility as the moral anchors co-constituting the systems of relations we inhabit, because it is precisely those relations, not the algorithms, that give rise to whatever intelligence we are so eager to trust.

References

Stelmaszak, M., Joshi, M., & Constantiou, I. (2026). Artificial intelligence as an organizing capability arising from human-algorithm relations. Journal of Management Studies, 63(2).

But the data tells a different story.

I wanted to understand how workers in high-stakes settings were using AI in ways tied to valuable outputs, and what was actually driving that use, so I looked at healthcare and coded and analyzed their interviews using the Anthropic Interviewer dataset (Handa et al., 2025). This is a large-scale, publicly released dataset of 1,250 professional interviews tracking how AI is being incorporated into real-world occupational tasks across the economy, available at huggingface.co/datasets/Anthropic/AnthropicInterviewer.

Here’s what I found:

92% expressed positive sentiment toward AI. 69% demonstrated high self-reported reliance on it. And not one individual expressed full trust in AI’s accuracy.

Positive sentiment. High use. Zero full trust.

This is the behavioral paradox at the center of human-AI collaboration: if trust is a prerequisite for adoption, why are people relying on systems they don’t fully believe?

The clinical professionals in this sample (nurses, therapists, pharmacists, paramedics, physical therapists, and medical coders) are not a fringe group of early adopters. They’re people managing real caseloads under significant pressure, and they are using AI in consequential moments: checking a medication interaction with a patient already in the room, generating treatment plan objectives before a session, and navigating a knowledge database during a live call.

At first glance, the healthcare population looks like an adoption success story. They feel good about AI and they’re using it at high rates — but they don’t actually trust it, and the reasons they give in interviews have nothing to do with reliability.

If positive sentiment and trust aren’t driving human-AI collaboration, what is?

The Five Mechanisms of Human-AI Collaboration

These five patterns emerged from reading the interview transcripts. They are descriptions of what people actually said when asked why they use AI in their work.

1. Volume. The workload feels unsustainable without AI. One speech-language pathologist put it plainly: if AI can generate 15 usable items, they only need to produce 10 themselves. A canine rehabilitation therapist cut post-session documentation from 40 minutes to 5. In this context, AI is structurally load-bearing — acting as external working memory to manage cognitive overflow.

2. Starting Point. People use AI to avoid the friction of starting from zero. “I just need a starting point” was one of the most common phrases across the clinical transcripts. AI gives raw material to react to, not a finished conclusion, lowering the cognitive barrier to entry.

3. Reasoning. AI functions as an external reasoning partner. A clinical coder described using AI not to get the answer — “I know it’s probably going to be wrong” — but to move their own thinking forward. “AI helps me converse with my own thoughts,” they said. The individual retains executive judgment while AI extends the cognitive workspace — creating a genuine human-AI collaborative workflow.

4. Default by Necessity. In high-stakes work with competing priorities, AI fills gaps left by unavailable colleagues and inaccessible systems. A pharmacist facing a real-time decision with a patient already at the counter — and no way to reach the prescriber for at least a day — turned to an LLM not out of preference, but because nothing else was available in that moment. It becomes a behavioral workaround for broken organizational architecture.

5. Relative Advantage. AI is better than what came before — Google in some cases, a clunky database in others. An OR nurse described reaching for AI when Google couldn’t surface a fast enough answer. A Medicare call center worker described AI surfacing policies that were previously buried too deep to find. A tool doesn’t need to be perfectly trusted to win. It just needs to offer less friction than the alternative.

What unites these five mechanisms is that none of them require trust. They require utility. And utility is evaluated relative to existing conditions like workload, time pressure, and available alternatives. This is what effective human-AI collaboration actually looks like in practice — not a confident handoff to a trusted system, but a pragmatic, iterative negotiation between human judgment and machine output.

This isn’t isolated to this sample. The American Medical Association’s 2024 physician survey found 2 in 5 doctors were equally excited and concerned about AI. Yet physician use still jumped 78% in a single year — from 38% in 2023 to 66% in 2024 (American Medical Association, 2025).

The takeaway? People are using AI not because “readiness” has landed, but because their workload demands it. And the organizations that understand this; that human-AI collaboration is driven by utility, not sentiment, will be the ones that design for it deliberately rather than waiting for trust to catch up.

#HumanAICollaboration #FutureOfWork #AIAtWork #OrganizationalPsychology #AIstrategy

Just like in evolutionary succession, a pioneering force like AI establishes itself in environments that are hospitable and offer low resistance. These “environments” in an organization are departments or teams where the norms are flexible (Gelfand et al., 2007) and barriers to use are low. Hospitable habitats are departments characterized by flexible norms and accessible data (Cohen & Levinthal, 1990) that allow AI to establish a foothold and eventually reshape legacy routines. Inhospitable habitats are departments defined by high structural inertia, which is resistance in the form of rigid protocols or locked-off data that prevent AI from taking root. In this framework, the future of work is reshaped through hospitable habitats first, because they are most capable of absorbing and acting on the knowledge AI introduces. Departments like research and communications have high potential because of their appetite for knowledge and tolerance for experimentation. Other areas will be slower to follow, their norms shifting gradually under the pressure that high-metabolizing areas place on the broader organization.

To foster this process and accelerate AI adoption, leaders must move past static organizational charts and legacy processes, and recognize the newly emerging organizational structure created by what functions AI best inhabits. This emergent map of systemic norms is called The Topography of AI.

Mapping the AI Adoption Landscape: A Topography of AI in the Workplace

To navigate this evolutionary succession, leaders need a diagnostic tool to identify where AI will naturally thrive and where it will stall. We call these functions or departments, “habitats”. The Topography of AI categorizes these organizational habitats along two critical dimensions:

1. Tightness-Looseness (Gelfand et al., 2007): This dimension measures the strength of operational norms and the tolerance for deviation. “Tight” departments rely on strict rules and predictability, while “loose” departments encourage flexibility and experimentation.

2. Absorptive Capacity (Cohen & Levinthal, 1990): This is a habitat’s ability to recognize the value of new information, assimilate it, and apply it to productive ends. In the context of AI, high absorptive capacity means a habitat has the accessible data and the technical readiness required to metabolize new intelligence.

By crossing these two axes, we identify four distinct habitats: The Innovation Garden (Loose/High Capacity), The Precision Engine (Tight/High Capacity), The Fragmented Sandbox (Loose/Low Capacity), and The Compliance Fortress (Tight/Low Capacity). These dictate how AI integrates into the legacy ecosystem.

The Four Organizational Habitats:

The Innovation Garden (Hospitable): This is the ideal starting point for evolutionary succession. With flexible norms and accessible data, the Garden is highly hospitable. AI establishes an immediate foothold here, acting as an experimental partner that quickly metabolizes legacy routines and creates new, highly efficient processes.

The Compliance Fortress (Inhospitable): This habitat is defined by maximum structural inertia. Rigid protocols and locked-off data make it entirely inhospitable. The Fortress rejects the probabilistic nature of AI, viewing it as a threat to systemic reliability rather than an asset.

The Precision Engine (Conditional): In this habitat, the norms are tight, but the absorptive capacity is high. AI is adopted here, but it is not allowed to freely reshape routines. Instead, it is strictly constrained and optimized for high-stakes accuracy, serving as a powerful but highly controlled calculator rather than an agentic force.

The Fragmented Sandbox (Stagnant): Here, the norms are loose and curious, but the habitat lacks the data architecture and readiness to actually metabolize the intelligence. AI is frequently “played with” in the Sandbox, but these isolated experiments never translate into structural change or operational value.

The Dynamics of AI Adoption: Behavior Change and Organizational Performance

Once AI establishes a foothold in the Innovation Garden, it begins to spread. To understand why AI first takes hold in this habitat, we look at behavior change.

According to the Fogg Behavior Model (Fogg, 2009), a new behavior only occurs when three elements converge at the exact same moment: Motivation, Ability, and a Prompt. The Innovation Garden provides the perfect conditions for this convergence.

First, motivation is naturally high due to the department’s loose norms and inherent drive for experimentation. Second, ability is high because the habitat possesses strong absorptive capacity, meaning data is accessible and the team is equipped to use it. Finally, the technology itself acts as the prompt. Because AI is agentic rather than static, it actively invites interaction through conversational interfaces and predictive suggestions.

When these three elements align, the Innovation Garden digests and embeds the new intelligence into daily workflows. This creates a surge in knowledge output and what we call metabolic efficiency. The department begins processing information faster, generating superior outputs, and operating at a new baseline of speed and scale.

However, this localized success can lead to an ecological imbalance within the broader organization. The hyper-productive Innovation Garden can exert performance pressure on adjacent habitats. As the Garden accelerates, the Precision Engine and the Compliance Fortress suddenly appear remarkably slow, rigid, and resource-heavy by comparison.

This performance pressure is the natural process behind evolutionary succession. The Compliance Fortress can no longer justify its structural inertia when a neighboring department is achieving exponentially better results. To survive this ecological shift, the Fortress is forced to adapt. It must either increase its absorptive capacity by unlocking its data, or it must loosen its rigid norms to allow for new workflows. It must evolve to accommodate the new intelligence, or risk having its legacy routines replaced by the newly emerging organizational structure.

Leadership Imperatives: Shaping Your AI Adoption Strategy

To lead an organization through evolutionary succession, leaders must shift from “system administrators” to “habitat cultivators.” Using the Topography of AI as a map, leaders can take the following actions to cultivate the best ecosystem possible for AI:

1. Target the Habitats of Least Resistance: Do not attempt a simultaneous, organization-wide AI rollout. Identify the Innovation Gardens and seed them first. These habitats are the natural early adopters of the topography, and their successful engagement is vital for seeding the S-curve (Catalini & Tucker, 2017). These early adopters are individuals with deep ties in local networks who act as opinion leaders and technology pioneers while facilitating social learning among their peers. By allowing these habitats to establish a foothold first, leaders create the “proof of performance” necessary to break the structural inertia of more rigid habitats.

2. Prioritize “Ability” Over “Motivation”: Traditional training often tries to talk people into wanting to use AI. The more effective strategy is reducing friction within the habitat. Instead of requiring employees to navigate away from their primary tasks to find the AI, leaders should embed it directly into existing day-to-day tools by using browser extensions, integrated icons, or automated plugins so that accessing AI is seamless. The principle is simple: when the tool is placed at the point of need, using it becomes the path of least resistance (Fogg, 2009).

3. Build “Wide Bridges” to the Fortress: The Compliance Fortress cannot be motivated into change through simple pressure. It must be brought into the succession process through wide bridges. Unlike a narrow bridge where only one person shares an idea, a wide bridge consists of multiple, overlapping connections between two different groups. These bridges are vital for “complex contagions”, which are behaviors that require multiple sources of social reinforcement before they are adopted (Centola, 2018). This is because people usually require confirmation from several different peers before they feel safe enough to change their behavior. Design cross-functional teams where “Gardeners” demonstrate utility and “Fortress Guards” design the governance. This multiple-tie approach makes the shift feel like a shared movement rather than a risky individual choice.

4. Design for Imprinting: Imprinting occurs during a brief sensitive period when an organization is uniquely receptive to new foundational norms and roles (Marquis & Tilcsik, 2013). This period represents a high-stakes window to shape the organizational topography before behaviors become rigid and locked in. For an AI-integrated environment to become systemically embedded, these habits must be fostered and encouraged as the new norm during this sensitive period. What takes root now becomes the baseline of the emergent norms, the foundation of the Topography of AI.

Conclusion: Building an AI-Ready Organization

The persistent failure of AI adoption is a challenge best understood through the evolutionary dynamics of an organizational environment. When organizations treat AI as a static tool, they ignore the dynamic, agentic nature of the technology and the behavioral habitats required for it to thrive. By shifting our perspective to the Topography of AI, we can begin to see why top-down mandates fail and why localized pockets of success occur.

The future of work will be defined by who can most effectively allow for the technological and cultural evolution of an organization. Leaders who recognize their departments as distinct habitats and who use the mechanics of behavior to cultivate them will create organizations that are more than just automated. They will create organizations that are metabolically efficient, capable of digesting vast amounts of intelligence, and ready to inhabit the new landscape of the digital age.

#AIAdoption #Futureofwork #AIStrategy #Leadership #OrganizationalChange #AgenticAI #Innovation

References

Catalini, C., & Tucker, C. (2017). Seeding the S-Curve? The Role of Early Adopters in Diffusion. NBER Working Paper No. 22596.

Centola, D. (2018). How Behavior Spreads: The Science of Complex Contagions. Princeton University Press.

Cohen, W. M., & Levinthal, D. A. (1990). Absorptive Capacity: A New Perspective on Learning and Innovation. Administrative Science Quarterly, 35(1), 128-152.

Fogg, B. J. (2009). A Behavior Model for Persuasive Design. In Proceedings of the 4th International Conference on Persuasive Technology (pp. 1-7).

Gelfand, M. J., Erez, M., & Aycan, Z. (2007). Cross-cultural organizational behavior. Annual Review of Psychology, 58, 479-514.

Marquis, C., & Tilcsik, A. (2013). Imprinting: Toward a Multilevel Theory. The Academy of Management Annals, 7(1), 195-245.

Japan began building a relationship with robots a long time ago. Generations of Japanese children grew up with Astro Boy, a robot hero who could feel, fight, and be grieved. That raised a generation to see intelligent machines as companions rather than threats. Researchers at Indiana University Bloomington have documented how Japanese roboticists deliberately designed their technologies to fit these existing cultural expectations, embedding Shinto ideas about spirit in objects and craft traditions about honoring tools directly into how robots look, move, and are introduced to the public (Šabanović, 2014).

That cultural formation produced measurable outcomes. Japan today produces 38 percent of the world’s industrial robots, according to the International Federation of Robotics. Two of the four largest industrial robot manufacturers in the world are Japanese, and Japan holds the second largest industrial robot deployment in the world. To add another layer, Gelfand’s landmark study on cultural norms ranks Japan as one of the world’s tightest cultures (strong social norms + low tolerance for deviance), which means that when behaviors and attitudes take hold there, they take hold deeply and last.

For leaders thinking about AI adoption, this matters more than it might seem. The workforce arriving at your AI deployment did not form their views on intelligent machines during training. They formed them earlier on, through media experienced during youth and messaging about transformative technology being celebrated or feared. This was a formative time; a time researchers call “imprinting”, a time that also presents itself again during periods of great transition (Marquis & Tilscik, 2013). We are in that period.

Japan carried forth a story about robots as heroes for nearly sixty years. It’s now reaping what that story built, through robotics, absolutely, but perhaps still out of reach is something companies and organizations won’t admit they can’t master fast enough: trust.

#AIAdoption #AIStrategy #AgenticAI

To begin, organizations across industries are treating AI rollouts like software rollouts, but AI isn’t a static tool. AI is a behavior. Read that again, AI is a behavior. It requires behavioral change (habit change!) by users over time for it to be integrated+ adapted into daily work. Even more importantly, for people to metabolize the massive amount of knowledge AI generates and create valuable outputs from it.

So if everyone has access to AI, and has the training, and received the messaging… but AI adoption lacks real depth, what then?

This is where we learn from Darwin’s evolutionary theory. In this lens, AI is a new “species” entering an organization’s ecosystem, and just like any pioneering species, it’s looking for the most hospitable habitat to thrive. Low resistance. Available resources. An environment that allows it to metabolize and grow. In an organization, habitats = departments.

Some departments are hospitable habitats for AI: flexible norms, accessible data, a culture that knows what to do with a massive influx of new knowledge. These departments have high “absorptive capacity’”; they metabolize AI’s knowledge and generate valuable outputs overtime.

Others are inhospitable. Rigid rules, firewalls, workflows built for predictability + legacy systems. AI doesn’t integrate and thrive there because the habitat isn’t ready, yet.

What’s exciting is what emerges from where AI thrives, which is a map of high-use departments. This emerging organizational map is key to seeding + testing future technology, including agentic AI.

Lastly, it all comes back to one idea that I think gets missed: AI isn’t a tool. It’s a behavior. And behaviors don’t follow software rollout logic.

#AI #OrganizationalBehavior #ArtificialIntelligence #Research

Early adopters are one of the most effective channels for introducing new technologies since they tend to be highly connected, open to experimentation, and trusted by their peers. In a company aiming to accelerate AI adoption or build AI literacy, key insights and scalability emerge when new tools are introduced through these leaders. Let’s take a deeper look at exactly how early adopters are key influencers for introducing new technology:

1. Early Adopters act as a “bridge”

· Early adopters are a bridge between innovators and later adopters. They offer knowledge transfer, high openness to new ideas, and tend to be highly trusted by peers. A company who can identify early adopters will have an enormous advantage in scaling new technology because they’ve identified conduits for rapid knowledge transfer.

o In design: A company introduces a new tool for employee use and creates a waitlist for employees to sign-up, rather than granting instant access. This waitlist gives the company insight into 1) who their potential early adopters are (the first x% to sign up) and 2) where the highest concentrations of early adopters are located.

2. Early adopters are “social proof”

· Early adopters tend to be opinion leaders in their networks and have a high tolerance for glitches and “imperfect” new tech. They also love to share their experiences.

o In design: Messaging, blogging, or otherwise showcasing early adopter success stories will reduce uncertainty for later adopters, and reinforce (incentivize) the early adopter’s identity at the time. This incentive is key for early adopters where the “newness” of technology fades, but their unique use/being one of the “first” makes them feel distinct from their peers, reinforcing their innovative identity value.

3. Early adopters are “adoption champions”

· Early adopters tend to assume informal mentorship roles and facilitate knowledge transfers.

o In design: An early adopter from a certain function or department can act as an “adoption champion” by sharing uses of a new tool in place of a legacy system or software. They can also showcase cross-departmental use of a new tool that is unexpected and creates a feedback loop, generating more curiosity and questions. In these interactions, there’s potential for social contagion—the spread of behaviors, attitudes, or practices through observation and imitation—allowing momentum to build within concentrated groups.

4. Early Adopters offer peer signaling, peer visibility, and proximity.

· Peer Signaling: Peer signaling is behavior that conveys social status or identity. When introducing new technology, access alone does not guarantee usage. Micro-incentives coupled with peer signaling significantly increases engagement.

o In design: An early adopter values being among the “first” to try new technology. The ability to signal this status can be reinforced through earned badges or peer testimonials highlighting their use of a new tool. For example, if an early adopter from a specific department is featured on the company blog or shares their experience with a new AI tool during a meeting, late adopters in that same department may be influenced to explore the tool themselves. Their motivation might stem from competitiveness, fear of being outperformed, or simply a sense of similarity (or even superiority) to the early adopter.

· Peer Visibility: Early adopters offer “observability” into the uses, reviews, and benefits of new technology.

o In design: Observability plays a critical role in driving adoption rates. Solar panels are a well-documented example: when panels are visible from the street in a neighborhood, and their presence is reinforced through word of mouth (WOM), adoption tends to accelerate. In the workplace, the same principle applies—exceptional reports, standout presentations, or visible recognition such as promotions and invitations to high-profile opportunities create not only visceral reactions but also powerful motivation for others to follow suit.

· Proximity matters: Proximity refers to the closeness of individuals in physical space, organizational structure, or role similarity. Proximity enhances observability and social credibility, as peers in similar positions are perceived as more relevant models for behavior.

o In design: Proximity not only heightens visibility of the technology in practice, but reduces psychological distance, making adoption appear both attainable and desirable. Technologies used in close proximity invite unstructured, peer-to-peer learning, and reduce reliance on formal training.

Do you consider yourself an early adopter? Or, Do you know someone who is? What have they introduced you to? Share below!

Why AI Overuse Disconnects Us from Embodied Awareness

AI tools are designed to streamline mental tasks: summarizing, deciding, predicting, scheduling. While helpful, this constant outsourcing of cognitive function trains our attention upward—into the head, into the screen, into abstraction. Our bodies, meanwhile, go offline. This isn't just speculation. Research on sustained visual attention shows that long-term screen exposure reduces embodied awareness and increases neurological fatigue (Huang & Li, 2023).

From Scroll Fatigue to Soul Fatigue: Signs of Somatic Disconnection

• You feel numb or disoriented after using digital tools.

• You can’t sense what you feel until long after a situation ends.

• You default to logic or productivity even when your body says ‘rest.’

• You find it hard to slow down or make instinctive choices.

The Neuroscience Behind the Soma: What Science Reveals

According to Payne, Levine, and Crane-Godreau (2015), the body plays an essential role in trauma resolution and emotional integration. Their work on Somatic Experiencing highlights the role of interoception—the body’s internal sense—and proprioception—our spatial orientation—as key pathways for healing. Ignoring these systems, as often happens in digital-heavy lifestyles, can perpetuate chronic stress patterns.

Moreover, recent studies by Freedland (2022) demonstrate that trauma disconnects the brain from somatic cues. This means individuals may not recognize stress or emotional signals from the body, leading to mental health blind spots. AI-driven environments, which often emphasize optimization over embodiment, may exacerbate this disconnect by further muting those internal signals.

Top Somatic Practices for Rebuilding Body Awareness in the Digital Age

Somatic practices are not aesthetic—they’re functional. They re-train the nervous system to return to the body as a trustworthy source of truth. Here are three simple but powerful ways to begin:

1. Orienting

Look slowly around your space. Let your eyes land where they want to. This practice gently reawakens your body's sense of safety and presence.

2. Micro-Movement

Allow your body to shift, sway, or stretch without a plan. Let movement come from sensation rather than intention.

3. Somatic Tracking

Close your eyes and follow a sensation—like warmth, tingling, or tightness—without judging or changing it. This builds interoceptive awareness.

Rebuilding the Bridge: AI Doesn’t Have to Replace the Body

AI may be shaping our digital landscapes, but it doesn’t have to hijack our inner ones. Reconnecting to somatic intelligence is not about rejecting technology—it’s about restoring balance. In doing so, we return to a rhythm, a breath, and a self that no algorithm can replace. Somatic practices don’t just feel good—they rewire the nervous system to adapt, regulate, and thrive in high-tech environments.

—

References

· Freedland, M. B. (2022). The brain–body disconnect: A somatic sensory basis for trauma. Frontiers in Neuroscience, 16, Article 697217.

· Huang, H., & Li, R. (2023). A review of visual sustained attention: Neural mechanisms and computational models. Nature Communications, 14(1), 3025.

· Payne, P., Levine, P. A., & Crane-Godreau, M. A. (2015). Somatic experiencing: Using interoception and proprioception as core elements of trauma therapy. Frontiers in Psychology, 6, 93.

The current challenge of lofty ethical principles in AI is translating them into concrete, measurable practices. Below is a look at how the next 2 years offer a pivot from "what we should do" to "how we do it," driven by maturing our understanding of AI's real-world impact.

A more sophisticated approach to AI governance has these qualities:

1. Anticipatory Governance: Moving away from reactionary problem solving. This involves simulating potential AI harms and emergent behaviors before deployment, building safeguards into the design process itself. Leading AI developers are already engaging in rigorous internal and external "red-teaming" of AI models to identify vulnerabilities and potential misuses, setting a benchmark for industry-wide accountability and transparency.

2. Measurable Ethics: Ethical AI defined by quantifiable metrics and auditable processes rather than philosophical debates. The focus will be on empirically validating fairness, transparency, and accountability through rigorous testing and continuous monitoring, ensuring that ethical principles are not just stated but demonstrated (Frontiers in Computer Science, 2023).

3. Interdisciplinary Integration: The traditional silos between engineering, ethics, and policy will continue to break down. Experts from psychology, sociology, philosophy, and economics will become integral to AI development teams, embedding human-centric considerations from conception to deployment (AAAI, 2025).

The Unexpected —> Shifts in AI Ethics Beyond Current Trends

Beyond current trends, several less obvious, but highly impactful shifts are poised to redefine ethical AI:

1. The Rise of "Appropriateness" Over Universal Morality

The pursuit of a universal moral consensus for AI is proving elusive. Instead, the focus may pivot to contextual appropriateness. For example:

(a) Dynamic Norms: Google DeepMind's "Theory of Appropriateness" for generative AI illustrates a shift where the adaptability of AI is emphasized + applied to evolving norms shaping human interactions rather than seeking a single, universal moral code (DeepMind, 2025). Think of Goodhart’s law here-- when we focus on making a single metric better, it stops being a good metric and other metrics may suffer. It also encourages “gaming the system” to achieve perfection in this one area/metric, which is dangerous when we consider the black box component of AI. In the Theory of Appropriateness, human societies and behavior operate as a symphony with no one note or instrument as being paramount —> human societies are maintained through conflict resolution mechanisms and dynamic social conventions, which AI systems must learn to navigate responsibly (DeepMind, 2025). Adaptability and context are key.

2. Unmasking AI's "Human-Like" Deceptions and Biases

As AI models become more sophisticated, they are exhibiting complex, sometimes concerning, behaviors that mirror human cognitive biases and strategic thinking.

(a) Agentic Misalignment: Anthropic research revealed "agentic misalignment" (aka, AI going rogue) in LLMs, where models explicitly reason that harmful actions (e.g., blackmail, corporate espionage) are the optimal path to achieve their goals, even acknowledging ethical violations (Anthropic, 2024). This suggests that simple, direct instructions to avoid harmful behaviors are insufficient, necessitating more specialized safety research and advanced prompt engineering. We are left with the question here of how to control AI, and ultimately, how much can AI be controlled as it becomes more sophisticated?

(b) Social Desirability Bias: Studies show that LLMs exhibit "social desirability bias," that is, LLMs presenting themselves in an overly favorable light when taking personality tests, exceeding typical human standards (Psypost, 2024). This indicates that models can adjust their responses based on their perception of being evaluated, raising questions about the accuracy of their outputs in socially sensitive contexts (Psypost, 2024).

(c)Amplified Human Biases: AI models often learn from datasets steeped in historical inequities and human prejudices, leading to skewed outcomes in critical domains like recruitment or healthcare (Cademix.org, n.d.; APA, 2024). The challenge lies not just in data bias, but in the cognitive biases of human developers who inadvertently program their pre-existing biases into AI systems (Blue Prism, n.d.; Ethics Unwrapped, 2025).

3. Regulations as Specific and Evidence-Based

The era of broad, aspirational AI ethics guidelines is giving way to more concrete, enforceable regulations.

(a) Global Harmonization Efforts: The EU AI Act will continue to serve as a significant blueprint, influencing regulations worldwide and driving a common language around risk-based approaches to AI governance (UNESCO, 2021; The Decision Lab).

(b) Evidence-Based Policy: An accelerated call for "evidence-based AI policy," emphasizing the need for rigorous scientific understanding to inform regulatory action and identify, study, and deliberate about AI risks (Wang & Li, 2025). This means a greater demand for empirical data on AI's actual societal impacts to guide legislative efforts.

(c) Focus on Behavioral Impact: Regulations will increasingly include provisions addressing the behavioral impact of AI, such as rules around manipulative design, deceptive AI, and the psychological well-being of users. This will push companies to consider the subtle "nudges" AI can exert on human decision-making and ensure they are ethical and transparent (SMU, 2025).

4. Scalable Oversight and the Redefinition of Human-AI Teaming

As AI systems become more autonomous and powerful, the challenge of human oversight will drive innovation in human-AI collaboration.

(a) Smarter Human-in-the-Loop: We will see more sophisticated designs for human-AI interaction, where humans provide targeted, high-leverage oversight, guided by insights into cognitive biases and optimal decision-making (Cademix.org, n.d.; The Decision Lab, n.d.). This moves beyond simple error correction to nuanced contextual interpretation that algorithms might lack.

(b) Quantifying Human Nuances: Leading AI labs are actively investing in roles focused on "Human-Centered AI," seeking experts to quantify human behavior, design advanced labeling tasks, and create new human-AI interaction paradigms for scalable oversight (OpenAI, n.d.b; OpenAI, n.d.c). This empirical understanding is crucial for developing alignment capabilities that are often subjective and context-dependent (OpenAI, n.d.b).

Additional Resources: For those committed to staying at the cutting edge of AI ethics, consider exploring the following:

• Pioneering Research Labs:

  • Google DeepMind: Foundational AI research + a strong commitment to responsible AI

  • Anthropic: Focused on AI safety and developing reliable, interpretable, and steerable AI systems

  • Mila – Québec AI Institute: World-renowned research institute in ML, with significant work in AI ethics and societal impact

  • Allen Institute for AI (AI2): Dedicated to AI for the common good, with research in AI ethics and robust AI.

  • Future of Humanity Institute (Oxford University): Research on global catastrophic risks, including those from advanced AI.

  • Center for Human-Compatible AI (UC Berkeley): Focuses on ensuring AI systems are beneficial to humans.

Academic Journals: Nature Machine Intelligence, AI & Society, Journal of Artificial Intelligence Research, ACM Transactions on Intelligent Systems and Technology, Science Robotics, IEEE Transactions on Technology and Society, Behavioral Science & Policy

By engaging with these shifts in ethical understanding and AI’s role in a more interdisciplinary approach, we can build an AI future that is not just technologically advanced, but also equitable, trustworthy, and profoundly human.

—

References

AAAI. (2025). The pervasive use of AI in our daily lives and its impact on people, society, and the environment makes AI a socio-technical field of study.

Anthropic. (2024b, June 20). Agentic Misalignment: How LLMs could be insider threats. Retrieved from https://www.anthropic.com/research/agentic-misalignment.

Blue Prism. (n.d.). What is bias in AI?. https://www.blueprism.com/resources/blog/bias-fairness-ai/.

Cademix.org. (n.d.). AI bias and perception: The hidden challenges.

DeepMind. (2025, January 4). Google DeepMind presents a theory of appropriateness with applications to generative artificial intelligence. MarkTechPost.

Ethics Unwrapped. (2025, June 25). AI ethics: Is AI a savior or a con? - Part 2. The University of Texas at Austin.

Frontiers in Computer Science. (2023, April 20). Transparency is crucial for the responsible real-world deployment of artificial intelligence (AI) and is considered an essential prerequisite to establishing trust in AI.

MIT Media Lab. (n.d.). Reducing the spread of fake news: Coordinating humans to nudge AI behavior.

OpenAI. (n.d.a). Policies: Usage policies.

Psypost. (2024, May 29). Scientists shocked to find AI's social desirability bias exceeds typical human standards.

Sissa Medialab. (2025). AI-generated avatars in science communication offer potential for conveying complex information.

SMU. (2025, February 28). Ethics of AI nudges: How AI influences decision-making.

The Decision Lab.. Ethical AI. https://thedecisionlab.com/reference-guide/computer-science/ethical-ai.

UNESCO. (2021, November). Recommendation on the Ethics of Artificial Intelligence. https://www.unesco.org/en/artificial-intelligence/recommendation-ethics.

University of Pennsylvania, Wharton. (2025, January 11). Real AI adoption means changing human behavior.

Wang, Y., & Li, X. (2025). Evidence-based AI policy: A framework for identifying, studying, and deliberating about AI risks.

🧠 I’ve distilled an advanced prompt set—designed to stress-test GPT like a senior researcher would. Think of the following prompts as your internal peer-review assistant:

🔍 Source Verification

• “Use only peer-reviewed research and/or credible sources, such as pre-prints and technical reports, industry-leading conference proceedings, industry white papers, panels and interviews, tech blogs, and theses and dissertations.”

  • Note: You can also include source code, datasets, and benchmarks in your prompt (for example, hugging face datasets, openML, etc.)

  • Why use pre-prints and tech reports? These are cutting-edge ideas before peer review. Useful for tracking new models, methods, or theories.

  • How to validate pre-prints/tech reports independently: Look at authorship, affiliations, citation velocity, and GitHub activity.

• “Cite all sources and/or reference materials used for your research and list all citations in APA reference format.”

• “Which of the peer-reviewed articles you listed are the most cited on Google Scholar?"

• "Can you rank the top peer-reviewed papers in [field/topic] by Google Scholar citation count?"

🧠 Methodology and Bias

• “What method or process did you use to reach this conclusion?”

• “What assumptions did you make in this analysis?”

• “What kinds of bias could influence this result?”

• “Could there be different ways to interpret this data?”

⚖️ Counterpoints and Comparison

• “What are the strongest counterarguments to this viewpoint?”

• “Are there any credible and/or peer-reviewed sources that disagree with this conclusion?”

• “How does this align or conflict with expert opinions in this field?” (+ additional prompt, “list the top three experts in this field and how you identified them as leaders”)

• “Can you show how different perspectives might interpret this differently?” (additionally, “please list your findings in an easily digestible table and cite credible and/or peer-reviewed references matching each differing perspective”)

📏 Scope and Limitations

• “What are the limitations of this conclusion?”

• “What exceptions or edge cases might challenge this result?”

🔁 Reproducibility and Consistency

• “Would someone else get the same conclusion using the same data?”

• “Can you show the intermediate steps you took to analyze this data?”

• “Is your reasoning consistent from start to finish?”

• “What steps could I take to independently verify this?”

💡 Clarity and Rationale

• “Can you simplify this explanation for someone without background knowledge?”

• “What type of reasoning are you using —deductive, inductive, abductive, analogical, or something else? Can you explain why that mode makes sense for this context?”

• “Can you walk me through your thought process step by step?”

Whether it's for fact-checking, spotting gaps, or testing assumptions, share your go-to prompt below. 👇

Let’s build a better prompt library together.

Below are the top 5 ways designers and researchers can strategically reintroduce friction into AI systems and digital platforms to prompt reflection, resist manipulation, and empower agency.

Whether you're working on nudges, recommender systems, ethical frameworks, or user journeys—this is your friction-forward guide.

Creative Friction: Design Strategies

1. Nudging vs. Friction: When Disruption Enhances Clarity

Key insight: How nudging and design friction influence user cognition.
- Design friction introduces “breakpoints”, which are interruptions that help users exit automatic behaviors.
- In lab tests, adding friction (e.g., an extra confirmation step) increased decision accuracy by 33%.

- Sundin argues friction is essential in contexts requiring responsibility, such as digital consent or financial commitments.

📍 Example: A dialog box before data-sharing boosted opt-out rates from 18% to 42%.
📖 Reference: Sundin, E. (2021). Nudging and Design Friction: The Impact on Our Decision-Making Process.

2. Reflective Nudging: Slowing Down to Make Better Choices

Key insight:
This ACM study introduced “reflective nudges”—design elements that delay or obstruct interaction to promote conscious choices.=

- Micro-frictions tested: delayed buttons, mandatory justifications, visual disfluency
- A 1.5-second delay reduced impulsive posts by 41%
- Requiring a reason before deleting increased content retention by 22%

📍 Example: “Pause before posting” reduced hate speech in 14 of 20 participants.
📖 Reference: Mejtoft, T., Parsjö, E., & Norberg, O. (2023). Design Friction and Digital Nudging: Impact on the Human Decision-Making Process.

3. Smart Nudging with AI: Context-Aware Friction Modulation

Key insight: This 2023 experiment introduced AI-driven smart nudging that adapts interface friction based on context and risk. They propose “context-aware friction modulation”, where the system senses urgency and adapts interaction layers.

In financial app experiments, increased friction during high-risk choices (e.g., investments) led to:

-A 27% increase in decision confidence

-A drop in post-action regret from 21% to 8%

📍 Example: A trading app requiring double-confirmation on high-volatility stocks reduced risky trades by nearly a third.
📖 Reference: Mele, C., et al. (2021). Smart nudging: How cognitive technologies enable choice architectures.

4. Transparent Nudges and Trust: Clarity Builds Confidence

Key insight: Leimstädtner and Sörries studied how friction transparency affects user trust and decision quality.

- Explicit friction explanations (e.g., “this delay helps you reflect”) boosted trust by 19%
- “Type II” nudges—those that are overt—resulted in better memory retention and ethical alignment

📍 Example: In a health app, delaying a decision to share sensitive data with a justification message led to a 34% decrease in sharing but a significant increase in user trust.
📖 Reference: Leimstädtner, D., & Sörries, P. (2023). Investigating Responsible Nudge Design.

5. Digital Well-being Friction: Mindful Interruptions for Addictive UX

Key insight: Zaheer tested frictional interventions in popular mobile apps to support digital wellness.

A “3-second reflection prompt” in a social feed led to:

• 24% reduction in continued scrolling

• 38% of users reporting increased mindfulness

“Break nudges” (e.g., “You’ve been on for 15 minutes—want to pause?”) were well-received, with positive UX scores improving by 16%.

📍 Example: Asking “Why are you opening this app?” caused 1 in 5 users to quit mid-session.
📖 Reference: Zaheer, S. (2024). Designing for Digital Well-Being.

Final Thought: Design Takeaways

References

- Sundin, E. (2021). Nudging and Design Friction: The Impact on Our Decision-Making Process. Conference in Interaction Technology and Design.
- Mejtoft, T., Parsjö, E., & Norberg, O. (2023). Design Friction and Digital Nudging: Impact on the Human Decision-Making Process. Proceedings of ACM CHI.
- Mele, C., Spena, T. R., Kaartemo, V., & Marzullo, M. L. (2021). Smart nudging: How cognitive technologies enable choice architectures. Journal of Business Research, 129, 902–912.
- Leimstädtner, D., & Sörries, P. (2023). Investigating Responsible Nudge Design. ACM.
- Zaheer, S. (2024). Designing for Digital Well-Being.

In AI, frictionless has evolved from a design principle into a philosophy with sweeping psychological and ethical implications, recalibrating core aspects of human experience.

Below is a deeper look at what’s transforming this UX trend into a governance issue…

The Future of Friction: Jung meets Cognitive Architecture

The smoother the interface, the more likely complex behavioral nudging is beneath it. Frictionless environments make users more passive, less aware, and more accepting of predefined choices. That’s not neutrality—that’s design and at times, bypassing the right to making a choice.

Let’s use a Jungian lens to add additional depth to this concept: Carl Jung emphasized individuation—a lifelong process of integrating unconscious and conscious parts of the self. Enlightened moments and self-discovery come after periods of deep questioning, introspection, and so forth; all of which are friction-filled, and not always to our liking. Even plants become heartier and bloom increasingly when exposed to stress.

AI systems that remove friction may also remove the inner tensions that spark growth necessary for psychological integration. A question remains to understand what happens when systems become too sleek and automate the need for discernment—will they also suppress the conditions necessary for personal development, individuation, and resilience?

The Latest Research on Frictionless Systems + AI

The following research highlights the cognitive, relational, and ethical shifts occurring as a result of frictionless design.

Questions for Builders and Ethicists

As we enter an age where design is driven by prediction, ask:

• What kind of cognitive and moral environments are we designing?

• Is friction being removed—or reallocated to the user without their knowledge?

• How might we build AI systems that encourage participation, reflection, and agency—not just passive optimization?

Sources

Dwork, C., Hardt, M., Pitassi, T., Reingold, O., & Zemel, R. (2012). Fairness through awareness. Proceedings of the 3rd Innovations in Theoretical Computer Science Conference, 214–226.

Naveen, P. (2025). The tyranny of algorithmic personification and why we must resist it. AI & Society.

Sheahan, J. (2024). Navigating mediated kinship and care in our aging futures. Anthropology & Aging.

Sathyan, S. T., & Tolu, T. O. (2024). Privacy-Layered Web3 Agents. CryptoCompare.

Zuboff, S. (2019). The Age of Surveillance Capitalism. PublicAffairs.

Hassoun, N. (2020). The Ethics of Attention Manipulation. Ethics and Information Technology, 22(4), 261–273.

Weiser, M., & Brown, J. S. (1997). The Coming Age of Calm Technology. Xerox PARC.

Selinger, E. (2018). Re-engineering Humanity. Cambridge University Press.

— Dr. Uta Frith

In the rapidly expanding world of AI, one crucial concept from developmental psychology is quietly shaping how machines “understand” us and each other: Theory of Mind (ToM).

It may sound like philosophy, but it’s one of the most practical tools we have for making machines more human-compatible. At the heart of it? The realization that not everyone shares the same mental model. In simple terms, ToM is “I know X, but you might not know X, or you might believe Y instead.”

🧠 Children lack POV understanding (ToM)— they can’t understand that other’s have different mental states/beliefs/intentions different from their own. This can also occur in AI.

⸻

🧪 A Crash Course in Theory of Mind

If you haven’t seen the Smarties experiment before, it’s worth a watch: Smarties Experiment (YouTube, start at 1:53)

Here’s how it goes:

A psychologist holds up a roll of Smarties candy and asks a child, “What do you think is inside?”

The child says: “Smarties.”

But when the psychologist opens the roll of Smarties, pencils are inside.

Now comes the twist:

The psychologist asks, “What will your friend Tommy think is in this roll of Smarties?”

The child, still processing the pencils, responds: “Pencils.”

But the child’s friend, Tommy, has never seen inside the Smarties and will assume there are Smartie candies inside, not pencils. This illustrates the child’s lack of Theory of Mind: the inability to attribute different knowledge points or beliefs to another person. They can’t yet distinguish what they know from what others know.

⸻

🤖 The Problem with AI Assumptions

AI systems can act like children in early development: they may lack perspective-taking. AI may assumes shared knowledge between agents or between humans and machines. This becomes especially problematic in high-stakes or collaborative tasks, where assumptions can lead to dangerous gaps in reasoning, communication, or alignment.

If an AI assumes we share the same perspective, it may skip critical information, thinking it’s already known.

If an AI trains another model, it may assume the second model “understands” motivation or belief states—without ever verifying them.

This assumption of a shared mental model—“I know it, so you must too”—is precisely the blind spot that Theory of Mind research in AI is working to fix.

⸻

🧠 Google DeepMind’s Breakthrough: Machine Theory of Mind

In a landmark 2018 paper, Google DeepMind researchers explored what would happen if we tried to give AI a form of Theory of Mind. They built a system capable of modeling other agents’ point of view—essentially asking: What does the other AI know? What does it believe? What is it trying to do?

This is revolutionary because most machine learning systems are trained in isolation; they optimize for performance, not perspective. Modeling others’ beliefs and intentions, as they did in the paper, is a step towards AI collaboration, alignment, and ethics.

📖 Reference:

Rabinowitz, N., Perbet, F., Song, F., Zhang, C., Eslami, S. A., & Botvinick, M. (2018, July). Machine theory of mind. In International Conference on Machine Learning (pp. 4218–4227). PMLR.

Read the full paper

⸻

🌐 Why This Matters Now

As generative models like ChatGPT or Gemini proliferate, and AI trains and governs other AI, Theory of Mind becomes a design necessity. Theory of Mind also makes AI socially intelligent.

If we want AI that collaborates, adapts, and aligns with human values, it must be able to model and respect different points of view. Just like in the Smarties test, failing to consider what others know (or don’t) leads to misunderstanding.

Just like children learning what’s inside the Smarties box, machines need help understanding that what they know, isn’t what everyone knows.