Why the training model is failing, what learning science already knew, and what replaces competence

Dr. Darcell L. Streeter

The organization had done everything the playbook prescribed. Licenses procured. Workshops delivered. A prompt library built and circulated. An internal champion named in every department.

Eight months later, most of the seats were still cold.

The leadership read this as resistance. That reading is nearly universal right now, and it is a category error with expensive consequences, because resistance and adaptation call for opposite interventions. Resistance is met with pressure. Adaptation is met with conditions. Apply pressure to a person who is adapting and you do not accelerate them. You push them into the worst available outcome, and you do it while congratulating yourself on your change management.

I have had a version of this conversation in a dozen rooms over the past two years, most often with learning and development specialists, consultants, and executives who are earnest, well resourced, and stuck. What is happening in those rooms is not refusal. It is acculturation, and it follows a pattern that has been documented in the research literature for nearly forty years.

A framework built for a different question

In 1997, cross-cultural psychologist John Berry published a synthesis describing how people navigate sustained contact with a culture not their own. His subject was immigration. I want to name that plainly, because I am about to borrow his framework for something else, and I am not equating the two experiences. An executive learning to work alongside a language model is not an immigrant. The stakes, the histories, and the consequences are not comparable, and I will not pretend otherwise.

What transfers is narrower and more useful than an analogy. It is a structural insight: when a person encounters a system that operates by rules they did not author, they must resolve two independent questions. Do I preserve what I already am? And do I participate in what is being asked of me?

Those two questions produce four positions, not one spectrum. That distinction is the entire argument, because a spectrum implies that everyone is somewhere on a line between refusing and adopting, and that the job of leadership is to move people rightward. Four positions means some of your people are not behind. They are somewhere else entirely.

Assimilation. Participate fully, preserve nothing. This is the worker who routes every judgment through the model and stops forming independent positions. Organizations frequently mistake this person for their success story. Their output volume rises. Their discernment does not.

Separation. Preserve fully, participate not at all. The practitioner who protects the integrity of their craft by refusing the tool outright. Usually the most experienced person in the room, and usually the one whose refusal is treated as obstruction rather than read as information.

Marginalization. Preserve nothing, participate not at all. This person has already absorbed the message that their expertise is depreciating, but has not acquired fluency to replace it. They are not resisting. They have gone quiet. This is the position organizations cannot see, because it does not generate a complaint, an escalation, or a help ticket. It generates attrition eighteen months later that gets attributed to something else.

Integration. Preserve and participate. Full fluency with the tool, judgment fully intact, and the discipline to know which of the two is answering at any given moment.

What the evidence says about which one you get

Berry and colleagues studied 5,366 immigrant youth across thirteen countries. Their cluster analysis produced four adaptation profiles, and the ordering was consistent: youth with an integration profile showed the best psychological and sociocultural adaptation, while those with a diffuse profile, the analogue of marginalization, showed the worst (Berry, Phinney, Sam, & Vedder, 2006).

Notice what that finding does not say. It does not say the people who participate most do best. Assimilation is maximum participation, and it does not win. What wins is the strategy that refuses the trade.

Now watch what happens when you look for that pattern in the workplace.

In March and April of 2026, SHRM surveyed 5,875 U.S. workers about AI at work. Buried in the section on engagement is a regression result the authors did not frame this way. Workers reported higher engagement when their organizations took an open approach to AI integration, built trust, and placed high value on critical thinking, and when workers relied less on AI to replace their own skills. Organizational commitment showed the same shape: greater openness to AI, and less reliance on it, appearing together.

Openness and non-dependence. Both at once. That is not a spectrum position. That is integration, measured in a workforce sample, by researchers who were not using the word.

The quadrant nobody is looking at is not where they think

Leadership assumes assimilation is the junior employee pasting model output into a deck. The data says otherwise.

SHRM asked workers what they would do if an AI recommendation conflicted with what they believed was the most ethical choice. Among directors and above, 55% said they would follow the AI. Only 45% said they would trust their own judgment. Individual contributors were more likely than their executives to hold their own line.

These are the same senior leaders who, by their own report, place the highest value of anyone in the organization on critical thinking when using AI.

A related finding explains the mechanism. Workers who scored highest on SHRM’s AI aptitude assessment reported significantly lower reliance on the tool than workers in the medium-low group, and rated critical thinking as more important. Understanding the system produces caution. Not understanding it produces deference.

Which means the organization’s most confident adopters may be its most assimilated, and it has no instrument pointed at them.

Meanwhile, the quiet ones

Marginalization does not announce itself, so I want to be careful about what can and cannot be claimed. No survey sorts workers into these four positions, and I am not going to pretend one does. What the data shows is that the conditions that produce this position are widespread.

Thirty-four percent of workers use no AI tools at all. Only a third of individual contributors recall hearing anything from leadership before AI arrived, against nearly three-quarters of directors. Trust in senior leadership on AI falls to 47% among individual contributors, and another 38% say they simply do not know whether they trust it.

That 38% is the group I want you to hold. They are not resisting. They have not committed. They have not complained, which is why they do not appear in any dashboard.

And here is the finding that should reframe how you read their silence. Across seven academic fields, workers rated their own field an average of 14 percentage points more likely to be made redundant by AI than workers who had not studied it did, with statistically significant differences in five of seven fields. Proximity to a body of knowledge increases the belief that it is being erased. The people who know the work best are the most convinced it is going away.

That is what expertise depreciation looks like before anyone quits.

Which brings us back to conditions

Berry’s framework is explicit that integration cannot be pursued unilaterally. It requires a receiving environment that permits a person to hold onto something while joining something (Berry, 1997). Where the surrounding institution demands that people shed what they were in order to belong, integration is not available. The individual does not get to choose it.

SHRM’s own conclusion arrives at the same place from the opposite direction: AI culture matters as much as AI adoption. Trust, openness, and a stated organizational value on human judgment predicted engagement and commitment. Mandates did not appear anywhere in that model.

One vendor-sponsored survey this spring, conducted by the AI platform company WRITER among employees already using AI, found 60% of companies planning to lay off workers who cannot or will not adopt. Treat the number with the caution its sponsorship deserves. Treat the strategy it describes with more.

Because an organization running that play has not motivated integration. It has eliminated it, and left its people to sort themselves among the three that remain.

Most will not choose separation. Separation requires enough standing to absorb the cost of saying no.

We Are Teaching It Like a Tool. They Are Learning It Like a Culture.

The first part of this argument held that AI adoption follows the structure of acculturation rather than resistance. This part takes up the harder question. Even where conditions permit integration, the training almost never produces it. That is not a failure of budget or effort. It is a design failure, and learning science named its component parts decades ago.

In Affirming Diversity, Sonia Nieto and Patty Bode examine what they call cultural discontinuities and school achievement: the mismatches that cause capable students to disengage from institutions that were built without them in mind. They treat discontinuities in language, and discontinuities in perspectives on achievement, as forces that suppress performance in students who are not, in fact, less able.

I extend the concept to five domains, because the mismatches now appearing in AI adoption are not confined to those two. The value of naming them separately is that it stops the conversation from collapsing into “they need more training.” Each discontinuity has a different cause and a different remedy.

I will use self-determination theory to explain why each one costs you, rather than simply naming it. Deci and Ryan’s account holds that motivation depends on three psychological needs: autonomy, competence, and relatedness. Their further contribution, less often cited in business settings, is the internalization continuum. Behavior can be externally regulated by reward or threat, introjected as internal pressure and anxiety, identified as personally valuable, or fully integrated into how someone sees themselves. Only the last two survive contact with difficulty (Ryan & Deci, 2000).

Most AI adoption programs are operating in the first two and cannot tell the difference.

The pedagogical discontinuity

A mismatch between how something is taught and how people actually learn it.

Nearly 40% of workers report their employer provided workshops on practical day-to-day AI skills, making the workshop the default intervention. But when workers rated effectiveness, the strategies that performed best were monetary incentives at 64%, multiple training sessions at 63%, and competitions at 62%. The single most common delivery method is the one-time session. The thing that works is repetition over time.

Gallup’s productivity data explains why. Among employees using AI for one or two purposes, 45% reported a positive effect on productivity. That rises to 66% at three or four purposes, 78% at five or six, and 90% at seven or more. Gallup is careful to note this does not prove that adding applications causes gains, since workers who find AI valuable may seek out more uses.

Read as learning science rather than as a productivity claim, the gradient is telling you something specific: the return does not come from acquiring the tool. It comes from repeatedly re-encountering it inside different parts of real work.

That is what situated cognition has held since the late 1980s. Lave and Wenger’s argument was that competence does not travel as a portable packet installed in a person. It develops inside the practices, tools, and relationships of a working community. Gee made the linguistic version of the same point: fluency lives in a discourse, in shared ways of talking, valuing, and acting that mark someone as a recognized participant.

You cannot deliver a discourse in a ninety-minute session. You can only introduce someone to one.

The linguistic discontinuity

Prompting is not a button. It is a register, and it is being learned in public.

Here is the condition I want to describe, because it is not confined to any one group. An adult is asked to perform, in front of people who can evaluate them, in a form of language they have not yet mastered, where every error is visible and legible as a statement about their intelligence.

That condition sits in a beginning English class, where a woman who managed a business in her first language is being asked to introduce herself in fragments. It also sits in a strategy meeting, where a director with twenty-five years of expertise types a prompt, gets something thin, and quietly closes the window rather than let anyone watch the second attempt.

The vocabulary differs. The physiology does not. Both are performing an unmastered register under evaluation, and in both cases the cost of visible failure is calculated against professional standing.

Self-determination theory names what breaks here as competence, and competence is the need most directly assaulted by a tool that makes fluency look effortless in other people’s hands.

The identity discontinuity

AI inverts the order of expertise, and the inversion is sharpest for the people who know the most.

The clearest evidence sits in the comparison already cited: workers rated their own field an average of 14 percentage points more likely to be made redundant by AI than outsiders rated it. Proximity to a body of knowledge increases the conviction that it is being erased. The people best positioned to see which tasks are automatable are the same people whose sense of professional self is built on those tasks.

Meanwhile the benefits accrue upward. Directors and above reported saving nine hours per week through AI, managers seven, individual contributors four. Directors reported improved career prospects at twice the rate of individual contributors and were two and a half times more likely to say AI improved their job security. Sixteen percent of managers said AI use actually hurts their professional standing.

An organization can be simultaneously delivering real gains to its executives and real identity loss to the people below them, and read the average as success.

The power discontinuity

This is the one most likely to be mistaken for motivation.

Entry-level and early-career professionals report the most pressure to use AI, at 45%. Leadership was named as the source by 38% and direct managers by 36%. But 37% said the pressure comes from within themselves, rising to 44% among midlevel workers. Only 15% reported that negative consequences were used to drive adoption.

So this is largely not coercion. Something worse is happening.

Pressure that originates inside the person, unaccompanied by a considered judgment that the behavior is valuable, is what Deci and Ryan call introjected regulation. It is internal, which makes it look like ownership. It is contingent on self-worth and anxiety, which makes it the most fragile position on the continuum and the one most reliably associated with depletion.

Your most anxious adopters will read as your most motivated ones, right up until they stop.

And you cannot internalize what was never explained to you. Autonomy, in self-determination theory, does not mean independence. It means understanding the reason well enough to endorse it. For a full year, while individual AI use climbed from 40% to 52%, the share of U.S. workers who say their organization communicated a clear plan has not moved off one in four. Across four consecutive quarters, roughly one in five could not say whether their own employer had adopted AI at all.

And when someone does explain it, the effect is not marginal. Employees whose manager actively supports their team’s AI use are 8.7 times as likely to strongly agree AI has transformed how work gets done, and 7.4 times as likely to say it gives them more opportunities to do what they do best. Just over a third have that manager.

The single strongest variable in the entire dataset is not the worker’s aptitude, the training budget, or the tool. It is whether one person above them made the case.

The cultural discontinuity

The shortest to state and the slowest to surface. These systems encode particular norms about what a good answer looks like: which register reads as authoritative, which structure reads as rigorous, which reasoning reads as sound.

For a practitioner whose professional culture reasons differently, and for anyone whose expertise lives in judgment that resists articulation, fluency with the tool can quietly require adopting someone else’s standard of good thinking. That is a real cost, and it is currently invisible because we have no instrument pointed at it.

What the discontinuities have in common

Every one of them is being treated as an individual skills gap and addressed with more training.

None of them is a skills gap. They are mismatches between a person and an environment, which means they can be closed from either side, and the side we are not working is the one that would actually move.

What Replaces Competence

If the discontinuities are mismatches rather than skill gaps, then more training cannot close them, and the industry’s current product cannot either.

Walk any conference floor and count what is being sold. AI mastery in six weeks. Certification. Prompt libraries. Fluency programs. The entire market is organized around a competence model, and the competence model has been tested before, in another field, on another kind of learning, and it did not hold.

The argument that was already made and lost

In health care, cultural competence training was the standard for two decades. Providers were taught to acquire knowledge about the cultures of their patients. The critique that eventually unseated it was not that the training was poorly delivered. It was that the concept was wrong.

Lekas, Pahl, and Fuller Lewis (2020) argued that training providers to become competent in various cultures carries the risk of stereotyping and othering the very patients it intends to serve, and that by disregarding intersectionality it undermines recognition that a person occupies many social positions at once. Competence implies mastery over a bounded object. Culture is neither bounded nor static, so training people to be competent in it produces exactly what you would expect: a false sense of having finished, applied to something still moving. Beach and colleagues’ systematic review had already found that such training improved provider knowledge, attitudes, and skills while showing little effect on patient satisfaction or health outcomes (Beach et al., 2005).

Lekas and colleagues proposed a replacement built on four elements: self-reflexivity, appreciation of the patient’s own lay expertise, openness to sharing power, and a commitment to continue learning from each encounter rather than arriving at it already competent. They called abandoning the term cultural competence a paradigm shift that was overdue.

Substitute the technology for the culture and the critique lands unchanged. These systems change monthly. Their failure modes are not enumerable. Certifying someone as competent in them is certifying them as finished with something that is not.

I want to define the alternative before it is misread. Humility here is not modesty, deference, or reduced ambition. It is a higher standard of rigor, and it is more demanding than certification, not less.

The data agrees, and it is counterintuitive

SHRM’s aptitude assessment produced the finding that should end the competence conversation.

Workers scoring highest on objective AI knowledge reported lower reliance on AI than workers in the medium-low group, and rated critical thinking as more important. Understanding the system produced caution. And workers scoring below the median admitted to breaking their organization’s AI policies at 38%, against 23% for those above it.

Confidence without understanding is the actual risk profile, and it is precisely what a six-week mastery program manufactures.

Then there is the finding from the opening. Asked what they would do if an AI recommendation conflicted with what they believed was the most ethical choice, 55% of directors and above said they would follow the AI.

That is the competence model’s finished graduate. Maximum fluency, maximum confidence, and judgment surrendered at the exact point where it was the only thing that mattered.

What humility looks like operationally

Four practices, mapped to the four elements Lekas and colleagues named. Each can be trained, observed, and assessed. None of them is a feeling.

Self-reflexivity becomes a written account of what is not delegable. Every function should be able to state which judgments belong to a person and why. The 55% finding is what happens in an organization that never wrote that document. Put the ethical override in it explicitly.

Appreciation of lay expertise becomes appreciation of the practitioner’s own experiential knowledge. The twenty-year practitioner holds judgment that has never been written down and cannot be prompted for. That is not a gap to be filled by the model. It is the asset the model has no access to, and the organization should say so out loud, because right now the people holding it believe the opposite.

Sharing power becomes treating output as a position rather than an answer. The model has a perspective, shaped by what it was trained on and what it optimizes. Working well with a position requires disagreeing with it sometimes. Working with an oracle does not.

Continuing to learn from the encounter becomes verification with a stated standard. Not “check the output,” which everyone nods at and no one does. A named threshold per task type: what must be independently confirmed, by whom, against what source, before this leaves the building. Workers already spend about four hours a week correcting AI output. That labor exists. It is simply undisciplined.

And what replaces the training program

Four shifts, each answering one of the discontinuities.

From event to practice. The productivity return rises with breadth of application, not with hours of instruction, which means the unit of design is repeated encounter inside real work rather than a session on a calendar.

From individual to environment. The 8.7x sits with the manager. If your AI budget has no line item for equipping the people one level above the users, you have funded the wrong thing.

From mandate to reason. Three in four workers have not been told the plan. Pressure without explanation produces introjected regulation, which looks like enthusiasm and behaves like attrition.

From sustaining nothing to sustaining something. Paris and Alim’s contribution to pedagogy was moving past accommodating what learners bring toward actively sustaining it (Paris & Alim, 2017). Applied here: an organization can treat professional judgment as friction to be minimized during adoption, or as an asset to be deliberately preserved. Only one of those makes integration available.

Which returns us to the four positions

Nobody in your organization is standing still. Everyone in contact with these systems is resolving the same two questions, whether or not anyone has asked them out loud. What do I keep. What do I join.

Four answers are possible. One of them produces people who are fluent and still themselves. The other three cost you something you will not see on this quarter’s dashboard: the judgment of your most experienced people, or the confidence of your quietest ones, or the discernment of your most senior ones at the precise moment it was needed.

Which answer your people reach is not a measure of their character or their aptitude. It is a measure of what you made available.

The environment is the intervention. And environments are built by people, which means they can be built differently.

References

Beach, M. C., Price, E. G., Gary, T. L., Robinson, K. A., Gozu, A., Palacio, A., Smarth, C., Jenckes, M. W., Feuerstein, C., Bass, E. B., Powe, N. R., & Cooper, L. A. (2005). Cultural competence: A systematic review of health care provider educational interventions. Medical Care, 43(4), 356–373. https://doi.org/10.1097/01.mlr.0000156861.58905.96

Berry, J. W. (1997). Immigration, acculturation, and adaptation. Applied Psychology: An International Review, 46(1), 5–34. https://doi.org/10.1111/j.1464-0597.1997.tb01087.x

Berry, J. W., Phinney, J. S., Sam, D. L., & Vedder, P. (2006). Immigrant youth: Acculturation, identity, and adaptation. Applied Psychology: An International Review, 55(3), 303–332. https://doi.org/10.1111/j.1464-0597.2006.00256.x

Gallup. (2026). Indicator: Artificial intelligence. https://www.gallup.com/699797/indicator-artificial-intelligence.aspx

Gee, J. P. (1996). Social linguistics and literacies: Ideology in discourses (2nd ed.). Taylor & Francis.

Kemp, A. (2026, July 20). Organizational AI adoption jumps six points. Gallup. https://www.gallup.com/workplace/712736/organizational-adoption-jumps-six-points.aspx

Lave, J., & Wenger, E. (1991). Situated learning: Legitimate peripheral participation. Cambridge University Press.

Lekas, H.-M., Pahl, K., & Fuller Lewis, C. (2020). Rethinking cultural competence: Shifting to cultural humility. Health Services Insights, 13, 1178632920970580. https://doi.org/10.1177/1178632920970580

Nieto, S., & Bode, P. (2018). Affirming diversity: The sociopolitical context of multicultural education (7th ed.). Pearson.

Paris, D., & Alim, H. S. (2017). What is culturally sustaining pedagogy and why does it matter? In D. Paris & H. S. Alim (Eds.), Culturally sustaining pedagogies: Teaching and learning for justice in a changing world (pp. 1–21). Teachers College Press.

Ryan, R. M., & Deci, E. L. (2000). Self-determination theory and the facilitation of intrinsic motivation, social development, and well-being. American Psychologist, 55(1), 68–78. https://doi.org/10.1037/0003-066X.55.1.68

Ryan, R. M., & Deci, E. L. (2017). Self-determination theory: Basic psychological needs in motivation, development, and wellness. Guilford Press.

SHRM. (2026). Navigating AI in the workplace: 2026. https://www.shrm.org/topics-tools/research/navigating-ai-in-the-workplace/full-report

WRITER. (2026). AI adoption in the enterprise (conducted with Workplace Intelligence). https://writer.com/blog/enterprise-ai-adoption-2026/