Perspectives · Sustainability & AI

The AI Slowdown Debate Leaves Out Human Adaptive Capacity

On September 12, Dario Amodei published an essay asking the AI companies, his own included, to slow the rate at which they improve their most capable models.¹ Sam Altman, Elon Musk, and Demis Hassabis said the same day that they agreed with the direction. I read a lot of AI headlines. That one stayed with me, for a reason most of the commentary walked past.

I work in sustainability, where the only question that finally matters is whether the demand placed on a system stays inside what that system can supply, absorb, or rebuild. You can pump an aquifer faster than it recharges for years before the wells tell you, and by then the decision that drained it is a decade old.

The executives are applying that test to their own side of AI. They are asking whether the ability to control and verify these systems is growing as fast as the systems are. It is a fair question and I am glad they asked it. The side they are not asking about is ours. Humanity is on the receiving end of everything already shipped, and our capacity to absorb it while keeping our own judgment intact is finite. No one on earth has the job of measuring it.

The two rates are nowhere near each other. Machines change in months. People change in years, because skills, judgment, curricula, professional standards, and laws all run on the slower clock. Stanford's 2026 AI Index reports that 88 percent of surveyed organizations used AI in 2025 and that four out of five university students use generative AI.² An OECD review of training catalogues in four countries found that between 0.3 and 5.5 percent of courses included any AI content.³ Almost everyone has the tool, and next to nobody is being taught to use it.

Regulators noticed the gap and dealt with it by lowering the bar. The EU AI Act originally required organizations to make sure their staff reached an adequate level of AI literacy. In July, the EU swapped that for a duty to take supporting measures, with no level guaranteed for any human being. The main obligations for high-risk systems, human oversight among them, moved from August 2026 to December 2027. The stated reason was that the standards and the national authorities were not ready. The systems those rules cover stayed in use. So a person is expected to be overseeing high-risk AI at the end of 2027, and nothing in the law now makes sure that person will be able to.

Take judgment. Thirty years of watching people get good at things has shown me one mechanism. You do the work, you get it wrong, someone corrects you, you do it again. Competence comes out of that loop and nowhere else. When AI does the reasoning and a person reviews the answer, the loop runs less often than the job used to make it run. Stanford's Digital Economy Lab reported in August, from ADP payroll records, that the number of workers aged 22 to 25 in the most AI-exposed occupations is 19 percent below where it would be if it had grown in line with less-exposed peers, mostly because fewer are being hired.⁴ Each firm that cuts entry-level hiring is making a cost decision that is rational for the firm. No firm anywhere is accountable for the supply of senior judgment ten years out, and the people who would have become that judgment are not in the building.

Inside organizations, adoption gets measured and people do not. Cost and output show up every quarter, while loss of human capability and growth in dependency have no line in any report I have seen. The unit that books the saving is rarely the unit that holds the risk. I call this an integrity problem in the governance sense. The name stays on the signature. The ability to know what you are signing moves to a system that answers to no one. Approvals keep going out on the assumption that someone could check them, and each year fewer people can.

The human adaptive ceiling

I use the term human adaptive ceiling for the upper limit, meaning the rate of technological change above which people lose competence, oversight, and the capacity to recover faster than they can rebuild them. The term borrows from Kate Raworth's Doughnut Economics, which places human development between a social foundation and an ecological ceiling.⁵ Her model has a lower boundary too, and for technology that boundary is access, so any slowdown carries a cost for the people who are still excluded. An ecological limit is set by physics. This one moves, rising with investment in training, governance, and shared human knowledge, and AI itself can help build all three. A limit that moves cannot be looked up in a reference table, so it has to be measured, and the model below is my attempt at how.⁶

The Human Adaptive Capacity Model for AI

The model has three zones and two boundaries, measured in units of us. The inner boundary is the human capability foundation, the minimum a person needs to take part in an AI-shaped world and keep their own judgment while doing it. Eleven dimensions sit on that ring: digital inclusion, equity and fair access, decision authority, accountability and redress, workforce mobility and transition support, institutional knowledge and resilience, professional competence and practice, education and lifelong learning, access to technology and infrastructure, AI literacy and critical understanding, and human agency and judgment. Below the ring is capability deficit: you cannot access the tool, you cannot evaluate what it tells you, or you have no standing to challenge it. A large share of humanity is there today, before any slowdown is discussed.

The outer boundary is the human adaptive ceiling, the point at which the rate, scale, autonomy, complexity, or concentration of AI pressure exceeds what people and their institutions can absorb. Thirteen pressures push against it: rate of capability change, scale of deployment, degree of autonomy, opacity and lack of explainability, concentration of control, dependency on AI systems, skills obsolescence, workforce displacement rate, governance lag, erosion of human agency, unequal access and power asymmetries, high-consequence decisions delegated to machines, and system complexity. Past the ceiling is technological overshoot, and its consequences are what you would expect once people can no longer keep up: dependency, skills erosion, loss of agency, governance failure, institutional lag, reduced resilience, concentration of power, and social harm and exclusion. The missing young workers and the EU's deferral of its own oversight rules are early readings on two of the thirteen, skills obsolescence and governance lag.

Between the boundaries is the regenerative AI operating space, the range in which AI leaves people more capable, better at judging, and quicker to recover, with accountability still in human hands. It exists only when the foundation is met and the pressure stays under the ceiling at the same time. A society can meet the foundation and still be in overshoot, and it can sit well under the ceiling while a third of its people are in deficit.

Two working expressions run the model. The first is the adaptive pressure ratio, technological pressure divided by human adaptive capacity. Under 1.0 there is room. At 1.0 we are at the ceiling. Above it we are in exceedance, and exceedance that lasts is overshoot. The second is the adaptation deficit, the rate of AI capability growth minus the rate at which humans adapt. A deficit that widens year on year raises the odds of crossing the ceiling later, even while today's pressure is still under it. Figure 2 shows that relationship, and of the two figures it is the one I would ask readers to sit with. The table below gives an illustrative measure for each side of seven dimensions. None of the thresholds is calibrated yet. I am offering the model as an instrument for analysis and measurement; the calibration is the work I think this debate should be about.

Table 1. Illustrative dimensions and metrics. Foundation metrics test whether the minimum condition is met for people; ceiling metrics test whether pressure exceeds their capacity. Thresholds are proposed starting points and have not been empirically calibrated.

What the measurement looks like

In my field, regeneration means a measurable increase in system capacity beyond baseline. Applied to AI, the people who use a system should be more capable afterward, measured with the system switched off. That is a test of human beings, and organizations can run it with the same discipline they already apply to other things they cannot afford to lose. Drill it on a schedule, the way a continuity plan is drilled, and record the share of AI-supported tasks a person can still complete and check without the system. Count the entry-level people doing unassisted work in each discipline, because that number is the future supply of judgment. Name a human for every AI-supported decision who has the authority to override it and the demonstrated ability to evaluate it. Time how long it takes people to restore a process when the system is gone.

Every one of those measures costs money the business case for adoption was written to save. Practice time, manual drills, and junior positions are paid for now. The benefit arrives later, in a different budget, on the day a system fails or a senior cohort retires. A company that funds human capacity will carry higher costs than one that does not, in every quarter in which nothing breaks. The labs face the same arithmetic. President Trump said the day after the essay that a slowdown was not needed and would put the American lead over China at risk. AI shares fell on the Monday.

The debate that opened this month is about how fast to build the machines, and it has not reached the question of how fast people can absorb them. Human adaptive capacity is finite. AI planning treats it as unlimited by leaving it out. There is no agreed measure, no threshold, no body assigned to set one, and the parties who decide the rate of adoption do not carry the cost of exceeding it. I do not expect any of that to change inside the current planning cycle.

¹ Amodei, D. Essay published September 12, 2026, calling on AI companies to slow the rate of improvement of their most capable models.

² Stanford Institute for Human-Centered AI. AI Index Report 2026.

³ OECD review of training catalogues in four countries; share of courses including AI content.

⁴ Stanford Digital Economy Lab, August 2026, analysis of ADP payroll records covering workers aged 22 to 25 in AI-exposed occupations.

⁵ Raworth, K. Doughnut Economics: Seven Ways to Think Like a 21st-Century Economist. Chelsea Green, 2017.

⁶ Carboni, J. B. (2026). The Human Adaptive Capacity Model for AI. Conceptual framework. The full register covers 11 foundation dimensions, 13 ceiling pressures, and 8 overshoot consequences, each with an illustrative indicator and unit.