Nobody Is Lying About AI
The debate about AI’s effects has been running for three years and producing mostly heat. Not because the researchers are wrong. Because they are not studying the same thing.
Here is the clearest version of what I mean. A major 2025 meta-analysis of 69 studies found that ChatGPT produces significant positive effects on academic performance, including something the authors called reduced mental effort, which they listed in the benefits column (Deng et al., Computers & Education, 2025). A separate MIT study from the same year measured neural connectivity during sustained LLM use and found that reduced mental effort is the neurological signature of cognitive debt: measurable EEG degradation, impaired recall, weakened critical thinking that persisted after the AI was removed (Kosmyna et al., MIT Media Lab, 2025).
The same variable. A benefit in one paper. A harm in the other. Both studies are correct.
The difference is time horizon. The meta-analysis measures this week’s assignment. The MIT study measures what happens four sessions later.
This is not a dispute about data quality. It is a disagreement about which level of analysis should count. And that disagreement runs through everything in the AI conversation, whether or not anyone names it.
The argument about AI is not one argument. It is five arguments, each conducted at a different level of analysis, by people who have largely stopped reading each other. Each produces a real picture. None is complete. And because no one has agreed on which level should have standing, every study confirming AI works and every study confirming AI harms is simultaneously right and talking past everyone else.
| Level | What AI Does | Where the Debate Lives |
|---|---|---|
| Task | +40% output gains on tasks inside its capability range (Dell’Acqua et al., 2023). But -19% degradation on tasks just outside that range, at a frontier the worker cannot see. The gains are real and so is the boundary. | Productivity researchers, management consultants, individual knowledge workers experiencing the gains directly. The “AI works” argument is made here and it has the most citations. The time horizon is this quarter. |
| Person | The same tool that made you faster quietly erodes the capacity to do the work without it. Skills atrophy when people return to unassisted tasks (Wu et al., 2025). Neural connectivity measurably weakens with sustained LLM use (Kosmyna et al., 2025). The damage is invisible in output metrics and shows up later. | Learning scientists, cognitive psychologists, and educators who notice that students who produced better papers cannot explain what they wrote. The deskilling argument lives here, alongside the evidence for it. |
| Organization | Rational organizations are trapped into automating even when they would collectively prefer not to. Each captures the full cost saving from reducing headcount and bears only a fraction of the demand it destroys (Falk & Tsoukalas, arXiv, 2026). They can see the trap and still cannot escape it. | CEOs, CHROs, operations leaders, and boards making adoption decisions in real time. The debate here is not academic: it is about rollout timelines, workforce transitions, and whether the productivity numbers justify the displacement. The people in these rooms are not economists. They are leaders trying to stay competitive. |
| Economy | Productivity gains are real but concentrated. They flow to capital, not labor. Automation explains 52% of U.S. wage inequality growth since 1980 (Acemoglu & Restrepo, 2022). For every measurable increase in regional AI adoption intensity, labor income share declines. | Labor economists, union leaders, policy advocates, and government officials watching what automation does to wages and employment at scale. The argument here is about distribution, not productivity. The gains are not in dispute. The question is who gets them. |
| Civilization | Threatens the conditions that make work worth doing. The connective labor of human recognition degrades (Pugh, 2021). AI removes the conditions for meaningful achievement even without removing the job itself (Danaher & Nyholm, 2020). Embodied expertise erodes because development requires internalizing what AI externalizes (Dreyfus, 2004). AI can replicate the output but not the formation that made the output worth producing. | Philosophers, humanists, theologians, and writers asking what work is for. This argument is the hardest to measure and the easiest to dismiss. It is also the one that, if correct, makes everything above it a footnote. |
The task-level gains documented by Dell’Acqua and colleagues are as real as the person-level deskilling documented by Wu and Kosmyna. The organizational trap modeled by Falk and Tsoukalas is as real as the productivity case made in every corporate AI adoption brief. The labor-market concentration documented by Acemoglu and Restrepo is as real as the output gains two rows above it.
Nobody is lying. Everybody is describing what they can see from where they are standing. From the task floor, the view is good. From the civilization summit, the same terrain looks different.
What makes the debate irresolvable is not the quality of the evidence. It is that no one agrees on which level has standing. The productivity researcher is not answering the philosopher’s question. The philosopher is not addressing the CEO’s quarterly decision. The labor economist is not helping the classroom teacher figure out whether to allow AI on tomorrow’s exam. Thus the argument continues, loudly, without going anywhere.
The levels that are easiest to measure tend to dominate the debate, and that is partly a visibility problem. Task-level output gains show up in dashboards. Cognitive debt accumulates in ways no dashboard captures. Labor income share declines over decades. The erosion of embodied expertise is nearly impossible to quantify until the expertise is already gone. But visibility is not the whole story. The levels that are hardest to measure also tend to be the ones where the costs, not the gains, concentrate. Task-level gains flow most directly to whoever deployed the tool. Civilization-level costs are distributed across everyone. The people arguing loudest from the levels that dominate are not always confused about methodology. Sometimes they are arguing from exactly where their interests sit.
Proximity is not the same as accuracy. The person closest to a technology is not automatically the one with the clearest view of what it is doing at scale.
The debate will not resolve by finding a single answer that works at all five levels. It will move when the right people are working at the right level, with enough expertise to know what their level can and cannot see.
The person arguing from the task level who does not understand the person-level costs is not fully right even when the data is correct. The philosopher who cannot engage the productivity evidence is not fully right either. Depth at one level, combined with awareness of the others, is what gives someone actual standing to argue. Most people in this debate have the depth. What is missing is the map.
This obligation lands on anyone in a position to develop other people’s capacity to think. K-12 teachers, corporate trainers, and managers carry a version of it too. But universities are where the stakes are highest, because the task is not just training people for their current jobs. It is training the people who will make decisions at all five levels for the next forty years. Some will make organizational decisions about AI adoption. Some will shape labor policy. Some will do the cognitive science. Some will ask the philosophical questions. If they leave without a working map of the levels, they will reproduce the same argument we are having now, with better data and the same confusion.
Thus the task is not just to use AI well. It is to understand it well enough, at the level where you operate, to prepare people who will operate at levels you will never reach. That is what makes AI expertise in education a civic obligation rather than a professional preference.
