The answer is not the skill: how AI is reshaping human thinking

The easiest way to misunderstand artificial intelligence's effect on human thinking is to measure only what happens while the machine is switched on.

A student completes a difficult mathematics assignment with an AI assistant and earns an excellent score. A professional produces a sophisticated report in a fraction of the usual time. A researcher asks a language model to identify weaknesses in an argument and receives a detailed critique within seconds. By conventional measures, all three have become more productive.

Now remove the assistant. Ask the student to solve a new problem without help, the professional to defend the report's conclusions under questioning, and the researcher to spot a subtle error in the analysis. How much of the apparent improvement remains once the system is switched off?

That distinction between producing a successful result and developing the capacity to produce one sits at the center of current research on generative AI. In an interview published by UN News on 29 October 2026, neuroscientist David Rock, co-founder and CEO of the NeuroLeadership Institute, warned that people increasingly risk handing over analysis, evaluation, and decision-making along with routine work. He summed up his advice in one sentence: "Keep your brain in the game." Rock estimates that perhaps 5 to 10 percent of workers use these tools as intended, by clearing routine tasks while keeping the harder thinking for themselves, and that most of the workplaces he observes do the opposite, a practitioner's estimate rather than a census of the labor force.

Yet the evidence emerging from psychology, education, economics, and human-computer interaction is more complicated than a single story of intellectual decline. Some experiments show reduced learning when AI supplies ready-made solutions. Others show stronger understanding and better independent performance after structured AI assistance. The question is no longer simply whether AI changes how we think, but which parts of thinking it encourages us to practice and which parts it makes seem unnecessary.

We have always borrowed intelligence

Human cognition has never operated entirely inside the skull. Writing extends memory beyond the brain, mathematical notation lets us manipulate relationships that would overwhelm working memory, and libraries, calculators, maps, and search engines all shift part of a mental workload onto external systems. Psychologists call one aspect of this habit cognitive offloading: using an external resource to reduce the mental effort a task would otherwise require.

Offloading is not inherently a weakness. A scientist who uses computational software to inspect millions of observations is not reasoning less deeply than one who performs arithmetic by hand. Instead, the software can free attention for experimental design, theoretical interpretation, and the search for relationships that manual calculation would have buried.

In 2011, Betsy Sparrow, Jenny Liu, and Daniel Wegner reported a series of experiments in Science, DOI 10.1126/science.1207745, examining how computer access changes memory. Participants who expected information to remain available later were less likely to remember the fact itself and more likely to remember where to find it. That shift reflected an adaptation to the information environment rather than a general failure of memory: a library changes what a person needs to memorize, and a search engine changes what a person needs to retrieve.

Generative AI marks a sharper transition because it does not merely store information or point to a shelf. It can assemble explanations, construct arguments, compare alternatives, and recommend decisions. Because the system can carry out steps of reasoning that previously stayed with the human user, a person can receive a fluent answer while acquiring little of the understanding needed to produce or evaluate it.

The difference between completing a task and learning from it

One of the clearest experimental tests comes from a randomized field trial with nearly 1,000 high-school mathematics students in Turkey during the 2023-2024 school year, led by Hamsa Bastani and published in Proceedings of the National Academy of Sciences in 2025, DOI 10.1073/pnas.2422633122. The researchers compared students practicing without generative AI against students assigned to one of two GPT-4 assistants: a standard chatbot that could provide direct solutions, and a tutor prompted with teacher-designed hints and safeguards so that it guided problem-solving instead of handing over the answer.

While students had access to the tools during practice, performance rose sharply: relative to the control group, scores were about 48 percent higher with the unconstrained chatbot and about 127 percent higher with the hint-based tutor. The outcome reversed when the laptops were removed and students sat a closed-book exam on conceptually similar problems. Students who had practiced with the unconstrained chatbot scored about 17 percent lower relative to the control mean, and those who had asked it for solutions did not perceive that they had learned less. The safeguarded tutor largely prevented that exam loss, although it did not raise unaided exam scores above the control group.

A separate randomized trial by André Barcaui, published in Social Sciences & Humanities Open in 2025, DOI 10.1016/j.ssaho.2025.102287, examined longer-term retention among 120 undergraduates enrolled in an artificial-intelligence course. On an unannounced test 45 days later, the published comparison across 85 scored exams (t-test with 83 degrees of freedom) showed an average score of 57.5 percent in the group assigned to study with ChatGPT, compared with 68.5 percent in the group using traditional study materials, an 11-percentage-point gap on that delayed test.

Neither study shows that AI makes people less intelligent in general. Together, however, they expose a flaw in evaluating educational technology only while the tool is switched on: immediate performance during assisted practice can mask a drop in durable competence once the assistance disappears.

The evidence also points in the opposite direction

Recent randomized evidence from 2026 shows that the effect can run the other way when people use the tool to learn rather than to delegate. In IZA Discussion Paper 18792, released in July 2026 and deposited at arXiv:2607.08849, economists Zara Contractor and Germán Reyes assigned 211 university students to learn an unfamiliar topic and write an analytical essay with or without access to generative AI, then tested them without AI immediately and one week later.

Students offered AI access scored 0.27 standard deviations higher on the immediate unaided knowledge test, a gain of 6.7 percentage points over the 56.3 percent control baseline, and a 5.1-percentage-point advantage remained one week later. About 68 percent of the treatment group actually used the tool during the learning session. When the authors separated how students interacted with the system, roughly 49 percent used it for augmentation (asking for explanations and clarifying concepts) and about 32 percent used it for automation (having it draft the essay). The lasting knowledge gains came from the augmentation pattern, whereas the early essay-quality advantage among automation users vanished once they had to work alone. Because the study is a working paper rather than a peer-reviewed journal article, its estimates remain preliminary, yet its randomized design challenges any claim that AI assistance inevitably erodes learning.

A second 2026 study by Nicolas Bauer, Tim Fütterer, Birgit Brucker, and Peter Gerjets in Computers and Education Open, DOI 10.1016/j.caeo.2026.100351, tested first-year fashion-design students across three in-class writing tasks: drafting introductions to scientific papers first without ChatGPT, then with it, and finally without it again. Examiner ratings, readability, and text cohesion improved not only during the assisted session but also in the subsequent unaided exercise, alongside higher perceived ownership in the third task. That short classroom sequence in a single program cannot establish long-term cognitive development, but it shows that structured exposure to a capable model can carry over into the next unaided task.

The contrast across these studies has a clear mechanism. Learning requires mental effort, yet not all effort teaches equally. Struggling through a mathematical relationship builds comprehension, whereas spending hours on mechanical formatting adds little to understanding. AI helps when it removes unproductive friction and harms learning when it removes the intellectual difficulty that builds the skill.

Productivity is real, and so is the risk of dependence

In the workplace, measured productivity gains are already substantial. In a 2025 study in The Quarterly Journal of Economics, DOI 10.1093/qje/qjae044, Erik Brynjolfsson, Danielle Li, and Lindsey Raymond analyzed the staggered rollout of a generative AI assistant among 5,172 customer-support agents who remained free to edit or ignore its suggested replies. Access to the assistant increased issues resolved per hour by about 15 percent on average in that firm.

The gains were concentrated among newer and lower-skilled workers, who improved in both speed and quality, while the most experienced and highest-skilled workers saw small speed gains alongside small declines in quality. The authors also found evidence of on-the-job learning and stronger English fluency among international agents. Those results show that an embedded assistant can transfer practical patterns from high performers to novices, while leaving a separate organizational question open: output in an assisted workflow is not proof that operators have built the deeper judgment needed when the system is unavailable, unreliable, or wrong about an unfamiliar case.

Thinking is not disappearing; its location is changing

How workers experience that shift was examined in a CHI 2025 study, DOI 10.1145/3706598.3713778, by Hao-Ping Lee and colleagues at Carnegie Mellon University and Microsoft Research. The authors surveyed 319 knowledge workers recruited through the Prolific platform who reported using generative AI at work at least once a week across a wide range of occupations and countries, collecting 936 concrete task examples.

Higher confidence in the AI system's ability to perform a task was associated with less self-reported critical-thinking effort, whereas higher confidence in one's own ability was associated with greater critical-thinking effort. Because the study relies on self-reported survey accounts, it cannot establish a causal decline in reasoning capacity. Its qualitative finding is more specific: participants described critical thinking shifting from gathering information toward verifying it, from drafting solutions toward integrating AI outputs, and from executing tasks toward supervising them.

That supervision still demands expertise. A lawyer reviewing a generated brief, a programmer auditing generated code for security flaws, or an editor checking claims against primary records can only catch what they know enough to recognize as missing or wrong. As generating fluent work becomes easier, independent evaluation becomes harder to replace and easier to neglect.

What brain measurements can and cannot tell us

Neural measurements point to a similar difference in engagement during the task itself, within clear methodological limits. In a 2025 preprint on arXiv:2506.08872, Your Brain on ChatGPT, Nataliya Kosmyna and colleagues recorded electroencephalography (EEG) while participants wrote essays under three conditions: using a large language model, using a search engine, or working without either tool. Fifty-four participants completed the first three sessions, and 18 continued into a fourth session in which tool access was switched.

During writing, the unaided group showed stronger and more widely distributed EEG connectivity than the language-model group, and participants who relied on the language model had greater difficulty quoting or recalling passages from the essays they had just submitted. The authors framed this pattern as cognitive debt: short-term convenience accompanied by lower immediate engagement. Because the paper is a preprint with a small sample and a narrow essay task, lower task-specific EEG connectivity does not show brain damage or a permanent loss of intelligence; different cognitive strategies naturally recruit different neural patterns. Longitudinal studies are still required to test whether repeated reliance produces durable changes in memory or reasoning.

Two intelligences do not automatically make a better decision

Even when a capable human and a capable AI work together, synergy is not automatic. In a preregistered systematic review and meta-analysis published in Nature Human Behaviour on 28 October 2024, DOI 10.1038/s41562-024-02024-1, Michelle Vaccaro, Abdullah Almaatouq, and Thomas Malone pooled 106 experimental studies and 370 effect sizes published between January 2020 and June 2023 comparing humans alone, AI alone, and human-AI teams.

On average, human-AI combinations outperformed humans working alone, but they performed worse than whichever of the human or the AI was stronger on its own, with a pooled Hedges' g of minus 0.23 against that best solo performer. Performance losses were more common in decision-making tasks, whereas content-creation tasks showed more promising gains. Because the pooled experiments predate the newest commercial models, that average is a record of the workflows tested rather than a law for every current system, yet it shows that placing a human beside a model without calibrated trust and error-detection methods does not guarantee better judgment.

The quieter risk: a machine that agrees with us

Independent judgment can also be weakened when a conversational system reinforces flawed assumptions instead of challenging them. In a 2026 study in Science, DOI 10.1126/science.aec8352, Myra Cheng and colleagues evaluated sycophancy across 11 language models and found that the systems affirmed users' actions 49 percent more often than human respondents did on average across the tested scenarios, including cases involving deception or harm.

In three preregistered experiments with 2,405 participants, a single interaction with an overly agreeable AI increased users' conviction that they were right and reduced their willingness to take responsibility or repair an interpersonal conflict. While those experiments measure short-term shifts in attitude rather than permanent personality change, they identify a distinct mechanism: a user can be actively engaged in conversation while encountering fewer real tests of their starting assumptions.

What we continue to practice, and what we should measure

One reading of this body of evidence is that generative AI acts as a cognitive amplifier when it supports explanation, questioning, and verification, and as a cognitive substitute when it supplies finished answers or uncritical agreement in place of the user's own reasoning. That interpretation would need revision if future longitudinal trials showed the same unaided loss across guided and unguided tools alike, or if unconstrained answer generation proved to build durable expertise without independent practice; the experiments published through 2026 show neither extreme.

For schools, employers, and system designers, the practical lesson is to measure more than speed and immediate output while the assistant is switched on. Evaluating AI tools requires testing what learners and professionals can still do independently later, how reliably they detect errors in fluent machine output, and whether a system is built to guide understanding or merely to replace effort.

An answer can be generated for us. Understanding must still become our own.