Is AI Making Us Smarter

A decade ago, the worry was that Google was making us shallow thinkers, content to skim rather than know. Today the same anxiety has a new target, and the stakes feel higher. Generative AI doesn't just retrieve information, it writes, reasons, summarizes, and decides alongside us. So the question has sharpened: is this technology extending human intelligence, or quietly hollowing it out?

The honest answer is that the evidence points in both directions at once, depending on what you measure and how the tool is used.

‍

‍

The case for dependency

‍

The strongest empirical warning comes from research on "cognitive offloading", the tendency to hand mental tasks over to an external system rather than doing them yourself. A widely cited 2025 study by researcher Michael Gerlich, surveying 666 participants, found a clear pattern: frequent AI users scored measurably lower on critical-thinking assessments than infrequent users, and cognitive offloading appeared to be the mechanism connecting the two. The effect was strongest among 17-to-25-year-olds, the demographic that has grown up treating AI assistance as a default rather than a novelty.

This isn't an isolated finding. A parallel line of research applies Cognitive Load Theory to the same question: AI can strip away the "extraneous" mental effort of a task, the tedious parts which in theory should free people to focus on deeper thinking. But researchers have found that it often strips away the "germane" load too: the productive struggle where learning actually happens. When AI resolves ambiguity before a person has to wrestle with it themselves, the mental muscle that would have done that wrestling never gets used.

There's also a trust problem layered on top of the skill problem. Studies on automation bias, the tendency to over-trust machine output, show that generative AI can intensify people's willingness to accept a plausible-sounding answer without checking it, especially under time pressure or in high-stakes fields like medicine. And unlike a calculator or a search engine, a language model will confidently produce a wrong answer with the same fluent tone as a right one, which makes the uncritical-acceptance problem harder to catch.

Newer research goes further, using randomized controlled trials rather than surveys, and finds causal evidence that AI assistance can reduce a person's persistence and independent performance on tasks, not just their self-reported reliance, but their actual ability to do the work unaided afterward.

‍

‍

The case for augmentation

‍

Set against this is a substantial body of research showing genuine gains. MIT Sloan's 2025 framework on human-AI complementarity argues that most knowledge work isn't cleanly automatable, it benefits more from augmentation, where AI handles a sub-task and a person supplies judgment, context, and integration. Anthropic's own workplace research, published in early 2026, estimated that tasks taking around 90 minutes without AI assistance were completed roughly 80% faster with it, and observed something more interesting than raw speed: people used the freed-up time to take on work that used to be outside their role entirely. Designers wrote code. Researchers took on engineering tasks. The tool didn't just do the old job faster, it let people attempt jobs they'd previously have avoided or deferred.

Labor-market studies tell a similar story at the macro level. An analysis of over five million U.S. patents found that generative AI capabilities in language, creativity, and decision-making were associated with more hiring and higher firm value, not job loss, a pattern of augmentation rather than replacement, at least so far. Other field experiments found productivity gains concentrated among lower-skilled or less-experienced workers, suggesting AI can act as a leveling tool, compressing the gap between novices and experts rather than just making experts faster.

Even some of the offloading research includes a hopeful caveat: cognitive offloading isn't fixed. A follow-up study found that interaction design matters enormously, prompts that force a user to justify their reasoning, or that build in retrieval practice, preserve most of AI's speed benefits while avoiding the slide into passive acceptance. The damage isn't inherent to the technology; it's a property of how the technology is used.

‍

‍

Two effects, not one

‍

The tension resolves somewhat once you separate two different things people mean by "smarter": raw output and underlying skill.

On output, the evidence for AI as an amplifier is fairly strong, better first drafts, faster analysis, broader task range, and real gains for people starting from a weaker skill base.

On underlying skill, the capacity to do the work well without the tool, the evidence is more worrying. Offloading, automation bias, and reduced persistence all point toward atrophy when AI use is heavy, passive, and unstructured, and the effect is most visible in younger users who never built the unassisted skill in the first place.

Not everyone loses skill at the same rate, though. Some research suggests the same person can be an amplified thinker in one context (using AI as a sparring partner, checking its work, pushing back on its reasoning) and a dependent one in another (accepting the first answer, skipping the step where they'd have had to think it through themselves). The tool doesn't determine the outcome; the habit of use does.

‍

‍

What seems to actually matter

‍

A few patterns show up across the research worth taking seriously:

• Struggle is doing something. If a task is easy enough that AI removes only tedium, offloading it looks safe. If the task is where real learning happens — working through an argument, debugging your own logic — outsourcing it removes the thing you were supposed to build.
• Verification habits matter more than usage frequency. The research on automation bias suggests the risk isn't using AI often, it's using it uncritically — treating a fluent answer as a checked one.
• Age and baseline skill shape the effect. People who built strong independent reasoning before relying on AI tools seem more resilient than those who never did.
• Design nudges work. Tools that prompt justification or force a moment of independent effort before revealing an answer measurably preserve critical thinking, compared with tools that just hand over the result.

‍

‍

So smarter, or more dependent?

‍

Probably both, distributed unevenly across the population and across the moments within a single person's day. The technology is powerful enough to make thoughtful people more capable and incurious people more passive, often at the same time, using the same tool. The open question isn't really about AI's capabilities, it's about whether the habits people build around it involve any friction at all, or whether ease quietly becomes the only setting anyone uses.

‍

‍

‍