Something subtle is happening every time we ask an AI to write our email, summarize our document, or answer a question we could have worked out ourselves. We get the output. We skip the process. And it turns out the process was the point.
A growing body of research from some of the world's leading institutions is beginning to quantify what many educators and researchers have suspected: routine dependence on AI tools may be quietly eroding the very cognitive capacities that make humans irreplaceable.
What the Research Shows
In early 2025, researchers at Microsoft and Carnegie Mellon University published findings from a study of 319 knowledge workers, examining how reliance on AI tools affected their critical thinking. The results were pointed: the more workers leaned on AI assistance, the less critical thinking they engaged in. Those who had high confidence in AI's abilities showed the steepest drop in cognitive engagement — essentially delegating judgment alongside the task.
The researchers described what they called a "key irony of automation": by mechanising routine tasks and leaving exception-handling to the human, you deprive that person of the routine practice that sharpens judgment. Cognitive muscle, left unused, atrophies. The study also found that workers using AI for critical thinking tasks produced less diverse outcomes — converging on similar answers, thinking in narrower channels.
"A key irony of automation is that by mechanising routine tasks and leaving exception-handling to the human user, you deprive the user of the routine opportunities to practice their judgement and strengthen their cognitive musculature, leaving them atrophied and unprepared when the exceptions do arise." — Microsoft & Carnegie Mellon University, 2025
Around the same time, a study out of MIT's Media Lab used EEG brain monitoring to compare three groups writing essays: those using ChatGPT, those using Google Search, and those using no AI assistance at all. The generative AI group showed the lowest levels of brain engagement across 32 measured regions — and consistently underperformed at neural, linguistic, and behavioural levels. The concern is especially acute for younger adults, whose reasoning and metacognitive habits are still forming.
A large peer-reviewed study by Swiss researcher Michael Gerlich, published in the journal Societies (2025), surveyed 666 participants across diverse age groups and educational backgrounds. Using standardized critical thinking assessments and in-depth interviews, it found a strong negative correlation between frequent AI tool use and critical thinking ability — mediated by what researchers call cognitive offloading: the habit of delegating mental work to an external system rather than engaging with it directly. Younger participants (ages 17–25) showed the highest AI dependence and the lowest critical thinking scores. Higher education served as a partial buffer.
In November 2025, the Harvard Gazette convened several of the university's leading researchers to examine the question head-on. The consensus was nuanced but consistent: AI can be a genuine thinking partner, but only when it augments rather than replaces cognitive effort. Harvard Kennedy School's Dan Levy put it plainly — the problem isn't AI per se, it's using it to do the work for you rather than with you. Learning requires active brain engagement. The output is a vehicle; the thinking is the destination.
Harvard's Karen Thornber drew an analogy most of us feel viscerally: just as GPS navigation has left many of us incapable of reading the streets of our own cities, LLMs may be systematically degrading the mental skills we most rely on. The difference is stakes: spatial memory is one thing; critical reasoning, ethical judgment, and the capacity for independent thought are foundational to everything else.
Cognitive Offloading: The Mechanism
The common thread across these studies is cognitive offloading — the practice of externalizing mental work to a tool. This isn't new; we've always used tools to extend our capabilities. The question is what happens when the tool is so capable, so frictionless, and so available that we stop doing the cognitive work entirely, not just occasionally.
When you struggle through a problem, you build a model of it. You encounter your own misconceptions and correct them. You develop heuristics and patterns that transfer to the next problem. When an AI resolves the problem for you cleanly and immediately, you receive the answer but skip the model-building. Over thousands of interactions, that difference compounds.
The researchers note that this effect is especially sharp among those with high confidence in AI systems. The less you doubt the tool, the less you engage with its output critically — and the more you delegate, the less you practice the very judgment needed to evaluate what you're delegating.
The Design Problem
Here is where this stops being a story about research findings and becomes a story about choices.
Most mainstream AI products are optimized for engagement and retention. They are designed to be maximally useful in the shortest possible time — to resolve your query completely, so that you return satisfied and return again soon. This is the logic of extraction: maximize interaction, minimize friction, capture attention. It is the same logic that shaped social media, and we are watching it play out at scale with AI.
But the Microsoft/CMU researchers themselves pointed toward a different possibility: "their discovery can help design AI tools that emphasize opportunities to practice critical thinking in an effort to stimulate development and prevent atrophy." This is not a fringe idea. It's simply one that doesn't serve engagement metrics.
What would an AI designed for long-term human benefit look like? It might sometimes respond with a question rather than an answer. It might surface the reasoning behind a conclusion rather than just the conclusion. It might notice when a user is developing a skill and offer friction rather than remove it. It might be honest about uncertainty in ways that invite the human to fill the gap. It might, in short, be more interested in what the human becomes through the interaction than in how satisfying the interaction feels in the moment.
"It depends on how we engage with it: as a crutch or a tool for growth." — ChatGPT, when asked by Harvard Gazette researchers whether AI can make us dumber
There's something quietly remarkable about that answer. The tool itself diagnosed the problem correctly. What it can't do is change its own incentive structure. That's a design and governance decision — made by the humans building these systems.
The Case for Restorative AI
AI is one of the most consequential tools humanity has ever built. Like any tool, it can be used restoratively or extractively — to build or to diminish. A hammer can build a house or break a window. The difference isn't the hammer.
The opportunity in front of us — and the one we're most interested in at Giesen Labs — is to build AI that treats human capability as something to compound, not consume. An AI companion that helps a child develop the habit of asking better questions is worth more to the long run than one that answers every question perfectly. An AI that helps a parent reason through a decision — rather than making the decision for them — leaves that parent more capable than it found them.
This is harder to build. It's harder to measure. It doesn't optimize cleanly for session length or NPS. But it is the design philosophy that follows directly from taking the research seriously.
The studies being published now are not a reason to fear AI. They're a specification document for how it should be built differently. The question for every AI product team isn't just "is this useful?" — it's "useful toward what?"
Convenience that compounds humanity is worth building. Convenience that hollows it out is not.