Key Takeaways

• Cognitive atrophy is the gradual weakening of memory, critical thinking, and analytical capacity that happens when those mental functions go unused.
• Generative AI accelerates this pattern because it can complete entire reasoning tasks on a person's behalf, which goes further than older tools like search engines that still required users to evaluate and assemble information themselves.
• Peer-reviewed research now links heavy AI reliance to measurably lower cognitive engagement, weaker recall, and reduced critical thinking, with the steepest effects among younger and more frequent users.
• For organizations, the real risk is that AI adoption quietly erodes original thinking while output still looks polished, which is why measuring Original Intelligence with Hupchecker gives leaders a way to see who is expanding their thinking with AI and who is offloading too much of it.

What is cognitive atrophy?

Cognitive atrophy is the decline of cognitive abilities, including memory, problem-solving, independent reasoning, and critical analysis, that follows from prolonged disuse or under-stimulation. The word pairs "cognitive," meaning the mental processes of knowing and understanding, with "atrophy," from the Greek for wasting away. Clinically, brain atrophy refers to the physical loss of neurons and the connections between them. In the context of technology, researchers use the term more broadly to describe the functional dulling of mental capacities when we stop exercising them.

The mechanism behind it is cognitive offloading: delegating mental work to an external tool. Offloading is not harmful by itself. Writing a grocery list, using a calculator, or saving a phone number all free up working memory, the limited mental workspace we use to hold and manipulate information in the moment, for more demanding work. The distinction that matters is what gets offloaded. Offloading a fact you would have looked up anyway preserves capacity. Offloading the reasoning itself, the comparing, weighing, and connecting that builds expertise over time, removes the practice that maintains the skill. That second kind of offloading is what drives cognitive atrophy, and it is the kind generative AI is uniquely suited to provide.

This is not a new worry. When search engines became ubiquitous, psychologists documented the "Google effect," also called digital amnesia. A 2011 study in Science by Betsy Sparrow, Jenny Liu, and Daniel Wegner found that when people expect information to remain available online, they remember where to find it rather than the information itself. Memory did not vanish, but what people chose to retain shifted from facts to retrieval paths.

How AI changes the game

Generative AI changes the equation in both degree and speed. A search engine returns sources that a person still has to read, weigh, and synthesize, which keeps some critical thinking in the loop. A large language model can produce a finished memo, analysis, or essay from a single prompt, bypassing the reasoning steps entirely.

That speed is compounding fast. According to Ifop's 2025 survey for Jedha AI School, 89% of French 16-to-25-year-olds have already used a generative AI tool such as ChatGPT, Grok, Gemini, Perplexity, or Claude, compared with 43% of the general public, and a quarter of those young people (25%) are daily users. A tool that can bypass reasoning entirely is now a daily default for a whole generation, at an age when the habits of independent thought are still forming.

That same speed shows up in a different form too. Research on AI and homogenization points to a related effect. A study of college admissions essays found that after generative AI became widespread, applicants' stories grew markedly more similar to one another, and many students came to see AI-generated versions of their own experiences as authentically their own. The research was conducted by Georgetown neuroscientist Adam Green, who also co-founded Hupside, and his broader finding is that diversity of ideas tracks with cognitive ability and stronger real-world outcomes. We see homogenization and atrophy as connected: when individual reasoning gets offloaded, the same offloading tends to pull many people's output toward the same average, since fewer people are left contributing the original judgment that would have pulled it apart. That is Hupside's read on why the two trends show up together, not a claim from the research itself. The original study is covered in more detail in Design Observer, and Hupside has written more on the broader pattern in Exploring AI Homogenization in 2026.

What the research actually shows

Beyond homogenization, a growing body of evidence speaks directly to atrophy itself, though some of it is still early. The 2025 MIT Media Lab study "Your Brain on ChatGPT" (Kosmyna et al.), currently a preprint pending peer review, monitored 54 participants aged 18 to 39 from five Boston-area universities with EEG headsets as they wrote essays. Participants assisted by ChatGPT wrote about 60% faster, but their relevant cognitive load fell by 32%, meaning they were doing measurably less of the active comparing, organizing, and synthesizing that turns information into understanding. Brain connectivity in the alpha and theta bands, frequencies the researchers associate with internal memory retrieval and unguided idea generation, was nearly halved compared with those who wrote unaided. Most strikingly, 83% of the AI group could not quote a single sentence from the essay they had just produced, while nearly everyone in the unaided and search-engine groups could. The researchers termed this pattern "cognitive debt," the idea that the effort saved today is borrowed against the skill you would have built by doing the work yourself.

In a CHI 2025 paper, researchers at Microsoft Research and Carnegie Mellon University surveying 319 knowledge workers found that higher confidence in AI was associated with less critical thinking, and warned that "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."

The pattern is even showing up in how people describe their own work. In Anthropic's December 2025 study of more than 81,000 Claude users across 159 countries, educators were 2.5 to 3 times more likely than average to report having witnessed cognitive atrophy firsthand, meaning skill loss, intellectual passivity, and declining critical thinking, presumably in their students.

What's at stake for organizations and individuals

For individuals, the cost is a slow erosion of the judgment and analytical range that make their contributions valuable. None of this means AI use itself is the problem. Plenty of AI use involves the harmless kind of offloading described earlier, and used well, AI can extend someone's thinking rather than replace it. The risk is concentrated in the cases where people stop doing the reasoning that AI can fake convincingly, and those cases are genuinely hard to tell apart from the productive ones just by looking at the finished work.

That difficulty is exactly what creates the organizational stakes. AI has made polished output cheap and abundant, which means the old signals of capability, credentials, clean deliverables, and confident analysis, no longer reliably prove that original thinking happened behind them. Hupside calls this signal collapse. When output alone cannot tell leaders who is genuinely expanding the idea space and who is recombining what the model already knows, the question stops being about individual effort and becomes a measurement problem at the organizational level.

McKinsey's State of AI survey, published in November 2025 and based on 1,993 respondents across 105 nations, found that 88% of organizations now use AI in at least one business function, up from 78% a year earlier, yet only about 6% qualify as high performers seeing significant financial impact, and just 39% report enterprise-level EBIT impact. McKinsey's survey was not designed to measure original thinking, but the gap between adoption and impact is consistent with a pattern where most teams are getting faster at producing work without anyone tracking whether that work still carries an original contribution. Speed without measurement tends to produce sameness rather than advantage. For a closer look at structuring AI rollout around people rather than tools alone, see How to Put Your People in Position to Succeed With AI.

Measuring Original Intelligence as a constructive response

The answer is not to reject AI or to lower expectations for the people using it. It is to measure what AI cannot replicate. Original Intelligence is the new standard for measuring value beyond AI, defined as the measurable capacity to create value beyond what AI-typical output would produce.

This is what Hupside, the Original Intelligence Infrastructure company, was built to make visible. Hupchecker measures Original Intelligence in your people through a short interactive experience that produces an OIQ (Original Intelligence Quotient) score calibrated against AI baselines, along with an OIQ archetype describing how each person naturally generates and shapes ideas. This is not for labeling purposes, nor is it a general verdict on how good of a worker someone is, but rather allows leaders to see what style of thinking each person brings to a team. 

With that data, leaders can build targeted training, smarter team composition, and role alignment around what they actually see, rather than what polished output suggests. Cognitive atrophy is a real and measurable shift, and the organizations that respond by measuring original value, rather than assuming it exists because the work looks good, are the ones that will keep it. Take the OIQ Challenge to see where your own Original Intelligence sits.

‍