Is AI Making Us Dumber?

There is a thought experiment worth sitting with. Imagine a first-year engineering student in 2026. Before she understands what a differential equation represents — what it feels like to struggle with one, to stare at it, to fail, and eventually to break through — she has already learned to type it into an AI assistant and receive a worked solution. Before a junior solicitor has memorised the structure of a tortious claim, he has learned to prompt a legal AI for the framework. Before a new medical resident has developed the clinical intuition to recognise sepsis from a constellation of subtle signs, her hospital's AI system has already flagged the alert.

The work gets done. The patient is treated. The brief is filed. The calculation is submitted. But something else has also happened — or rather, failed to happen. The hard, frustrating, neurologically formative process of struggling with difficult material, getting it wrong, and building a mental model through effort has been quietly bypassed. The question this article confronts is a serious one: at scale, and over time, is artificial intelligence routing around the very cognitive processes that make expertise possible?

This is not a Luddite argument. AI is a genuinely transformative technology, and many of its educational and productivity applications are real and valuable. The Harvard physics study we discuss below demonstrated that a well-designed AI tutor can double learning gains. The question is not whether AI is good or bad. The question is whether we are using it in ways that systematically undermine the development of human judgment — and whether we have thought through what that means for the next generation of professionals.

How We Actually Learn: Why the Struggle Matters

To understand what AI risks short-circuiting, it helps to understand what cognitive scientists know about skill acquisition. The dominant model in educational psychology, developed through decades of research by scholars including K. Anders Ericsson, Robert Bjork, and John Sweller, is that durable learning requires desirable difficulty. That is, learning proceeds fastest — and sticks most reliably — when the learner is operating at the edge of their current ability, making errors, and being forced to retrieve and reconstruct knowledge rather than simply recognising it.

Robert Bjork's concept of desirable difficulties — techniques like spaced retrieval practice, interleaving problem types, and reducing immediate feedback — produce slower apparent progress in the short term but dramatically superior long-term retention and transfer. The mechanisms are neurological: when you struggle to retrieve a memory, the hippocampus strengthens the encoding. When you are given the answer, the encoding pathway is bypassed.

This is the fundamental tension with AI assistance. When a student asks an AI to solve a maths problem, explain a concept, or draft an argument, the product is generated outside the student's working memory. There is no effortful retrieval. There is no error, no correction, no neural reinforcement of the correct pathway. The student receives an output that may be excellent. They learn almost nothing.

"We distinguish between the experience of learning, which should feel challenging and even uncomfortable, and the experience of being informed, which feels effortless. AI tools are extraordinarily good at informing. They may be actively harmful as learning tools when used to replace effortful practice." — Frontiers in Psychology, The Cognitive Paradox of AI in Education: Between Enhancement and Erosion, 2025

This distinction matters enormously in professional formation. The junior lawyer who reads fifty contracts doesn't just learn what contracts contain — they develop a pattern-recognition system for anomalies, unusual clauses, and risks that cannot be easily articulated or transferred. The junior engineer who manually calculates dozens of drainage designs doesn't just get the numbers — they develop an intuitive sense of when a result is plausible, a professional instinct that is the foundation of responsible engineering judgment. These capacities are not separable from the struggle of acquiring them.

The Evidence: What Research Is Finding

The empirical literature on AI and cognition is still young — generative AI tools only reached mass adoption in 2022–2023 — but a clear and consistent pattern is emerging across multiple independent research programmes.

The MIT Media Lab EEG Study (2025)

Perhaps the most striking study to date comes from MIT's Media Lab, where researchers fitted 54 students with electroencephalography (EEG) headsets and monitored their neural activity during writing tasks over four months. Students were divided into three groups: one used ChatGPT, one used search engines, and one wrote without AI assistance.

The results were stark. Students who used ChatGPT showed significantly weaker brain connectivity compared to the other groups — the kind of neural patterning associated with reduced memory consolidation and lower-order processing. Most alarmingly, 83% of the AI-assisted group could not accurately recall passages from essays they had just written. The essays existed as text; the learning had not occurred.

The most troubling finding was the persistence of the effect. Even after students stopped using AI, their brain activity remained sluggish relative to baseline — suggesting that the cognitive habits formed during AI use may not be easily reversed. It is worth noting that this study is small (54 participants) and has not yet completed peer review, and should be treated as preliminary; however, it is consistent with the broader direction of the literature.

The MDPI / Society Studies on Cognitive Offloading (2025)

A broader review published in the MDPI journal Society synthesised evidence on cognitive offloading — the phenomenon of outsourcing mental work to external tools — and found a significant negative correlation between frequent AI tool usage and critical thinking abilities, with the relationship mediated by increased cognitive offloading. Younger participants showed higher AI dependence and lower critical thinking scores than older participants, consistent with developmental vulnerability during the years when professional reasoning skills are being formed.

The University of Technology Sydney Report (2026)

A white paper from UTS published in early 2026 warned specifically of "cognitive atrophy" in school-age learners where AI use is unstructured. The paper argued that consistent offloading of cognitive work to AI systems during critical developmental windows creates a dependency that progressively reduces the learner's capacity for independent reasoning — not through any dramatic failure, but through a gradual, asymptomatic erosion of mental muscle.

The Positive Case: Harvard Physics (2024)

Not all the evidence is negative, and it is important to be honest about this. A well-designed randomised controlled trial at Harvard, led by Gregory Kestin and Kelly Miller, found that students using a custom AI tutor in an introductory physics class achieved approximately double the learning gains of students in standard active-learning lectures — in less time. The key design choice was that the AI (called "PS2 Pal") was explicitly programmed not to provide direct answers. It guided students one step at a time, asked questions, and required the student to do the cognitive work. The AI functioned as a Socratic tutor, not an answer machine.

This finding is critically important. It demonstrates that AI can enhance learning — but only when it is designed to require cognitive engagement rather than replace it. The Harvard result is a reproach not to AI in education but to the manner in which most AI is currently used in education.

Summary of Key Studies

We Have Been Here Before: The GPS Precedent

For those who find the AI-cognition research alarming, there is both comfort and warning in a directly analogous story that played out over the preceding two decades: the adoption of GPS navigation.

A peer-reviewed study published in Scientific Reports in 2020 examined the relationship between habitual GPS use and spatial memory, using a large longitudinal dataset. The findings were unequivocal: people with greater lifetime GPS experience show worse spatial memory during self-guided navigation. More significantly, greater GPS use since initial testing was associated with a steeper decline in hippocampal-dependent spatial memory over time. The mechanism is straightforward — GPS-based navigation removes the cognitive requirement of building a mental map. Because the external tool does the work, the brain's spatial processing circuits receive less activation, and their capacity gradually diminishes.

This is cognitive offloading made manifest in the hippocampus. It is not hypothetical. It is documented, longitudinal, and neurological. The GPS analogy is instructive in another way: we accepted the trade-off largely without discussion. Almost nobody argues that we should ban GPS. But we also did not, as a society, think carefully about which cognitive capacities it was eroding or design compensating practices into education systems.

We are making the same mistake with AI, at far greater scale and across far more cognitively central domains.

The Writing Skills Crisis in Higher Education

Writing is not merely a communication skill. It is a thinking technology — the process of committing ideas to text forces clarification, exposes logical gaps, and generates the kind of slow, deliberate reasoning that complex problem-solving requires. When we outsource writing to AI, we do not merely save time. We lose the cognitive product that writing produces.

The scale of AI adoption in student writing is now beyond dispute. A 2025 survey by the UK's Higher Education Policy Institute found that 92% of students used AI in some capacity, up from 66% the previous year. A large majority reported using it for essay drafts, literature summaries, and argument construction. A systematic review in ScienceDirect (2026) on ChatGPT's cognitive impact in higher education found that "students who regularly delegated essay composition to AI showed significantly weaker performance on timed analytical writing tasks where AI was unavailable" — suggesting that the skill is not developing even as AI-assisted output improves.

The Brookings Institution, in a widely cited 2025 analysis, argued directly that "AI has rendered traditional writing skills obsolete" and that "education needs to adapt." But there is a meaningful distinction between adapting assessment frameworks and accepting that the underlying cognitive capacity need no longer be developed. The ability to construct a coherent, evidence-grounded, logically sequenced argument is not merely an assessment box to be ticked. It is the fundamental cognitive product of a university education — and arguably of professional life.

"They can do B-plus work all the time. The worry is not that they are writing badly. The worry is that they are not developing the capacity to think at all." — University of Edinburgh writing instructor, quoted in Tandfonline / Taylor & Francis: AI & the End of College Writing as We've Known It, 2025

The problem is compounded by what researchers have called "AI-induced cognitive dissonance": students know, at some level, that the work is not theirs, but they also feel productive and capable. Over time, the gap between their perceived competence and their actual independent capability grows — often invisibly, until a high-stakes moment (a professional presentation, a complex client deliverable, a technical examination) exposes the deficit.

The Overconfidence Trap: A Reverse Dunning-Kruger Effect

Among the most troubling findings in the recent literature is research from Aalto University suggesting that AI use is generating a systematic and paradoxical overconfidence in users' own capabilities.

The classic Dunning-Kruger effect holds that people with limited knowledge in a domain tend to overestimate their competence, while genuine experts tend toward under-confidence. AI appears to invert this. Researchers at Aalto found that AI-literate users — those most experienced with AI tools — were the most likely to overestimate their own independent performance. The mechanism is cognitive: using AI to produce high-quality output generates a feeling of competence that is not tethered to any genuine capability. The student feels like a good writer because the AI produces good writing under their direction. The junior engineer feels competent because the AI's structural calculations are sound. The confidence is real; the underlying skill may not be.

A separate study published in TechXplore (October 2025) confirmed this pattern: AI use makes users overestimate their cognitive performance on tasks completed with AI assistance, and the overestimation does not diminish with experience — it grows. Users become more confident the more they use AI, even as their independent performance remains static or declines.

This has direct professional safety implications. A structural engineer who cannot sanity-check an AI-generated load calculation is not merely less capable — they are a liability. A doctor who cannot form an independent differential diagnosis when the AI system gives an implausible result is a patient safety risk. A lawyer whose legal reasoning has never been formed through effortful practice cannot recognise when the AI's case summary contains a material error. The competence illusion is not merely an academic problem. It is a safety and ethics problem.

The Vanishing Apprenticeship: AI and the Entry-Level Crisis

Professional formation in most fields has always relied on a structured progression: entry-level roles provide the foundational experience — the tedious debugging, the routine client correspondence, the standard contract drafts, the routine site inspections — through which young professionals develop the pattern recognition, professional judgment, and cultural understanding that cannot be taught in a classroom. This apprenticeship model, whether formalised or informal, is how expertise transfers from one generation to the next.

AI is disrupting this model at its base. The tasks that were routinely assigned to graduates and junior staff — precisely because they were instructive, even if not complex — are now increasingly automated. And the hiring decisions that follow reflect this.

The Numbers

The data on entry-level hiring in AI-exposed sectors are stark:

• Entry-level tech job postings fell 60% between 2022 and 2024, according to LinkedIn data reported by CNBC.
• At the fifteen largest tech companies globally, fresh graduate hiring fell more than 50% over three years, with entry-level hiring at those firms down 25% from 2023 to 2024 alone.
• The unemployment rate for recent US computer science graduates reached 6.1% in 2024, significantly above the national average and up sharply from historical norms.
• Internships across all industries fell 11% year-over-year, with tech-specific internships down 30% since 2023 (Handshake data).
• A 2024 LeadDev survey found 54% of engineering leaders planned to hire fewer junior engineers due to AI copilot productivity gains among senior staff.
• In a 2024 survey, 37% of hiring managers said they would rather "hire" AI than take on a recent graduate, and 57% said they trusted AI's output more than that of an intern.

Salesforce CEO Marc Benioff confirmed in late 2024 that the company would not be hiring any new software engineers in 2025, citing a 30% productivity increase from its internal use of AI tools. While Salesforce later reversed course — hiring 1,000 graduates and interns in a high-profile announcement — the initial decision reflected a calculation that many other companies have quietly executed without announcement: if one senior engineer with AI assistance can now produce the output of three, why maintain a junior cohort?

The Structural Problem

The answer to that question — one that many organisations are failing to ask — is that junior roles are not merely cheap labour. They are the mechanism by which senior talent is produced. The five-to-seven-year development pipeline from graduate intake to independent senior professional cannot be compressed or skipped. It can only be deferred — at which point it becomes a crisis.

"Entry-level, junior-level roles are the breeding ground for the leadership of the future. If you overcut that junior layer, you will have a talent bottleneck at some point in the business that leads invariably to an increase in hiring costs." — Quoted in CNBC, In Defense of Junior Staff: Why Replacing Young People with AI Could Spark a 'Talent Doom Cycle', November 2025

The CNBC reporting on the "talent doom cycle" captures the dynamic precisely: companies that eliminate junior hiring to capture short-term AI productivity gains will find themselves, in five to seven years, unable to staff senior roles from internal progression. They will then be forced to compete in an external senior talent market that is itself constrained, because their competitors made the same decision. Hiring costs surge, institutional knowledge is lost, and the very productivity advantage that justified the original cuts is eroded by the cost of external recruitment.

AI Weekly reported in early 2026 that analysts are projecting a significant senior talent shortage beginning around 2031 as a direct consequence of junior hiring freezes since 2023 — precisely the five-to-seven-year pipeline timeline playing out.

Case Studies Across Professions

The Microsoft Warning: When Industry Sounds the Alarm

It is one thing for academics and educators to raise concerns about AI and cognitive development. It is another thing entirely when the people building AI tools issue warnings about the professional pipeline they are inadvertently destroying.

In early 2026, Microsoft Azure CTO Mark Russinovich and VP of developer community Scott Hanselman published a peer-reviewed opinion piece in Communications of the ACM — one of computing's most prestigious journals — arguing that agentic AI coding tools are creating a structural crisis in the software engineering profession. The core argument was clear: AI gives senior engineers a massive productivity boost (what the authors call an "AI boost") while simultaneously imposing what they term an "AI drag" on early-in-career (EiC) developers who lack the judgment to steer, verify, and integrate AI output effectively.

The result, as the authors described it, is a new incentive structure in which companies hire senior engineers and automate junior functions — while the talent pipeline that produces the next generation of senior engineers quietly collapses.

"We must keep hiring EiC developers, accept that they initially reduce capacity, and deliberately design systems that make their growth an explicit organisational goal." — Mark Russinovich & Scott Hanselman, Communications of the ACM, 2026

Russinovich cited research showing that after GPT-4's release, employment of 22-to-25-year-olds in AI-exposed jobs including software development fell by roughly 13%, even as senior roles grew. Separate data placed the decline in entry-level developer hiring at 67% since 2022. The authors also made a striking intervention on the education side: "You need some classes where using AI is considered cheating." Not because AI is bad, but because there are foundational skills that can only be built through independent struggle — and that struggle must be deliberately protected.

The proposed solution — a "preceptorship" model adapted from medical residency training, in which senior engineers take formal responsibility for mentoring cohorts of junior developers within product teams — reflects a recognition that the apprenticeship pipeline does not regenerate automatically. It must be actively maintained, even at short-term productivity cost, or it will fail.

What Good AI Use Looks Like: The Force Multiplier Case

Intellectual honesty requires engaging with the counter-argument. There is genuine, documented evidence that AI, used appropriately, makes workers more capable rather than less — and that the dichotomy between "AI dependence" and "genuine skill" is not as absolute as the alarm literature implies.

AI copilots demonstrably allow senior engineers to produce more output, handle more complex problems, and document their work more comprehensively than they could without assistance. In a 2023 randomised controlled trial involving professional software developers, GitHub Copilot increased coding speed by up to 55% — with no significant reduction in code quality among experienced users. The RAND Corporation published research in 2024 suggesting that AI writing tools could reduce inequality in professional outcomes, allowing workers with strong ideas but weaker writing skills to produce professional-quality communications.

The critical variable in nearly all the research is the existing skill level of the user. Studies consistently find that AI amplifies what is already there. For a senior developer with twenty years of experience, Copilot is a productivity tool — they can verify its outputs, correct its errors, and integrate it into a workflow built on deep understanding. For a first-year student who has never written a working programme independently, Copilot removes the cognitive struggle that would have created that understanding. The tool is identical; the developmental consequences are opposite.

What Needs to Change: Recommendations for Institutions and Individuals

The challenges outlined in this article are structural, not individual. They require responses at the level of educational institutions, professional bodies, employing organisations, and policy — not merely the choices of individual students or graduates. The following recommendations reflect the emerging consensus in the research literature.

For Educational Institutions

• 1Distinguish explicitly between AI-permitted and AI-prohibited assessments, and design both types thoughtfully. Some assessments should prohibit AI to develop independent capability; others should require AI use to develop AI literacy. The mix matters.
• 2Invest in AI literacy education — not just how to use AI tools, but how to validate their outputs, recognise their failure modes, and maintain independent capability alongside AI use. As of 2025, only 28% of UK graduates reported that their institution "meaningfully integrated" AI into their programmes.
• 3Redesign high-stakes assessments to be impossible to complete by AI delegation alone — oral examinations, live problem-solving sessions, collaborative design critiques, and portfolio defences are all formats that require genuine independent capability.
• 4Take the Harvard model seriously: where AI tutoring tools are deployed, ensure they are designed to require effortful student thinking rather than to provide answers. The tool design is not neutral.

For Employers and Professional Bodies

• 5Resist the temptation to eliminate junior hiring on the basis of short-term AI productivity gains. Model the five-to-seven-year pipeline. The talent doom cycle is a quantifiable risk; price it into workforce strategy.
• 6Formalise mentorship. Russinovich and Hanselman's preceptorship model — in which senior engineers take explicit responsibility for the development of junior cohorts — is a response to the disappearance of informal learning that junior roles previously provided. It should be evaluated seriously across industries.
• 7Define AI-free zones in professional development. Some tasks should be assigned to junior staff precisely because doing them builds capability — not because the result is optimal. A junior engineer who manually checks the drainage calculations they will later rely on AI to produce is building judgment. A junior engineer who skips straight to the AI output is not.
• 8Evaluate junior staff on their reasoning process, not just their output quality. If the output can be generated by AI, it is not a measure of professional formation. Competency frameworks need updating to capture independent analytical capability explicitly.

For Individual Professionals and Students

• 9Develop what researchers are calling "metacognitive AI awareness" — the conscious habit of distinguishing between work you have done and work AI has done for you, and maintaining honest accounting of where your independent capability actually sits.
• 10Treat AI like a senior colleague whose work you must review, not an authority whose output you accept. The skill of validating AI outputs is itself a professional competence — and it requires genuine domain knowledge to exercise.
• 11Deliberately seek out the difficult, foundational tasks in your field that AI could do but that you need to do yourself, at least until you have built the underlying understanding. Write the first draft. Do the manual calculation. Read the primary source. The marginal time cost is an investment in independent capability.

The Uncomfortable Verdict

Sources & Further Reading

All sources cited in this article were verified at the time of publication (May 2026).

• 1 MIT Media Lab EEG Study — ChatGPT weakens cognitive performance: madinamerica.com — ChatGPT Weakens Your Ability to Think, MIT Study Finds (Jul 2025)
• 2 Nextgov/FCW — MIT cognitive decline study: nextgov.com — New MIT study suggests too much AI use could increase cognitive decline (Jul 2025)
• 3 Frontiers in Psychology — The cognitive paradox of AI in education: frontiersin.org — The cognitive paradox of AI in education: between enhancement and erosion (2025)
• 4 MDPI Society — AI tools, cognitive offloading and critical thinking: mdpi.com — AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking (2025)
• 5 University of Technology Sydney — Cognitive atrophy in schools: uts.edu.au — AI, cognitive offloading and implications for education (Mar 2026)
• 6 Harvard Gazette — Professor tailored AI tutor, engagement doubled: news.harvard.edu — Professor tailored AI tutor to physics course. Engagement doubled. (Sep 2024)
• 7 Nature / Scientific Reports — Harvard AI tutoring RCT: nature.com — AI tutoring outperforms in-class active learning: an RCT (2025)
• 8 Harvard Gazette — Is AI dulling our minds?: news.harvard.edu — Is AI dulling our minds? (Nov 2025)
• 9 Aalto University / Live Science — AI causes reverse Dunning-Kruger effect: livescience.com — The more people use AI, the more they overestimate their abilities (2025)
• 10 Inc.com — Science warns AI is causing reverse Dunning-Kruger effect: inc.com — Science Warns That AI Is Causing a 'Reverse Dunning-Kruger Effect' (2025)
• 11 TechXplore — AI use makes us overestimate cognitive performance: techxplore.com — AI use makes us overestimate our cognitive performance, study reveals (Oct 2025)
• 12 Nature / Scientific Reports — GPS use negatively impacts spatial memory: nature.com — Habitual use of GPS negatively impacts spatial memory during self-guided navigation (2020)
• 13 UK Higher Education Policy Institute — 92% of students use AI: k12dive.com — What impact are AI chatbots having on writing skills? (2025, citing HEPI data)
• 14 ScienceDirect — Cognitive impact of ChatGPT in higher education: sciencedirect.com — The cognitive impact of ChatGPT in higher education: a systematic review (2026)
• 15 Brookings Institution — AI has rendered traditional writing skills obsolete: brookings.edu — AI has rendered traditional writing skills obsolete. Education needs to adapt. (2025)
• 16 Tandfonline — AI and the end of college writing: tandfonline.com — AI & the End of College Writing as We've Known it (2025)
• 17 Springer Nature — AI-induced deskilling in medicine: link.springer.com — AI-induced Deskilling in Medicine: A Mixed-Method Review (2025)
• 18 ScienceDirect — AI and risk of deskilling among physicians: sciencedirect.com — AI in medicine: scoping review of risk of deskilling and loss of expertise (2026)
• 19 MDLinx — Doctors relying too much on AI and losing skills: mdlinx.com — Doctors are relying too much on AI — and losing their skills fast (2025)
• 20 NCBI / PMC — Skill decay with AI assistance: ncbi.nlm.nih.gov — Does AI assistance accelerate skill decay and hinder development? (2024)
• 21 CNBC — AI is not just ending entry-level jobs, it's the end of the career ladder: cnbc.com — AI is not just ending entry-level jobs. It's the end of the career ladder (Sep 2025)
• 22 CNBC — Why replacing junior staff with AI could spark talent doom cycle: cnbc.com — In defense of junior staff: Why replacing young people with AI could spark a 'talent doom cycle' (Nov 2025)
• 23 CNBC — Why AI may kill career advancement for young workers: cnbc.com — Why AI may kill career advancement for many young workers (Nov 2025)
• 24 Fortune — AI entry-level jobs uncertainty for college grads: fortune.com — As AI cuts entry-level jobs, young workers are left wondering what's next (Sep 2025)
• 25 Rest of World — AI wiping out entry-level tech jobs, graduates stranded: restofworld.org — AI is wiping out entry-level tech jobs, leaving graduates stranded (2025)
• 26 Stack Overflow Blog — AI vs Gen Z, junior developer pipeline: stackoverflow.blog — AI vs Gen Z: How AI has changed the career pathway for junior developers (Dec 2025)
• 27 IEEE Spectrum — AI shifts expectations for entry-level jobs: spectrum.ieee.org — AI Shifts Expectations for Entry Level Jobs (2025)
• 28 InfoQ — Microsoft Russinovich and Hanselman warn AI hollowing junior pipeline: infoq.com — Microsoft's Russinovich and Hanselman Warn AI Is Hollowing out the Junior Developer Pipeline (Apr 2026)
• 29 The Register — Microsoft execs worry AI will eat entry-level coding jobs: theregister.com — Microsoft execs worry AI will eat entry level coding jobs (Feb 2026)
• 30 DevOps.com — Microsoft executives warn AI could limit developer talent pipeline: devops.com — Microsoft Executives Warn AI Could Limit the Developer Talent Pipeline (2026)
• 31 Visual Studio Magazine — GitHub Copilot research finds downward pressure on code quality: visualstudiomagazine.com — New GitHub Copilot Research Finds 'Downward Pressure on Code Quality' (Jan 2024)
• 32 Knowledge at Wharton — AI pushing us to break the talent pipeline: knowledge.wharton.upenn.edu — Is AI Pushing Us to Break the Talent Pipeline? (2025)
• 33 AI Weekly — Junior dev hiring freeze sets up 2031 senior shortage: aiweekly.co — Junior dev hiring freeze sets up 2031 senior shortage (2026)
• 34 Muse by Clios — AI could be a disaster for junior staff development: musebyclios.com — AI Could Be a Disaster for Junior Staff Development. Here's a Remedy (2025)