Christopher Ketcham’s Los Angeles Times opinion — a brisk, apocalyptic meditation that frames AI as the next evolutionary step in making humans intellectually lazier — crystallizes a fear that has migrated from elite theory to everyday worry: the machines that augment our thinking will, if unregulated and unexamined, hollow it out. The piece stitches together a personal anecdote about an e‑biker, references Nicholas Carr’s influential critique of the internet’s cognitive effects, and cites several recent empirical studies that suggest heavy use of large language models (LLMs) like ChatGPT correlates with reduced neural engagement, lower performance on writing tasks, and measurable declines in critical thinking. Those studies are real, their findings headline‑worthy, and their implications profound — but the story is more complicated than a straightforward march toward “stupidity.” This feature examines the evidence, verifies the claims, assesses strengths and weaknesses, and lays out the practical risks and policy choices that follow.
In his 2010 book The Shallows, Nicholas Carr argued that the internet’s constant interruptions and incentives for speed rewire attention and reduce capacity for sustained, deep thought. That argument has become a standard frame for critiques of digital culture. Carr’s thesis — that environment shapes neural pathways, favoring quick switching over prolonged concentration — remains influential in popular and academic debates about technology and cognition.
The past two years have seen a dramatic shift. Generative AI tools that can write, summarize, code, plan, and answer questions in natural language moved from novelty to mass adoption. OpenAI’s ChatGPT alone reported explosive user growth in early 2025, with major outlets documenting a leap to hundreds of millions of weekly users in a matter of months. Reuters reported 400 million weekly active users in February 2025; subsequent company statements and media reports placed weekly usage in the high hundreds of millions by spring and summer 2025. These adoption curves transformed what had been intermittent tool use into system-level dependence for many students, professionals, and consumers.
At the same time, a cluster of empirical studies and preprints emerged that explicitly measure how LLM assistance affects cognitive engagement, learning, and critical thinking. The headlines — “Your Brain on ChatGPT,” “Generative AI inhibits critical thinking,” and “GenAI users underperform neurally and linguistically” — are accurate summaries of those papers’ central claims, but they require careful unpacking. The studies use diverse methods (EEG, surveys, self‑report, performance scoring), populations (students, knowledge workers), and durations (from single sessions to multi‑month protocols). Taken together they raise urgent questions; treated as definitive proof of irreversible cognitive decline, though, they overreach.
Strengths of the MIT work:
Strengths:
What the evidence does not yet establish:
The Microsoft–Carnegie Mellon paper’s conclusion that interface and trust signals influence whether people engage critically is a crucial counterweight to pure market logic: product defaults matter. Designers can choose defaults that require confirmation or expose uncertainty rather than hiding it. The choice is not technical inevitability but design and business strategy.
But “stupidity” is too crude a verdict. The evidence shows altered cognitive strategies and behavioral dependencies that can be mitigated by design, policy, and pedagogy. Calling for a return to pre‑digital life is neither realistic nor desirable; instead, the imperative is deliberate integration. Design systems that require accountability; teach the skills of stewardship and verification; and fund the careful, large‑scale science needed to understand long‑term effects.
Unchecked, AI risks becoming the e‑bike of the mind: attractive, efficient, and ultimately weakening if used as total replacement. Used well — with active supervision, curricular adaptation, and interface constraints that promote learning rather than convenience — AI can instead be a sophisticated trainer: a tool that amplifies human capacities while preserving the mental exercises that sustain them.
Community conversation about these trade‑offs is already happening in forums and workplaces, where practitioners and users debate how best to balance convenience and competence. Those local conversations matter because the design choices made today — by product teams, educational institutions, and IT managers — will determine whether AI promotes flourishing or erosion of human cognitive skills. The research gives clear direction: prioritize explainability, verification, and pedagogy. The rest is a design problem we still have the capacity to solve.
Source: Los Angeles Times Contributor: The internet made us stupid. AI promises to make it worse
In his 2010 book The Shallows, Nicholas Carr argued that the internet’s constant interruptions and incentives for speed rewire attention and reduce capacity for sustained, deep thought. That argument has become a standard frame for critiques of digital culture. Carr’s thesis — that environment shapes neural pathways, favoring quick switching over prolonged concentration — remains influential in popular and academic debates about technology and cognition. The past two years have seen a dramatic shift. Generative AI tools that can write, summarize, code, plan, and answer questions in natural language moved from novelty to mass adoption. OpenAI’s ChatGPT alone reported explosive user growth in early 2025, with major outlets documenting a leap to hundreds of millions of weekly users in a matter of months. Reuters reported 400 million weekly active users in February 2025; subsequent company statements and media reports placed weekly usage in the high hundreds of millions by spring and summer 2025. These adoption curves transformed what had been intermittent tool use into system-level dependence for many students, professionals, and consumers.
At the same time, a cluster of empirical studies and preprints emerged that explicitly measure how LLM assistance affects cognitive engagement, learning, and critical thinking. The headlines — “Your Brain on ChatGPT,” “Generative AI inhibits critical thinking,” and “GenAI users underperform neurally and linguistically” — are accurate summaries of those papers’ central claims, but they require careful unpacking. The studies use diverse methods (EEG, surveys, self‑report, performance scoring), populations (students, knowledge workers), and durations (from single sessions to multi‑month protocols). Taken together they raise urgent questions; treated as definitive proof of irreversible cognitive decline, though, they overreach.
The studies the op‑ed relies on — what they actually found
MIT Media Lab — “Your Brain on ChatGPT” (arXiv preprint; Media Lab project page)
Two connected outputs from a Media Lab team reported EEG and behavioral evidence from an essay‑writing experiment. Participants were split into three groups: Brain‑only (no external tools), Search Engine (allowed to use Google), and LLM users (allowed to use ChatGPT). Across sessions the researchers measured EEG connectivity and linguistic/behavioral outcomes. The headline result: the Brain‑only group showed the most distributed neural networks and highest engagement; Search Engine users were intermediate; LLM‑assisted writers showed the weakest EEG connectivity and lower linguistic/behavioral performance. Over a four‑month sequence, LLM users continued to underperform on neural, linguistic, and behavioral measures relative to the other groups. The project page and the arXiv preprint present these as preliminary findings and explicitly note the study’s limitations (small sample size, geographic concentration, model specificity, EEG spatial limits). The Media Lab cautioned that the paper had not yet been peer‑reviewed at the time of posting and that the findings should be treated as early evidence requiring replication.Strengths of the MIT work:
- Direct neural measures (EEG) tied to a concrete learning task (essay writing) provide an objective signal beyond self‑report.
- A longitudinal element (multiple sessions over months) lets the study observe accumulation effects rather than a single snapshot.
- Small sample (54 participants initially, 18 in the final reassignment session) limits statistical power and generalizability.
- The study used one LLM implementation (ChatGPT at the time), so outcomes may differ across models or user interfaces.
- EEG connectivity differences indicate altered neural strategies, but mapping connectivity directly to "stupidity" or permanent skill loss is an inferential step that requires more evidence.
Microsoft Research + Carnegie Mellon — CHI 2025 survey of knowledge workers
A collaborative CHI paper surveyed 319 knowledge workers and analyzed 936 first‑hand examples of generative AI use in workplace tasks. The quantitative and qualitative results indicate that reliance on generative AI alters the nature of critical thinking: higher confidence in AI correlates with less enacted critical thinking, while higher user self‑confidence predicts more engagement and oversight. The paper emphasizes changes in task stewardship (verification, integration) rather than outright disappearance of judgment. Designers are urged to build features that nudge verification, offer explainability, and calibrate assistance according to user skill.Strengths:
- Large sample for a workplace study and direct, real‑world task examples.
- Balanced interpretation: gen‑AI reduces cognitive load in some routine areas but increases the need for verification and stewardship.
- Self‑reported behaviors and examples are subject to reporting bias.
- The study measures changes in the expression of critical thinking (what people do) more than underlying cognitive capacity (what the brain can do).
Other corroborating reporting and surveys
- Multiple news outlets and education surveys report massive uptake of AI tools among students worldwide. In the U.K., major surveys found dramatic year‑over‑year increases in student AI usage and a large share of students now using genAI in coursework. These surveys document behavioral change (tool use) but vary in measurement and scope.
- Media coverage and company statements corroborate explosive adoption of ChatGPT and other LLMs across demographics, further amplifying the potential for widespread cognitive offloading.
What the evidence does and doesn’t prove
The strongest and most defensible claim from the body of evidence is narrow: frequent, heavy, and unquestioning reliance on LLMs changes how people perform specific tasks and can reduce observable engagement during those tasks. Across methodologies, use of LLMs is associated with:- Reduced overt cognitive effort on the target task (less searching, less incremental drafting).
- Behavioral patterns consistent with “cognitive offloading” — delegating synthesis and composition to an external system.
- Changes in the type of skills exercised: verification, integration, and stewardship become more important than raw generation.
What the evidence does not yet establish:
- Permanent, irreversible neurological decline caused by AI use. EEG connectivity differences show altered engagement during specific tasks; they do not by themselves demonstrate permanent atrophy of cognitive capacities.
- Uniform effects across ages, tasks, or models. Several studies note that younger users or less confident users may be particularly prone to passive use, but the magnitude of effects varies and context matters.
- That LLMs cannot be designed to teach or scaffold learning. Design choices — prompting structure, forced justification steps, graded assistance — can convert AI into training wheels rather than crutches.
Critical analysis: strengths, blind spots, and risks
What researchers and critics get right
- Mechanism clarity: The papers are explicit about mechanisms (cognitive offloading, reduced engagement), not merely alarmism. They map observable behaviors to plausible cognitive processes. That methodological transparency strengthens the warnings.
- Design focus: Several authors call for product design that supports critical thinking: explainable outputs, provenance signals, built‑in verification prompts, and graduated assistance. These are practical, actionable mitigations.
- Population sensitivity: Studies identify vulnerable groups (novices, less confident users, students) and note that effects are not uniform — a necessary nuance that counters blanket claims of “AI makes everyone stupider.”
Blind spots and weaknesses
- Causality vs. correlation: Survey work (including the CHI paper) relies on self‑report and cross‑sectional associations. While plausible causal mechanisms exist, stronger causal claims require randomized, larger‑scale interventions and replication.
- Small samples in neural work: The MIT EEG study provides provocative neural evidence, but small n and unblinded designs (participants know whether they’re using tools) complicate interpretation.
- Model and interface variance: LLMs evolve rapidly; a study that uses ChatGPT in March 2025 may not generalize to a substantially different model or a future UI that forces verification.
- Overextension in popular discourse: Op‑eds that leap from short‑term behavioral change to dystopian biological elimination (e.g., Darwinian purges) mix metaphor with empirical claims in ways that must be flagged as speculative.
Notable social and educational risks
- Skill atrophy in routine tasks: If students and workers rely on AI for core steps (idea generation, structuring arguments), training in those foundational skills may decline.
- Erosion of epistemic habits: Accepting AI outputs without verification can normalize lax source evaluation and reduce resistance to misinformation.
- Equity implications: When AI becomes a proxy for instruction or scaffolding, unequal access and differential teacher training could widen achievement gaps.
- Credential and assessment challenges: Universities and certification bodies face a new burden: designing assessments that measure durable competence, not tool‑assisted performance.
Design and policy levers: how to avoid the “muscle atrophy” scenario
The studies point to practical mitigations; they make clear that technology need not be destiny. Converting AI from a crutch into a coach is feasible if product designers, educators, and employers adopt deliberate constraints and scaffolds.- Build accountability into workflows
- Force explanation steps: require students/workers to justify AI outputs in their own words.
- Incremental assistance: permit AI for ideation but not for final composition, or reduce assistance as user competence grows.
- Design for verification and provenance
- Surface source attributions and confidence intervals.
- Offer traceable chains of reasoning so users must evaluate intermediate steps.
- Train users in stewardship skills
- Integrate verification and prompt literacy into curricula and workplace onboarding.
- Teach how to interrogate model outputs: cross‑check, re‑prompt, and triangulate.
- Reconfigure assessment
- Emphasize oral examinations, in‑person problem solving, and multi‑stage projects that capture process over product.
- Use AI‑aware rubrics that value critique and integration.
- Regulatory and institutional steps
- Encourage transparency standards for model capabilities and limitations.
- Fund large‑scale longitudinal research into cognitive effects across ages and socioeconomic groups.
The corporate angle: incentives that favor convenience
A sober reading of the landscape shows why dependence proliferates. Platform providers have commercial incentives to maximize usage and remove friction. Bundling AI into ubiquitous tools, surfacing one‑click completions, or monetizing premium “copilot” features encourages users to accept outputs rather than interrogate them. Market incentives push toward convenience — which can be valuable but also increases the risk of habitual offloading.The Microsoft–Carnegie Mellon paper’s conclusion that interface and trust signals influence whether people engage critically is a crucial counterweight to pure market logic: product defaults matter. Designers can choose defaults that require confirmation or expose uncertainty rather than hiding it. The choice is not technical inevitability but design and business strategy.
Practical guidance for Windows users, educators, and IT managers
- For individual users:
- Use LLMs to accelerate low‑level tasks (summarization, formatting) but reserve complex synthesis for unaided work.
- Practice “active prompting”: ask the model to reveal its reasoning, then evaluate it.
- Periodically complete tasks offline to keep the “muscles” exercised.
- For educators:
- Redraw assessments to prioritize process: drafts, in‑class reflections, and oral defense.
- Teach prompt literacy and verification as core competencies.
- Pilot scaffolded AI tools that require students to annotate how they used AI.
- For IT and product teams:
- Offer AI features that are opt‑in and transparent about data handling.
- Implement audit trails for AI assistance to support academic integrity and accountability.
- Collaborate with pedagogy teams to deploy AI features that scaffold rather than replace learning.
Where the evidence needs to go next
The current research program is promising but early. Priority next steps include:- Larger, multi‑site randomized trials that vary model, interface, and task to establish causal effects and boundary conditions.
- Longitudinal studies spanning years to test whether short‑term engagement differences accumulate into persistent skill changes.
- Exploration of remedial design: can AI be configured to teach critical thinking rather than supplant it? Controlled trials of such “pedagogical AI” are essential.
- Equity‑focused research to test differential effects across socioeconomic and educational backgrounds.
Conclusion
Christopher Ketcham’s LA Times piece captures a visceral truth: tools that make thinking easier also create the temptation to stop thinking. The recent body of research reinforces that temptation’s reality — generative AI is changing how people perform cognitive tasks, and heavy reliance can reduce engagement and maneuver the locus of skill from human generation to machine supervision.But “stupidity” is too crude a verdict. The evidence shows altered cognitive strategies and behavioral dependencies that can be mitigated by design, policy, and pedagogy. Calling for a return to pre‑digital life is neither realistic nor desirable; instead, the imperative is deliberate integration. Design systems that require accountability; teach the skills of stewardship and verification; and fund the careful, large‑scale science needed to understand long‑term effects.
Unchecked, AI risks becoming the e‑bike of the mind: attractive, efficient, and ultimately weakening if used as total replacement. Used well — with active supervision, curricular adaptation, and interface constraints that promote learning rather than convenience — AI can instead be a sophisticated trainer: a tool that amplifies human capacities while preserving the mental exercises that sustain them.
Community conversation about these trade‑offs is already happening in forums and workplaces, where practitioners and users debate how best to balance convenience and competence. Those local conversations matter because the design choices made today — by product teams, educational institutions, and IT managers — will determine whether AI promotes flourishing or erosion of human cognitive skills. The research gives clear direction: prioritize explainability, verification, and pedagogy. The rest is a design problem we still have the capacity to solve.
Source: Los Angeles Times Contributor: The internet made us stupid. AI promises to make it worse