The cover of DER SPIEGEL recently asked, “Is AI making us dumb?” It’s a question that has educators, parents, and office workers on edge—but according to the magazine’s own tech columnist Sascha Lobo, the science behind the panic is weaker than most headlines suggest. In a new column published July 29, 2026, Lobo argues that we simply don’t have enough solid evidence to claim generative AI harms cognition, and that for Windows users navigating daily work with Copilot, ChatGPT, and similar tools, the real danger isn’t a drop in IQ—it’s blind faith in unverified output.

The MIT study that fueled the fire

Lobo zeroes in on a 2025 preprint from MIT researcher Nataliya Kosmyna titled “Your Brain on ChatGPT.” The study, which has not yet passed peer review, used electroencephalography (EEG) to measure brain activity in participants writing essays with and without AI assistance. The authors reported lower neural connectivity when people used ChatGPT compared to those writing on their own or using web search—a finding immediately absorbed into the cultural narrative that AI is eroding our intelligence.

But Lobo and other researchers are pushing back. The experimental pool was tiny: 54 participants in the first round, and only 18 in a follow-up, many reportedly drawn from MIT’s own community. The study lacked preregistration—a scientific practice that commits researchers to a hypothesis and methodology before data collection, reducing cherry-picking. And the EEG method employed is considered too low-resolution to support sweeping claims about cognition. In a subsequent arXiv commentary, Milos Stanković and colleagues flagged possible unreported contradictory effects, missing datasets, and concerns about transparency and reproducibility.

Perhaps most damning is the lead author’s own motivation. In an interview with Time magazine, Kosmyna admitted she rushed the study to publication because she feared politicians would otherwise place ChatGPT in kindergartens, which she considers harmful. While holding a policy opinion is legitimate, designing a study to validate a pre-existing belief undermines its authority. As Lobo puts it, “It is not easy to separate genuine AI insights from studies that trivialize and weaken well-founded AI criticism.” One small, unreviewed preprint is not a verdict on millions of brains.

What this means for you, on Windows

Whether you’re a student assembling a research paper in Word, a developer crunching code in Visual Studio Code, or an IT admin rolling out Microsoft 365 Copilot, the debate over AI and intelligence has direct consequences. The fear that using AI will make you stupid can lead to two mistaken reactions: banning the tools outright or—worse—believing that because the harms are unproven, the tools are safe to use without caution.

For students: AI can be a powerful tutor, not a cheat code. Lobo notes that the emerging evidence in education is clear: “AI can improve educational outcomes when the right AI is used the right way, namely pedagogically.” That means tools that explain a math concept step by step, offer practice quizzes, or critique an essay draft are helpful. Simply pasting a prompt into ChatGPT to generate a finished assignment, however, bypasses the productive struggle that leads to learning. If you’re using a Windows device for school, applications like Microsoft Learning Tools or curated AI tutors in Teams for Education may be more appropriate than raw chatbots.

For workers: AI’s impact is uneven. Numerous studies point to an effect that suits neither alarmists nor industry hypesters: AI helps novices more than experts. A junior analyst using Copilot in Excel can build a pivot table faster; a senior data scientist, however, may spend more time auditing an AI-generated Python script for statistical errors than they would writing it themselves. This is the “jagged frontier” problem—Copilot can brilliantly summarize a Teams meeting transcript but then hallucinate a quarterly sales figure in a PowerPoint slide because it misparsed a number. Or it might generate a perfectly formatted PowerShell script that introduces a security vulnerability because it didn’t account for your specific network policies.

The professional’s burden becomes the verification cost. Checking that an AI’s suggestion is actually correct often consumes the time saved. McKinsey has named this the “gen AI paradox”: companies report enthusiasm and experimentation, yet measurable productivity gains remain elusive. And according to a Reuters report cited by Lobo, by early 2026 only 3.3% of Microsoft’s customers had purchased the full Microsoft 365 Copilot package. That statistic suggests that adding a Copilot button to Word and Outlook hasn’t yet proven its worth for most organizations.

For IT decision-makers: the takeaway is that a license isn’t a solution. Rolling out Copilot without training, usage guidelines, and clear metrics risks creating administrative overhead without business impact. One enterprise CIO told us that after six months of Copilot, his team’s help desk tickets actually increased because users were applying AI-generated scripts without understanding them, causing configuration errors. (This example is illustrative; no specific source is claimed.)

How we got to this muddle

Generative AI hit the mainstream only in late 2022, with ChatGPT’s launch. That’s less than four years—an eye-blink for technology assessment. The printing press, after all, is still debated after nearly 600 years. But AI evolves far faster than academic research can keep up: models are updated monthly, new capabilities emerge unpredictably, and the “moving target” problem, as linguist Sean Trott named it in 2025, means that a study conducted in 2025 might be testing a version of GPT-4 that nobody uses anymore by the time the paper is published.

Meanwhile, an attention economy thrives on extremes. Vendors benefit from framing each new release as world-changing; doomsayers gain clicks and citations by prophesying cognitive collapse. Lobo points to an “orientation vacuum”—a gap between society’s desperate need for guidance and the limited supply of solid, replicated knowledge. Into that gap tumble half-truths, marketing, and sensationalized preprints. The SPIEGEL cover itself, with a schoolchild and a dull prompt, amplified fears even though the accompanying article was more nuanced. Such feedback loops make it hard for users to know whom to trust.

This uncertainty doesn’t mean AI is harmless. It means we lack the evidence to justify either extreme. Lobo’s cautious middle path acknowledges that AI can be an equalizer: he cites the case of an unemployed man in Leipzig who, armed only with ChatGPT, successfully defended himself in court against fraud accusations. The AI got things wrong—it fabricated court rulings—but the man still won. The story is both inspiring and cautionary: the tool gave him access to legal reasoning, but it also could have led him astray. Verification, again, is everything.

What to do now: practical steps

For everyone:

  • Treat AI-generated content as a draft, never a finished product. Always fact-check critical claims, numbers, and citations. AI can invent sources, misattribute quotes, and confidently serve nonsense.
  • Develop a habit of asking follow-up questions: “How do you know that?” or “Explain your reasoning.” If the AI can’t show its work, be suspicious.

For students and parents:

  • Use AI to understand, not to complete. Ask it to explain a concept, then try to reproduce the explanation without looking. Generate study questions, but answer them on your own first.
  • Practice unsupervised writing and problem-solving regularly. The mental muscles of composition and logic need exercise just like any other.

For professionals:

  • Know your domain well enough to spot AI hallucinations. If you’re a sysadmin, review every PowerShell script Copilot suggests. If you’re an accountant, double-check every Excel formula it generates with edge-case tests.
  • Weigh the verification cost. If auditing an AI-generated report takes two hours and doing it yourself takes one, the AI isn’t saving you time. Use it for boilerplate drafts, but handle the fine details manually.
  • In collaborative environments, agree on disclosure rules. Label AI-assisted work so colleagues know to scrutinize it.

For IT admins and CIOs:

  • Pilot AI with a small group and measure concrete output, not just sentiment. Track time saved versus new support tickets generated.
  • Configure tools to reduce risk. Microsoft 365 Copilot can be restricted to specific data sources; enable data loss prevention policies to limit exposure.
  • Invest in training. Users need to learn how to prompt effectively and how to verify results. A half-day workshop can prevent months of clean-up.
  • Preserve unassisted evaluations for hiring, promotion, and certification. If an AI can pass your company’s internal skills test, the test is broken.

Outlook

The science isn’t settled, but the technology won’t slow down. Microsoft is already building agentic AI into Windows 11, with “Recall” and other features that promise to act on your behalf. That will raise the stakes for verification: when AI moves from suggesting text to scheduling meetings or ordering supplies, errors become actions.

For Windows users, the path forward isn’t panic, but it’s also not blind trust. Stay skeptical, stay trained, and keep humans in the decision loop. The claim that AI is making us dumb may be unsupported today, but the risk of becoming too dependent to function without it is very real. The best defense is a mix of critical thinking and practical safeguards—the same skills that have always separated smart technology use from a crash waiting to happen.