DER SPIEGEL columnist Sascha Lobo is pushing back on the idea that generative AI is demonstrably “making us dumb,” arguing that the evidence is far too young, uneven and often overstated to support the panic now surrounding ChatGPT in schools and workplaces.
Writing in The German View, Lobo’s core point is not that AI is harmless. It is that a technology deployed at mass scale only since late 2022 cannot yet support sweeping claims about intelligence, learning or long-term social damage—especially when the models, products and usage patterns change faster than research can catch up.
Lobo focuses on Nataliya Kosmyna and colleagues’ widely circulated 2025 preprint, “Your Brain on ChatGPT,” which used EEG readings during essay-writing tasks and reported lower neural connectivity among participants using an LLM than among people writing unaided or using search.
The paper has become shorthand for a far broader proposition: that AI use weakens the brain. But it remains a preprint rather than peer-reviewed research, and a subsequent arXiv comment by Milos Stanković and co-authors identified concerns around the small participant pool, transparency, EEG methodology, reporting inconsistencies and reproducibility.
That does not prove the original findings are wrong. It does mean that headlines claiming ChatGPT is making people stupid reach far beyond what a small, task-specific experiment can establish. Lower measured activity while completing one kind of writing task is not a diagnosis of diminished intelligence, and it says little on its own about whether a learner retains knowledge, develops judgment or can work independently later.
The more useful question for Windows users, educators and IT leaders is narrower: which tasks should AI assist, and where must people continue to practice without it?
That distinction matters for organizations rolling out Microsoft 365 Copilot, ChatGPT Enterprise or similar tools. A chatbot that supplies finished answers can reduce the productive struggle involved in learning. A tool configured to explain an error, generate practice problems, ask follow-up questions or critique a draft can support learning—provided the student or employee remains responsible for the work.
For administrators, that points toward practical controls rather than blanket bans:
This is the jagged frontier problem. An assistant may produce an excellent PowerShell explanation, Excel formula or policy draft in one case, then fail badly on a nearly identical request because a hidden detail changes the answer. Professionals with deep domain knowledge are better equipped to spot that failure, but they also pay the cost of auditing it.
McKinsey has described a related “gen AI paradox”: widespread experimentation and deployment without material bottom-line impact for most organizations. Its argument is that generic copilots and chatbots can deliver scattered, hard-to-measure gains, while higher-value, workflow-specific deployments often remain stuck in pilot stages.
Lobo also cites Reuters reporting that only a small fraction of Microsoft’s customer base had paid for the full Microsoft 365 Copilot package by early 2026. Whatever the precise adoption figure, the operational lesson is clear: adding AI buttons to Word, Excel, Outlook and Teams is not the same thing as redesigning a business process around reliable automation.
The evidence so far supports neither a universal productivity turbocharger nor a settled case for cognitive collapse. It supports an uncomfortable middle ground: AI can be highly useful, particularly for people who need guidance, but its value depends on task design, domain knowledge, verification and the discipline to keep humans accountable for consequential decisions.
Writing in The German View, Lobo’s core point is not that AI is harmless. It is that a technology deployed at mass scale only since late 2022 cannot yet support sweeping claims about intelligence, learning or long-term social damage—especially when the models, products and usage patterns change faster than research can catch up.
One viral MIT preprint is not a verdict on cognition
Lobo focuses on Nataliya Kosmyna and colleagues’ widely circulated 2025 preprint, “Your Brain on ChatGPT,” which used EEG readings during essay-writing tasks and reported lower neural connectivity among participants using an LLM than among people writing unaided or using search.The paper has become shorthand for a far broader proposition: that AI use weakens the brain. But it remains a preprint rather than peer-reviewed research, and a subsequent arXiv comment by Milos Stanković and co-authors identified concerns around the small participant pool, transparency, EEG methodology, reporting inconsistencies and reproducibility.
That does not prove the original findings are wrong. It does mean that headlines claiming ChatGPT is making people stupid reach far beyond what a small, task-specific experiment can establish. Lower measured activity while completing one kind of writing task is not a diagnosis of diminished intelligence, and it says little on its own about whether a learner retains knowledge, develops judgment or can work independently later.
The more useful question for Windows users, educators and IT leaders is narrower: which tasks should AI assist, and where must people continue to practice without it?
Assistance helps most when it is designed as instruction
Lobo argues that the emerging education picture is conditional rather than apocalyptic. AI can improve outcomes when it is used in a structured, pedagogical way; simply dropping a general-purpose chatbot into a classroom is unlikely to do much beyond make copying easier.That distinction matters for organizations rolling out Microsoft 365 Copilot, ChatGPT Enterprise or similar tools. A chatbot that supplies finished answers can reduce the productive struggle involved in learning. A tool configured to explain an error, generate practice problems, ask follow-up questions or critique a draft can support learning—provided the student or employee remains responsible for the work.
For administrators, that points toward practical controls rather than blanket bans:
- Require disclosure and review for AI-assisted work where accuracy matters.
- Build training around prompting, source checking and recognizing fabricated output.
- Preserve unaided assessments and hands-on exercises for foundational skills.
- Treat AI-generated content as a draft or recommendation, not an authoritative result.
Enterprise AI still has a verification problem
The column’s strongest workplace point is that productivity effects are not evenly distributed. Less experienced workers can gain substantial help from AI in well-defined tasks, while seasoned specialists can encounter a different problem: the output looks plausible enough to demand careful verification, and that checking cost can erase the apparent time saving.This is the jagged frontier problem. An assistant may produce an excellent PowerShell explanation, Excel formula or policy draft in one case, then fail badly on a nearly identical request because a hidden detail changes the answer. Professionals with deep domain knowledge are better equipped to spot that failure, but they also pay the cost of auditing it.
McKinsey has described a related “gen AI paradox”: widespread experimentation and deployment without material bottom-line impact for most organizations. Its argument is that generic copilots and chatbots can deliver scattered, hard-to-measure gains, while higher-value, workflow-specific deployments often remain stuck in pilot stages.
Lobo also cites Reuters reporting that only a small fraction of Microsoft’s customer base had paid for the full Microsoft 365 Copilot package by early 2026. Whatever the precise adoption figure, the operational lesson is clear: adding AI buttons to Word, Excel, Outlook and Teams is not the same thing as redesigning a business process around reliable automation.
The useful skepticism is procedural
The AI debate is crowded with incentives to exaggerate. Vendors benefit from presenting every new model as transformative; critics, researchers and media outlets can benefit from framing each concerning result as a civilizational warning. Neither impulse helps a school deciding how to assess students or an IT department deciding whether Copilot deserves another license renewal.The evidence so far supports neither a universal productivity turbocharger nor a settled case for cognitive collapse. It supports an uncomfortable middle ground: AI can be highly useful, particularly for people who need guidance, but its value depends on task design, domain knowledge, verification and the discipline to keep humans accountable for consequential decisions.
References
- Primary source: DER SPIEGEL - The German View
Published: 2026-07-29T08:23:05+00:00
Loading…
derspiegel.substack.com