
How AI Makes People Dumber
Title history
- How AI Makes People Dumber — observed on 20261112-0400 UTC (YouTube)
Keywords
Summary
196 words
Critical Evaluation
Value of the Information & Strength of the Argument
The video provides significant value by compiling a broad range of studies across different contexts (education, professional work, medicine, neuroscience) that collectively point to a concerning pattern. It carefully distinguishes between correlation and causation, noting limitations of individual studies. The argumentation is balanced, presenting counterarguments and nuanced interpretations, such as the distinction between specific skills and general cognitive capacities. The emphasis on the ‘how’ of AI use (e.g., tutor vs. direct answer) offers a constructive angle. However, some studies are preliminary or rely on self-report, and the video could more explicitly address potential publication bias or industry-funded research.
Scientific Rigor, Source Quality, Title Accuracy
The video cites numerous studies with dates, institutions, and journal names, enhancing credibility. It correctly identifies that some findings are observational or not yet peer-reviewed. The title accurately reflects the central question, and the content stays on-topic. While the video does not provide direct URLs or links, the references are sufficiently detailed for viewers to locate them. The production appears well-researched, though a few names are misspelled in the transcript (e.g., ‘Carnegie Melon’ instead of ‘Carnegie Mellon’). Overall, the sources are high-quality and appropriately contextualized.
198 words
Title / Content Match
The title directly matches the content, which examines whether AI reduces human cognitive abilities, presenting both evidence and nuance.
Quality & Reliability
8/10
The video synthesizes multiple peer-reviewed and preliminary studies with balanced caveats, though some research is not yet peer-reviewed and effects are largely correlational.
Key Moments
Markers derived by PSI from the transcript: the creator did not define chapters.
- Opening study: fraction problems and chatbot removal leads to worse performance and increased skipping.
- Historical context: Socrates' warning about writing, and how similar fears recurred with each technology.
- Satellite navigation study at McGill: heavy app users show worse spatial memory and hippocampal decline.
- MIT Media Lab brain-activity study: ChatGPT group shows up to 55% lower brain connectivity during essay writing.
- Microsoft and Carnegie Mellon survey: higher confidence in AI correlates with less critical thinking; workers produce less diverse results.
- Wharton study in Turkey: students using standard GPT-4 scored 17% worse on exam than those with textbook, while tutor version showed no harm.
- Large-scale Chinese study: AI use initially improves homework but reduces exam scores by 18-24% after years.
- Anthropic coding experiment: AI-assisted engineers scored lower on learning quiz, but conceptual-use group fared better.
- Counterarguments: Trends in Cognitive Sciences review argues basic cognitive capacities are resilient, but specific skills matter.
- Places a positive spin: tutor-style AI, asking conceptual questions, and thinking first protect against decline; NYC announces moratorium.
Cited Sources
- Grace Louu and Brian Christian fraction study (2026) — Three experiments with 1,222 online workers showing lower persistence and accuracy after AI removal.
- Natalia Cosmina, 'Your Brain on ChatGPT' (MIT Media Lab, 2025) — EEG study of 54 young adults showing reduced brain connectivity when writing with ChatGPT.
- Microsoft Research and Carnegie Mellon survey (2025) — Survey of 319 knowledge workers about AI use and critical thinking, collecting 936 examples.
- Michael Gerlick, 'Societies' journal (2025) — Survey and testing of 666 UK participants; correlation of -0.68 between AI use and critical thinking.
- Hamsa Bastani et al., PNAS (2025) — Randomized experiment in Turkish high school with ~1,000 students using GPT-4 or tutor version.
- Stockholm University and University of Hong Kong study (2026) — Followed 26,811 students in China; AI adoption led to 18-24% lower college entrance exam scores.
- Brookings Institution premortem (2026) — Report based on interviews and review of 400+ studies; risks of AI in schools outweighed benefits.
- Anthropic randomized trial (2026) — 52 junior engineers learning Python; AI group scored 50% vs 67% without, no speed advantage.
- Meter nonprofit study (2025) — 16 experienced developers, 246 tasks; AI made them 19% slower, yet they believed it made them faster.
- Lancet Gastroenterology and Hepatology study (2025) — Doctors at 4 Polish centers; AI assistance reduced detection of precancerous growths from 28.4% to 22.4% when used without AI.
- Trends in Cognitive Sciences review (2026) — Cash et al. argue basic cognitive capacities are resilient; only task-specific skills decline.
- Pew Research survey (2026) — 54% of American students use AI for schoolwork; 12% use it for every assignment.
- College Board survey (2026) — 84% of high schoolers use AI for homework.
- New York City moratorium (September 2026) — Mayor announced one-year ban on student-facing generative AI in public schools.
Concurring Sources
- Grace Louu and Brian Christian fraction study (2026) — Shows immediate drop in performance and persistence after AI removal.
- Natalia Cosmina's MIT brain study (2025) — Demonstrates reduced brain connectivity when writing with ChatGPT.
- Microsoft Research and Carnegie Mellon survey (2025) — Finds less critical thinking with more reliance on AI.
- Michael Gerlick's Societies paper (2025) — Correlates high AI use with lower critical thinking scores.
- Wharton Turkey experiment (2025) — Shows that standard chatbot use impairs later exam performance.
- Chinese large-scale study (2026) — Documents substantial exam score declines after years of AI use.
- Brookings premortem (2026) — Reports teacher and student narratives of cognitive decline and dependency.
- Anthropic randomized trial (2026) — Indicates slower learning and lower quiz scores for AI-assisted engineers.
- Polish colonoscopy study (2025) — Shows reduced detection ability after AI tool use, indicating deskilling.
Dissenting Sources
- Trends in Cognitive Sciences review (2026) — Argues that basic cognitive capacities like working memory and attention are resilient to task-specific offloading, suggesting the concern may be overstated.
- Historical argument (Socrates, printing press, calculator) — Past fears of technology-induced cognitive decline often did not materialize or were outweighed by adaptive gains; society adapted.
- Meter study paradox — While showing slower performance, it also reveals that developers believed they were faster, hinting at complex perceptual effects that might not reflect real cognitive decline.
Contribution & Novelties
The video uniquely aggregates recent 2025-2026 research across multiple fields, highlighting both the universal mechanism of cognitive offloading and the moderating role of usage patterns. It offers a nuanced view that AI is not inherently harmful but that certain uses (delegation vs. conceptual engagement) lead to very different outcomes. It connects laboratory neuroscience with real-world educational and professional data, and provides actionable insights for individuals and policymakers.
Pour aller plus loin :
- Cognitive offloading — Central concept explored in the video; Wikipedia page explains the phenomenon of relying on external tools to reduce mental effort.
- Digital amnesia — The tendency to forget information easily when it is readily accessible via digital devices; directly related to the ‘digitally induced amnesia’ mentioned in the Brookings report.
- Critical thinking — The core skill measured in several studies; Wikipedia provides definition and assessment methods.
- Hippocampus — The brain region implicated in spatial memory decline from navigation app use; Wikipedia article details its role in memory and mapping.
163 words
Radar Profile
The scores indicate high quantity and quality of information with moderate technical depth. Reliability is strong given the balanced presentation of evidence and limitations. The video excels in breadth and nuance, making it a valuable synthesis for general audiences.