Let’s do this.

AI and us · Learn

What’s true about AI and us.

Ten minutes. Nine facts with their sample sizes attached, six words that get blurred, and the places where the research disagrees with itself. Most of what you’ll read about AI and the mind rests on studies of a few dozen people; here’s which ones are bigger.

Do the thing Practice

Nine facts

  1. How you use it decides what it does to you. About a thousand high-school students in Turkey practised maths with a chatbot. Those with unrestricted access scored 17% worse on the exam, once the bot was removed, than students with no bot. Those whose bot gave only hints scored the same as the no-bot group. The students with the answers didn’t notice they’d learned less.1
  2. The same pattern shows up in working adults. Engineers who delegated a task to AI learned less of it than those who worked it through (52 people). Office workers who trusted AI more reported doing less critical thinking of their own (319 people, self-reported).2
  3. The viral “brain scan” study is small and unreviewed. “Your Brain on ChatGPT” is a preprint with 54 participants and 18 in its key session; a comment posted on arXiv, itself unreviewed, disputes its methods. Cite it as a hint, not a finding.3
  4. People can’t tell AI writing from human writing. In one study, they judged AI text to be human 77% of the time. Open-source deepfake detectors lose about half their score on the standard measure (AUC) when tested on fakes found in the wild.4
  5. Deepfakes mostly spread doubt, not belief. In experiments they caused more uncertainty than deception, and that uncertainty lowered trust in news on social media. Labelling true headlines “AI-generated” made people believe them less. In the 2024 UK, French and EU elections, researchers found viral AI disinformation but no evidence it changed a result.45
  6. Engagement-ranked feeds measurably raise anger. Changing them doesn’t change politics much. On Twitter, engagement ranking amplified out-group hostility and anger more than any other emotion, and users rated its political picks lower than a time-ordered feed. In Meta’s three-month experiment, time-ordered feeds cut time on the apps and didn’t move anyone’s polarisation.6
  7. AI floods what’s made more than what’s seen. On one music service, fully AI-generated tracks were over half of daily uploads on peak days in 2026 and one to three percent of streams. Of the first 500 Shorts shown to a new YouTube account, 104 were AI-generated.7
  8. AI companions are rated unacceptable for anyone under 18. Common Sense Media found age gates easy to pass and bots that gave harmful advice and claimed to have feelings. The pediatricians’ advice: talk to your child’s doctor if they seem withdrawn and prefer the bot to people.8
  9. There is no federal AI law. The states are writing them, and a federal task force exists to sue over some of them. The EU’s labelling rules took effect on 2 August 2026; its strictest rules slipped to 2027 and 2028. New York’s law on the largest developers takes effect on 1 January 2027, with penalties cut tenfold on the way. A December 2025 executive order set up a task force to challenge state AI laws; the Senate had voted 99 to 1 against banning them.9

Six words that get blurred

Hints vs answers
The single distinction the classroom trial turned on. A tutor that asks you questions leaves your skill intact. A tool that hands you the output doesn’t.
Training vs retention
Turning off training keeps your future chats out of future models. It doesn’t delete anything, and the company still stores chats for a while (72 hours to several years, depending on the service and what you clicked).
Deepfake vs cheapfake
A deepfake is synthesised. A cheapfake is a real clip, cropped, slowed or captioned to mislead. In the 2024 UK, French and EU elections, viral disinformation leaned on old methods, bot comments and influencers, more than on generative AI.5
Engagement-ranked vs chronological
Ranked by what people click and linger on, or by time. The first is what every major feed defaults to; the second is what the research compared it with.
Content Credentials
A signed record of who made a file and how it was edited, travelling with the file. It proves the file hasn’t changed, not that it’s true. Credentials can be separated from the file or stripped, by accident or on purpose, so their absence proves nothing.12
Frontier model
The handful of largest models from the largest developers. New York’s and California’s new laws apply to these, not to the apps built on them.

Where the sources disagree

Is AI making people dumber? Nobody has long-term data. The best trial says it depends entirely on the mode of use; one 2026 preprint found a gain from tutoring-style use that still held a week later. The dramatic claims rest on tiny samples.13

Do the meta-analyses show ChatGPT helps students learn? They disagree. A 2026 pooling of 35 studies finds a moderate average gain. A widely cited 2025 meta-analysis that found a large gain was retracted in April 2026. Earlier, a preprint re-analysing its data found no evidence of a benefit once publication bias was accounted for, while the 2026 pooling found no significant bias in its own studies. A 2026 preprint that pools 67 meta-analyses of education studies puts the bias-adjusted effect at about a third of the unadjusted one. Don’t quote “meta-analyses show” either way.11

Do algorithmic feeds polarise us? They measurably raise anger and hostility in the moment (preregistered audits). The one large field experiment on switching feeds found no change in political attitudes after three months, though critics note the platform changed its own systems during the study.6

Is regulation being lobbied away? Weakened and delayed, often: EU high-risk rules pushed back, Colorado’s law replaced before it started, New York’s penalties cut. Erased, rarely: the labelling rules started on time, a dozen states passed chatbot-safety laws in 2026, and the attempt to ban state laws lost 99 to 1.9

Does inner work change what gets built? No study tests it. The evidence on mindfulness, time offline and time with people is about your own wellbeing: modest for mindfulness across 47 trials, real in one two-week trial of blocking mobile internet, and strong for social ties. The claim that a builder’s inner state shapes the product is an argument, not a finding.10

Read the originals

Next

Do the thing: 30 minutes Practice: the daily habit

Sources

  1. Bastani et al., Generative AI without guardrails can harm learning, PNAS 122(26), 2025 (about 1,000 students; open access via Europe PMC) research; Contractor and Reyes, preprint, October 2026 preprint.
  2. Anthropic, AI assistance and coding skills, 29 January 2026 (52 engineers; the company’s own trial, not peer-reviewed) industry research; Lee et al., The impact of generative AI on critical thinking, CHI 2025 (319 workers, self-reported) research.
  3. Kosmyna et al., Your Brain on ChatGPT, preprint, 2025, not peer-reviewed; the comment disputing its methods, arXiv 2601.00856, December 2025, is also unreviewed. preprint
  4. International AI Safety Report 2026, 3 February 2026 (an expert report, not peer-reviewed) primary; Deepfake-Eval-2024, preprint, v5 May 2026 preprint; Altay and Gilardi, People are skeptical of headlines labeled as AI-generated, PNAS Nexus, October 2024 (4,976 people) research.
  5. Vaccari and Chadwick, Social Media + Society, 2020 (abstract read) research; Schiff, Schiff and Bueno, American Political Science Review 119(1), 2025, online 2024 (the “liar’s dividend”; five experiments, 15,000+ adults) research; Alan Turing Institute CETaS, 2024 elections review, September 2024 (16 viral AI cases in the UK election, 11 in the EU and French; “no evidence that AI-enabled disinformation or deepfakes meaningfully impacted” results) primary.
  6. Milli et al., Engagement, user satisfaction and divisive content, PNAS Nexus, March 2025 (806 users, preregistered) research; Guess et al., How do social media feed algorithms affect attitudes?, Science 2023 research; Piccardi et al., reranking an X feed, Science, November 2025 (1,256 users) research.
  7. Deezer, 2026 statements on AI uploads and streams company’s own figures; Kapwing, YouTube Shorts audit, October–November 2025 (one account, manual coding) industry research.
  8. Common Sense Media, AI companions decoded, 30 April 2025 primary; American Academy of Pediatrics, How AI chatbots affect kids: benefits, risks and what parents need to know, 27 August 2025 primary. Common Sense’s institute says it is funded by philanthropy and by industry, including makers of some products it rates.
  9. European Commission, AI Act page, updated 19 October 2026 primary; New York Senate, S8828 (RAISE Act amendment, chapter 96 of 2026) primary; White House, Executive Order 14365, 11 December 2025 primary; Inside Privacy on the Senate’s 99–1 vote, July 2025 news; the earlier RAISE penalties ($10M/$30M) are in the signed bill, S6953B, chapter 699 of 2025 primary; Arnold & Porter, Colorado narrows its AI law, July 2026 (SB 26-189 replaced the 2024 act before it took effect) news; Multistate, state AI companion chatbot laws, 2026 (twelve states by mid-year) industry research.
  10. Goyal et al., Meditation programs for psychological stress and well-being, JAMA Internal Medicine 2014 (47 trials); Castelo et al., Blocking mobile internet on smartphones, PNAS Nexus, February 2025; Holt-Lunstad et al., Social relationships and mortality risk, PLOS Medicine 2010 (148 studies). research
  11. Wang and Fan, the 2025 meta-analysis, Humanities and Social Sciences Communications, retracted 22 April 2026 research; Wu et al., Humanities and Social Sciences Communications 13, 2026 (35 experimental studies, no significant publication bias found, 4,193 participants; the authors call the effect “moderately positive”; abstract and bias tests read) research; Bartoš, Martinková and Wagenmakers, PsyArXiv, May 2025, a re-analysis of Wang and Fan’s data (once publication bias is accounted for, “the evidence in favor of the benefits disappears”; abstract read) preprint; Bartoš, Bujak, Martinková and Wagenmakers, PsyArXiv, January 2026, a pooling of 1,840 effect sizes from 67 meta-analyses (bias-adjusted effect about 0.2 standard deviations, credible interval 0.0 to 0.3; abstract read) preprint.
  12. C2PA, Content Credentials FAQ, read 6 Nov 2026 (credentials “can be separated”; metadata can be “accidentally or intentionally stripped”). primary
Contest a fact, suggest something, or tell us what’s broken

If someone is in danger right now, don’t use this form. Call your local emergency number (911 in the US). For trafficking: 1-888-373-7888, or text 233733.

What is it?

No name, no account. Your note goes into the site’s private records, nothing about your browser or connection is stored with it, and the email box is the only way we’d know who you are. The weekly review reads every note, re-reads the original source, and fixes the page if it should. A contested fact and its outcome are listed on Contested, in the review’s words, never yours.

No tracking on any link.Ticks stay on this device.Facts checked 5–6 Nov 2026.What changedContested