AI and us · Do

Make it your advisor, not your dictator.

In crisis, or worried about someone? Call 988 or text 988, any time. What 988 does.

Three habits, no cost: think first and ask for hints, not answers; find the original before you share; take a break from the feed, then keep it on Following. How you use AI decides whether you get sharper or duller. Students who let a chatbot hand them answers scored 17% worse once it was taken away; students given a version that only offered hints lost nothing.¹ Every action here carries a label for how well it’s proven.

What’s true Keep it going

The whole thing

1. Think first, then ask for hints, not answers.

   Write your own rough pass before you open the chat. Then ask it to question you, or to argue against you. moderate Four trials point the same way: one of about a thousand students, one of 194, one across nine Nigerian schools, and one of 211 retested a week later. Used as a tutor, AI helped or did no harm. Used to do the work, it cost students once it was gone, or left them no lasting gain.¹⁴ The habits.

   A habit · $0

2. Find the original before you share.

   You can’t tell a fake by looking. Before you repeat or forward anything striking, find where it first appeared and ask yourself whether it’s accurate. moderate Twenty experiments with 26,863 people found that a pause for accuracy cuts people’s stated intent to share false headlines by about a tenth. After six class sessions on “lateral reading”, 61% of students checked a source elsewhere and judged it right on at least one of four problems, up from 6% before.⁵ How to check.

   A habit · $0

3. Take a break from the feed, then keep it on Following.

   Two to four weeks off, then switch every feed to the people you follow. moderate: with a cost Four weeks off Facebook raised wellbeing a little and cut news knowledge a little; two weeks without mobile internet improved attention and mood; a time-ordered feed cut time on the apps and, on average, left political attitudes where they were.⁶ How to set it up.

   An afternoon to set up · $0

Pick the day and time you’ll do step one:

Near you

Enter your ZIP code and this page adds the steps that depend on where you live: your state’s own programs and laws, and the finders that take a ZIP. The code stays in your browser, like the ticks.

The habits

• Ask for questions, not answers, when you’re learning something. moderate The hint-only tutor in the classroom trial did no harm; the answer machine cost 17%.¹
• Write first. Bullet points are enough. weak: one 18-person session In one small unreviewed study, people who drafted on their own before using a chatbot remembered more of their own work.³
• Make it argue against you. none but plausible: untested Office workers who trusted the AI more said they did less critical thinking of their own; the survey didn’t test whether asking for objections fixes that.²
• Check every fact you’ll repeat. Open the cited page and find the sentence; if it isn’t there, don’t repeat the claim. moderate The accuracy-pause evidence above.⁵
• Keep one skill AI-free. First drafts, sums, a kind of code, a drawing. weak: mechanism only The harm in the trials appeared when the crutch was removed; keeping one skill unassisted is the obvious response, untested as such.¹
• Never run someone’s writing through an AI detector. moderate: evidence of harm Seven detectors flagged, on average, 61% of essays by non-native English writers as AI-written. Don’t use one on a student, an employee or a friend.¹³

The settings

Turning these off keeps your future chats out of future training. It doesn’t remove anything from models already trained, and companies still keep chats for a while.⁷ none but plausible: privacy only

• Claude: Settings → Privacy (claude.ai/settings/data-privacy-controls) → the model-training toggle, which Anthropic’s help pages call “Model Improvement” → off. Incognito chats are never used for training. Thumbs-up or thumbs-down feedback keeps that whole chat for five years.⁷
• ChatGPT: profile icon → Settings → Data controls → Improve the model for everyone → off. For a one-off private chat, use Temporary Chat. OpenAI’s page blocked our check, so confirm the label.⁷
• Gemini: myactivity.google.com/product/gemini → Keep Activity → Turn off. Chats are still kept 72 hours; chats people already reviewed, up to three years.⁷
• Meta AI (Facebook, Instagram, WhatsApp): since 16 December 2025 Meta uses your Meta AI chats to personalise content and ads; its announcement offers no opt-out, only the existing ad-preference controls.⁸ Privacy guides say there’s no US opt-out from training on public posts either; Meta’s own page wouldn’t load for us. The only levers are not using it and keeping posts private.

Don’t put anything into a chatbot you wouldn’t put in an email to the company that runs it.

Telling what’s real

• Don’t trust your eyes. strong Across 56 papers, people spotted deepfakes about 55% of the time, close to a coin toss. In one study, people took AI-written text for human 77% of the time. Open-source detectors lose about half their score on the standard measure (AUC) when tested on fakes collected online.¹¹
• Find the original. moderate Reverse-search the image and look for the oldest copy and where it first appeared. Read what other sources say about the site before you read the site.⁵
• Don’t swing to “it’s all fake”. strong Calling real things fake is its own harm: in two experiments with 4,976 people, an “AI-generated” label made true, human-written headlines less believed.¹²
• Check Content Credentials. none but plausible: can be stripped Drag an image or video into verify.contentauthenticity.org. If it carries credentials you’ll see who signed it and how it was edited. A credential proves the file hasn’t changed, not that it’s true. Credentials can be stripped from a file, by accident or on purpose, so no credential doesn’t mean fake.⁹
• Ask Gemini whether Google made it. none but plausible: Google’s output only Upload an image, video or audio clip and ask “Was this created with Google AI?” It checks for Google’s SynthID watermark, so other companies’ AI output won’t be flagged.¹⁰
• Agree a family safe word tonight. none but plausible: official advice A cloned voice can sound like your kid. The FBI’s advice: a secret word or phrase. The FTC’s: hang up and call them back on a number you already know.¹⁴

Your feeds

Switching to a feed of only the people you follow gives you your time back. Don’t expect more: in Meta’s three-month experiment, time-ordered feeds cut time on the apps but didn’t change political attitudes on average, and a 35,000-person study of leaving Facebook and Instagram before the 2020 election found polarisation effects near zero.⁶ Engagement-ranked feeds do push anger up, and in the study that showed it, users liked the algorithm’s political picks less than the time-ordered ones.⁶

• Instagram: open the Feed dropdown at the top → Following: the last 30 days from accounts you follow, in time order.
• Facebook: the Feeds tab → Friends or Favorites, newest first.
• X: Following is ranked by default now. On the web, click the arrow next to Following → Sort by → Most recent. X’s help pages blocked our check; confirm in the app.
• YouTube: the Subscriptions tab. Turning off and clearing watch history empties the recommendations on Home.
• TikTok: the Following tab. TikTok’s help pages blocked our check of whether a non-personalised feed exists in the US.

Feed menus change often; these came from guides and help pages and may have moved. Look for the same words in your app.

If you have a kid

Common Sense Media rates AI companion apps like Character.AI and Replika “unacceptable” for anyone under 18: age gates were easy to get past, and the bots gave harmful advice and claimed to have feelings. Its research institute says it is funded by philanthropy and by industry, including makers of some of the products it rates.¹⁵ weak: expert judgment

• Ask which chatbots they use, including the ones inside apps they already have: Snapchat’s My AI, Meta AI in Instagram and WhatsApp.
• Keep chatbot use in a shared room for younger kids, and read a few chats together.¹⁶
• Say it out loud: bots are built to keep you talking and to tell you what you want to hear. Private things go to real people.
• The pediatricians’ advice: talk to your child’s doctor if they seem withdrawn and prefer the bot to people. They warn that kids turn to bots when frustrated with the people in their lives, and that bots may say false, harmful, highly sexual or violent things.¹⁶
• Delay the smartphone, with other parents. none but plausible: no studies The Wait Until 8th pledge activates once ten families in a grade at a school sign, so nobody’s kid is the only one without. Its site cites studies of phone harms, but none of the pledge itself, and we found none.¹⁸
• Back a phone-free school day. moderate: contested In England, test scores rose after schools banned phones, most for the weakest students, but only where the ban was kept. In Norway, bans had no average effect on grades or mental health, but they cut bullying, and girls used less mental-health care and got better grades. Two other studies found no clear gain in wellbeing; one of them found no difference in attainment either.¹⁷

Inner work, honestly

The evidence here is about your own wellbeing and attention; it doesn’t test whether a builder’s practice changes what they build.

• An eight-week mindfulness course. moderate Across 47 trials, it eased anxiety, depression and pain modestly, though no better than exercise, medication or therapy.¹⁹
• Two weeks without mobile internet (calls and texts still work). weak: one trial Wellbeing, mental health and measured attention all improved, though only a quarter of people managed ten days.¹⁹
• Time with real people. moderate: association, not a trial Across 148 studies, people with stronger social ties had 50% higher odds of survival.¹⁹

Ask for one thing

There’s no federal AI law; the states are writing them. Pick one live bill in your state and write your legislator in your own words: one ask, and a line about why it matters to you. moderate In a field experiment, state legislators who got a few dozen calls from constituents about a bill were about 12 percentage points more likely to support it. In a 2010 survey, about 20% of congressional staff said an individual letter or email would have “a lot of influence” on an undecided member; 1% said so of a form message.²⁰

• Find your state legislators with Plural, a company’s free lookup tool that takes your address, or on your state legislature’s own site. Federal: house.gov and senate.gov.
• What’s live, as of 5 November 2026: New York’s RAISE Act on the largest developers and its law on AI companions for minors both take effect on 1 January 2027, as does Colorado’s replacement AI law. The EU’s labelling rules apply since 2 August 2026, with 2 December 2026 as the deadline for generative systems already on the market. A December 2025 executive order set up a federal task force to challenge state AI laws; the Senate had earlier voted 99 to 1 against a ten-year ban on them.²¹

What happens after

• After you turn off training in Claude, Anthropic says it won’t use your past or new chats to train future models. Chats already in a training run, or in a model already trained, stay there. A chat you delete leaves its servers within 30 days. Chats its systems flag for breaking its rules are kept up to two years.⁷
• After you turn off Keep Activity in Gemini, your future chats aren’t used for training unless you send feedback. Each chat is still kept for 72 hours. Chats human reviewers already saw are kept up to three years, even if you delete your activity. Some features stop working while it’s off, such as Google Workspace apps.⁷ OpenAI’s help page refused our reader, so we can’t say what ChatGPT keeps.
• After Gemini checks an image, a SynthID watermark means Google’s AI made or edited all or part of it. No watermark means Google’s AI didn’t make it; another company’s AI still could have. Sometimes the answer is unclear, as with very simple images or very small edits. You get about 10 image checks in any 24 hours.²⁴
• After you check Content Credentials, you see which company signed the information and the history it records. If the file or its record changed after signing, the check shows that. It doesn’t tell you whether the picture is real, and the name shown may not be the owner’s.⁹²⁴
• After a break from the feed, people who left Facebook for four weeks in 2018 reported more wellbeing, knew less factual news and were less polarised. Afterwards they used Facebook a lot less, the drop lasted, and they valued it less. In a bigger six-week study before the 2020 election, the effect on polarisation was close to zero, and people may have known less general news. In the mobile-internet study, wellbeing and mental health were still better two weeks after the block ended, and screen time stayed lower. Only a quarter of the people who signed up kept the block on for ten of the fourteen days.⁶
• After you sign the Wait Until 8th pledge, nothing is made public. Within one to two business days you get an email with tips for recruiting other parents. Once ten families from your child’s grade and school sign, the pledge is active: you all get one group email with the list of families and the parents’ emails, then updates when more join. Until then your pledge stays anonymous.²⁵
• After a school goes phone-free, the ban has to hold. In England, schools that enforced it weakly saw no significant change in test scores. In Norway, bans changed no averages but cut bullying.¹⁷
• After you write your state legislator, who you are matters most. In interviews with 77 state legislative offices in ten states, constituents came first, and many offices, though not most, threw away messages from outside the district. Roughly half the states kept tallies; most offices did so only for major bills. Most offices were annoyed by form letters but often counted them the same as personal ones if a constituent sent them. The study is undated (its file is from 2004) and was done for the ASPCA. No study we read says how fast state offices reply.²⁶
• After you call, text or chat 988, a live counselor says hello, asks whether you’re safe, then listens and shares resources. 988 tries to route calls to the nearest center by area code. You don’t have to say who you are or where you are. Confidential isn’t anonymous: the counselor takes notes, a center may record the call, and centers keep records at least three years. Your details can be used to arrange further care, to look into complaints or deaths, for court orders and, in some states, to bill insurance. Otherwise nothing is shared without your consent unless the law requires it or someone is at imminent risk of harm. 988 can’t trace you the way 911 can. In the rare case emergency services must be called and you can’t or won’t say where you are, the counselor gives 911 your phone number, or your IP address in a chat.²⁷

Questions people ask

With an AI beside you

These are the habits above, written as prompts.

Send this page to one person:

Sources

1. Bastani et al., Generative AI without guardrails can harm learning, PNAS 122(26), 2025 (open access via Europe PMC) research; Contractor and Reyes, preprint, October 2026 (211 people; the gain from tutoring-style use held a week later) preprint.
2. Anthropic, How AI assistance impacts the formation of coding skills, 29 January 2026 (52 engineers; the company’s own trial, not peer-reviewed) industry research; Lee et al. (Microsoft Research and Carnegie Mellon), The impact of generative AI on critical thinking, CHI 2025 (self-reported, 319 workers) research.
3. Kosmyna et al., Your Brain on ChatGPT, preprint, not peer-reviewed; the comment disputing its methods, arXiv 2601.00856, is also unreviewed. preprint
4. Kestin et al., AI tutoring outperforms in-class active learning, Scientific Reports, 2025 (194 students; a designed tutor, 0.63 to 1.3 standard deviations) research; World Bank, From chalkboards to chatbots, working paper, 2025 (a randomized trial in nine schools in Benin City, after school, with teachers guiding; +0.31 standard deviations overall, 0.23 in English) preprint.
5. Pennycook and Rand, Accuracy prompts are a replicable and generalizable approach for reducing the spread of misinformation, Nature Communications, 2022 (20 experiments, 26,863 people) research; Brodsky et al., Improving college students’ fact-checking strategies through lateral reading instruction in a general education civics course, Cognitive Research, 2021 (230 students, a quasi-experiment) research.
6. Allcott et al., The welfare effects of social media, American Economic Review, 2020 (four weeks off Facebook: wellbeing +0.09, news knowledge −0.19 standard deviations, in the authors’ 2019 version) research; Allcott et al., The effects of Facebook and Instagram on the 2020 election, PNAS, 2024 (19,857 Facebook and 15,585 Instagram users; polarisation effects near zero) research; Castelo et al., Blocking mobile internet on smartphones improves sustained attention, mental health, and subjective well-being, PNAS Nexus, February 2025 research; Guess et al., How do social media feed algorithms affect attitudes and behavior in an election campaign?, Science, 2023 research; Milli et al., Engagement, user satisfaction, and the amplification of divisive content on social media, PNAS Nexus, March 2025 research; Piccardi et al., reranking an X feed, Science, November 2025 research.
7. Anthropic, Is my data used for model training? and How long do you store my data?; Google, Gemini Apps privacy hub, updated 26 October 2026; OpenAI, Data controls (blocked automated reads; read through a search index). primary
8. Meta, Improving your recommendations on our apps with AI at Meta, 1 October 2025 (in effect 16 December 2025). primary
9. C2PA, Content Credentials FAQ; Content Authenticity Initiative, How it works (credentials “aren’t intended to prescriptively indicate whether a piece of content is ‘real’”). primary
10. Google DeepMind, SynthID, read 5 Nov 2026. primary
11. Diel et al., Human performance in detecting deepfakes: a systematic review and meta-analysis of 56 papers, Computers in Human Behavior Reports, 2024 (deepfakes spotted 55.54% of the time, not significantly above chance; read in the authors’ preprint and the journal’s abstract) research; International AI Safety Report 2026, 3 February 2026 (77%, from one study it cites) primary; Deepfake-Eval-2024, preprint, May 2026 (open-source detectors’ AUC fell by 45% to 50% on real-world fakes) preprint.
12. Altay and Gilardi, People are skeptical of headlines labeled as AI-generated, PNAS Nexus, October 2024. research
13. Liang et al., GPT detectors are biased against non-native English writers, Patterns, 2023 (61% of TOEFL essays flagged). research
14. FBI, public service announcement, 3 December 2024; FTC, family emergency scams, March 2023. primary
15. Common Sense Media, AI companions decoded, 30 April 2025, and its institute’s methodology and funding page, read 6 Nov 2026. primary
16. American Academy of Pediatrics, How AI chatbots affect kids: benefits, risks and what parents need to know, updated 27 August 2025. primary
17. Beland and Murphy, Ill communication: technology, distraction and student performance, LSE Centre for Economic Performance, 2015 (four English cities; +6.4% of a standard deviation, +14.2% for the lowest quartile) research; Abrahamsson, Journal of Human Resources, 2026, doi:10.3368/jhr.0224-13403r2 (Norway; abstract read) research; Goodyear et al., school phone policies and adolescent wellbeing, 2025 (30 English schools, cross-sectional; no difference in wellbeing or attainment) research; Vereecke et al., one school’s ban, Frontiers in Developmental Psychology, 2026 (no clear improvement in social wellbeing) research; UNESCO, How many countries have phone bans in school?, 2026 (58% of countries by March 2026) primary.
18. Wait Until 8th, the pledge, read 6 Nov 2026 (“more than 155,000 parents”; its “Why wait” page cites studies of phone harms, none of the pledge) charity’s own figures; its 2025 filing at ProPublica primary.
19. Goyal et al., meditation programs for stress and well-being, JAMA Internal Medicine 2014; Castelo et al., blocking mobile internet, PNAS Nexus, February 2025; Holt-Lunstad et al., social relationships and mortality, PLOS Medicine 2010. research
20. Bergan and Cole, Call your legislator, Political Behavior, 2015 (a field experiment on Michigan state legislators; constituent phone calls) research; Congressional Management Foundation, Communicating with Congress: perceptions of citizen advocacy on Capitol Hill, 2011 (an online survey of 260 congressional staff, October to December 2010) primary; Broockman and Skovron, Bias in perceptions of public opinion among political elites, American Political Science Review, 2018 research.
21. White House, Executive Order 14365, “Ensuring a National Policy Framework for Artificial Intelligence”, 11 December 2025 primary; Inside Privacy on the Senate’s 99–1 vote, July 2025 news; New York Senate, S8828 (RAISE Act amendment) and S6953B primary; IAPP, New York’s governor signs youth AI chatbot safety bill, 28 October 2026 (S9051B; in force 1 January 2027) news; Arnold & Porter, Colorado narrows its AI law, July 2026 (SB 26-189, effective 1 January 2027) news; European Commission, AI Act page, updated 19 October 2026 primary; the 2 December 2026 deadline is in Regulation (EU) 2026/1744, which adds Article 111(4) to the AI Act, read on EUR-Lex 6 Nov 2026 primary; Multistate, state AI companion chatbot laws, 26 June 2026 (twelve states by mid-year, though its own list names eleven) industry research.
22. Stop Killer Robots, on the Seventh Review Conference, 22 October 2026; the official dates were on no page we could read. news
23. 988 Suicide & Crisis Lifeline, Get help, Help someone else and Is the 988 Lifeline available in other languages?, read 6 Nov 2026. primary
24. Google, Verify AI-generated images, videos, and audio, Gemini Apps Help, read 6 Nov 2026. primary
25. Wait Until 8th, FAQs, read 6 Nov 2026. charity’s own figures
26. HRC for the ASPCA, ASPCA legislative alerts study: interview analysis, hosted by Faunalytics (77 phone interviews with state legislative offices in ten states; undated, the file was created in 2004; not peer-reviewed), read 6 Nov 2026. industry research
27. 988 Suicide & Crisis Lifeline, What to expect, Confidentiality statement, Does the 988 Lifeline trace callers? and on call recording, read 6 Nov 2026. primary

App menus and AI laws move fast. The autonomous-weapons line goes stale after the November 2026 meeting; the EU marking deadline is 2 December 2026.

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