# C6 research: the "trust funnel" (hear, understand, check at the official account, share) and the habit claim

Date of research: 19 Nov 2026. About 33 searches, then sources opened.
Marks: READ = I opened the paper or page and saw the figure (some through the fetch tool's extract; Pennycook 2021 and Amar 2025 I extracted locally with pdftotext). SECOND-HAND = only a summary, press page or abstract-level search result. Nature.com, PMC and Science.org blocked direct fetches, so several papers are SECOND-HAND.

## (a) Triage, scepticism, the liar's dividend

1. Hoes, Aitken, Zhang, Gackowski, Wojcieszak, Nature Human Behaviour 8, 1545-1553, 2024. Three survey experiments (US, Poland, Hong Kong), n = 6,127. Fact-checking, media-literacy tips and news about misinformation all cut belief in false claims AND lowered the credibility rating of accurate information. https://doi.org/10.1038/s41562-024-01884-x . SECOND-HAND (university news releases; paper not opened). Supports the worry that "doubt everything" training spills onto true content.
2. Guay, Berinsky, Pennycook, Rand, "How to think about whether misinformation interventions work", Nature Human Behaviour 7, 1231-1233, 2023. Argument: judge an intervention by discernment (true vs false), not only by lower belief in false items. https://www.nature.com/articles/s41562-023-01667-w . SECOND-HAND (paywalled; abstract fragments only). No numbers seen.
3. Ternovski, Kalla, Aronow, "The Negative Consequences of Informing Voters about Deepfakes: Evidence from Two Survey Experiments", Journal of Online Trust and Safety 1(2), 2022. Telling people about deepfakes did not improve detection of fakes; it made them judge videos as fake even when real ("regardless of whether the video is fake or real"). https://tsjournal.org/index.php/jots/article/view/28 . READ (abstract page only; sample sizes and effect sizes NOT seen, the PDF link was 404).
4. A UK deepfake-warning experiment (n = 1,093; arms 361/354/365): in the warning arm 45% flagged more than one video, so wrongly called at least one real video a deepfake. Cited inside PMC10679876, READ via Europe PMC. Authors and year not captured; say only "a UK experiment".
5. Schiff, Schiff, Bueno, "The Liar's Dividend", APSR 2024 (print 119, 2025). Five pre-registered experiments, 15,287 US adults, four real scandals. Claiming "it's fake news" raised support for the politician for text scandals: pooled 0.17 SD (informational uncertainty) and 0.21 SD (oppositional rallying). For VIDEO scandals it mostly did not work (Study 1 none; Study 4 one tactic 0.14 SD). https://www.cambridge.org/core/product/687FEE54DBD7ED0C96D72B26606AA073/core-reader . READ (first half). Video was harder to dismiss than text. A summary says a fact-check removes the dividend (SECOND-HAND; not in the text I read).
6. Köbis, Doležalová, Soraperra, iScience 2021: 210 participants, people cannot reliably spot deepfakes; awareness and money incentives did not help; they were overconfident. SECOND-HAND. Groh et al. report 66-72% crowd accuracy on video (SECOND-HAND); lab figures range 57.6% to 88.9% across studies. "Just look closely" is not a safe step to teach.

For step (a): no study tests "verify only what you will act on". Closest facts: people share without attending to accuracy (see b); blanket scepticism has a measured cost (1, 3, 4); the liar's dividend is real for text and weak for video (5). The triage idea is reasoned, not proven.

## (b) Sharing friction and accuracy prompts

7. Pennycook, Epstein, Mosleh, Arechar, Eckles, Rand, Nature 592, 590-595, 2021. READ (pdftotext of the paper). Study 1, n = 1,002 MTurk, 36 headlines: people rated true headlines accurate 55.9 percentage points more often than false ones, but veracity moved sharing intention only 5.9 points (partisan fit 19.3 points). Study 7, Twitter field experiment, n = 5,379 users who had shared Breitbart/Infowars links: one DM asking them to rate one non-political headline raised the average quality of news shared in the next 24 h by 4.8% (2.1% counting all user-days), and sharing discernment was 2.8 times higher. The effect was in low-effort re-shares (72.4% of data), not in posts with a comment. Authors' limits: US political news, 24 h only, cross-cultural untested.
8. Pennycook and Rand, Nature Communications 2022: pooled 20 experiments, N = 26,863; accuracy prompts cut sharing of false headlines by about 10% relative to control. https://ideas.repec.org/a/nat/natcom/v13y2022i1d10.1038_s41467-022-30073-5.html . SECOND-HAND. A later adversarial collaboration (21 surveys, over 27,000 people) found prompts reduce sharing for left and right (SECOND-HAND, year unconfirmed).
9. Ceylan, Anderson, Wood, PNAS, Jan 2023, "Sharing of misinformation is habitual, not just lazy or biased": four studies, 2,476 participants; 15% of the most habitual users shared 37% of the false news; accuracy prompts appear short-lived because the old sharing habit returns. SECOND-HAND (summaries; PMC blocked). Supports "the share-habit is what to change"; cuts against "one reminder is enough". A 2025 Transparent Replications report re-tested its Study 2; verdict not read.
10. WhatsApp forward limits. Melo et al., "Can WhatsApp counter misinformation by limiting message forwarding?", 2019, arXiv 1909.08740. READ (page extract). Data: public political groups; India 5,839 groups / 362,739 users (Mar-Jun 2019). Simulated limits of 5, 20, 256: limits "slightly diminish the velocity of spreading, but does not stop it"; slowdown one order of magnitude (Section 6) vs "up to two" (conclusion), so the paper contradicts itself; for highly viral content "ineffective in blocking" spread through public groups. A model, not a before/after measurement.
    WhatsApp's own claim: limits cut "highly forwarded messages" by over 70% (spokesperson quoted by Dhaka Tribune; SECOND-HAND; no method). In 2022 forwarded-labelled messages were limited to one chat at a time (same source). I found no independent before/after study. A Singapore interview study (SECOND-HAND) found users read the "forwarded" tag as originality and sincerity, not as a warning.
11. Meta/Instagram labels. Indicator's audit: 516 AI posts across five platforms, only 169 (just over 30%) correctly labelled; a March 2026 follow-up said all platforms except Instagram improved. https://indicator.media/p/tech-platforms-fail-to-label-ai-content-c2pa-metadata . SECOND-HAND. The Quint (2026): 19 AI videos tied to the 2026 Assam/West Bengal/Kerala/Tamil Nadu elections, "none ... labelled as AI-generated". READ (https://www.thequint.com/news/webqoof/ai-videos-state-elections-bjp-congress-meta-labelling-analysis). So "wait for the platform label" fails.

For the share step: friction and accuracy prompts are the best-evidenced piece, but evidence is US, English, text headlines, small effects (5-10%), short horizon. No independent India measurement of WhatsApp limits.

## (c) Lateral reading, SIFT, Civic Online Reasoning, toolbox

12. Wineburg and McGrew, Teachers College Record 2019 (report 2017): 10 historians, 10 professional fact-checkers, 25 Stanford undergraduates thinking aloud. Fact-checkers left the page, opened new tabs, and "arrived at more warranted conclusions in a fraction of the time". READ (Stanford summary page; no seconds figure found). Small; experts, not laypeople; US websites.
13. Wineburg, Breakstone, McGrew, Smith, Ortega, Journal of Educational Psychology 2022, district-wide field study. About 500 students (271 treatment, 228 control per a search summary), urban mid-western district; six 50-minute lessons in a semester; treated students "did twice as well" as before, yet earned only half the possible points. https://ed.stanford.edu/news/it-doesn-t-take-long-learn-how-spot-misinformation-online-stanford-study-finds . READ (press page; no effect size seen). No follow-up weeks later.
14. Kozyreva et al., "Toolbox of individual-level interventions against online misinformation", Nature Human Behaviour 8, 1044-1052, 2024: 81 papers, nine intervention types (accuracy prompts, debunking, friction, inoculation, lateral reading/verification, media-literacy tips, social norms, source-credibility labels, warning/fact-check labels). SECOND-HAND (abstract-level); per-type effect sizes not read. Framing: individual tools are one layer, not a system answer.
15. Guess et al., PNAS 2020, Facebook "tips to spot false news": discernment up 26.5% (US nationally representative) and 17.5% (highly educated online India sample); NO effect in a rural northern-India sample (low social-media use); US gain shrank but remained weeks later; authors say lessons need regular reinforcement. SECOND-HAND (Princeton/Exeter press). Amar et al. cite it as 0.11 SD in India (READ, footnote).

Limit: all lateral-reading tests are on websites and text, not on a video clip. I found no test of "search the claim in another tab" on a deepfake video, and none with a Tamil layperson on a phone inside WhatsApp or Instagram.

## (d) Checking at the official account

16. No controlled study found of "go to the person's own account to check". Not found does not mean absent, but the literature I reached is about fact-checkers, not primary-source checking.
17. Official accounts get taken over: SEC (US) X account Jan 2024 (fake bitcoin approval); UK Leader of the House account 2025; Peru economy ministry, Kenya DCI, Ireland's Garda; Modi's Twitter account Dec 2021 (bitcoin message). https://www.malwarebytes.com/blog/news/2024/01/sec-x-account-hacked-to-hawk-crypto-scams . SECOND-HAND (search summaries). None was a deepfake video; they show the account can be wrong.
18. Official accounts post AI content themselves. The Quint (READ): of 19 AI election videos in 2026, several came from official handles of BJP (BJP Assam: a since-deleted video of Himanta Biswa Sarma shooting at photos), Congress (Assam unit's Instagram) and DMK (a governance video with no AI disclosure; Quint says it did not appear to spread misinformation). Tamil Nadu (search summaries, SECOND-HAND, verify before using on screen): TVK released an AI video of C. N. Annadurai in Sept 2025, which the DMK called a deepfake; [withheld: complaints and police cases about makers of AI clips]; the state fact-check account called a clip of Udhayanidhi Stalin a deepfake.
19. Tamil Nadu has its own checkers: TN Fact Check Unit (state government; Udhayanidhi claimed 47,000 people reached by workshops and fewer fake messages; a government claim, SECOND-HAND, and a government unit checking for its own side). MCA/Meta Deepfakes Analysis Unit WhatsApp tipline, live 25 Mar 2024, English, Hindi, Tamil, Telugu: forward audio or video, get an assessment (SECOND-HAND, news reports; request counts not found).

## (e) Habit formation and training that lasts

20. Lally, van Jaarsveld, Potts, Wardle, European Journal of Social Psychology 40, 998-1009, 2010. 96 volunteers (mean age 27) chose one daily behaviour (eating, drinking, exercise) for 84 days; 82 gave enough data; 39 had a good curve fit; for those 39 the median time to 95% of the automaticity plateau was 66 days, range 18 to 254. Source: a close reading at https://www.thebehavioralscientist.com/articles/how-long-to-form-a-habit , so SECOND-HAND for the paper. Cautions: 66 days is for 39 of 96; 254 is a model projection beyond the 84-day study; behaviours were eating/drinking/exercise, not a mental check; one missed day did not matter much. Safe wording: "weeks, for many about two months".
21. Gollwitzer and Sheeran 2006: 94 independent tests, if-then plans d = 0.65 on goal attainment. A later 642-test review (Sheeran, Listrom, Gollwitzer, 2024) found effects .27 to .66, but publication-bias corrections shrank it (d = 0.35 or 0.15 per a blog's account). SECOND-HAND (search results, abstract, blogs). Do not quote 0.65 as settled.
22. Badrinathan, APSR 115(4), 2021. Bihar, 2019 elections, N = 1,224. One-hour in-person training (inoculation, corrections, why verify): no significant average effect; BJP supporters who were treated became significantly worse at spotting pro-party misinformation. READ (Cambridge abstract page).
23. Amar, Badrinathan, Chauchard, Sichart, APSR 2025. READ (PDF text). 583 Bihar villages, over 13,500 students aged 13-18; four 90-minute sessions over 14 weeks with homework; placebo (English classes). Accuracy discernment +0.32 SD, source credibility +0.21 SD, sharing discernment +0.21 SD, health preferences +0.21 SD. A random 2,059 resurveyed about 4 months later: discernment +0.26 SD (health) and +0.31 SD on political items never taught. Parents of treated students also better. Session 3 taught judging sources, verifying, "how to decide whether to share", reverse image search, the "forwarded many times" tag, Indian fact-check sites; session 4 taught talking to an elderly relative. BUT: (i) the gain came from rejecting false items (false statements +0.31 to +0.33 SD), while belief in true statements did not rise (health -0.06, politics -0.01, not significant); (ii) at the first survey belief in true items had briefly dropped, "no longer detectable" at 4 months; (iii) only 19% of the sample used the internet (an author's Scroll piece, SECOND-HAND); (iv) the authors argue that in India "the risk is not that people are too sceptical, it is that they are too trusting". Strongest India support for "slowly becomes behaviour", and it took 14 weeks of teaching, not a reminder.
24. A 2024 meta-analysis of 49 experiments (SECOND-HAND, authors not captured): multiple sessions beat a single session. A German randomised study of 2,666 adults (SECOND-HAND): small gains, indistinguishable from control at follow-up. A Taiwan school study (n = 326, four 50-minute sessions, one-week follow-up; READ) is on weight-loss ads, not deepfakes.
25. India school programmes. Kerala "Satyameva Jayate" (KITE): 19.72 lakh students of grades 5-10 in 2022, 5,920 trainers, a 2.5-hour course, fact-checking modules in Class 5 and 7 ICT textbooks (SECOND-HAND, news and government announcements). I found NO independent evaluation. Tamil Nadu: an awareness week on misinformation in government schools each year (Deccan Herald 2023, SECOND-HAND; current status unconfirmed); TN Fact Check Unit workshops; the India Institute "Tamil Nadu Critical Thinking Curriculum" randomised trial is complete, report "forthcoming" (https://tnct.indiai.org/the-issue/, SECOND-HAND). No Tamil Nadu deepfake-literacy programme found. Google "Be Internet Awesome": not examined; no claim.

## (f) Prebunking and inoculation

26. Roozenbeek, van der Linden, Goldberg, Rathje, Lewandowsky, Science Advances 2022. READ (Europe PMC full text). Six randomised studies (n = 6,464) plus a YouTube field study: 22,632 responses (11,432 saw the video, 11,200 control); about 967,000 watched 30 s or more; recognition of the manipulation technique up "about 5%", Cohen's h = 0.09, which "narrowly failed" the pre-registered smallest effect of interest (h = 0.10); met only for the false-dichotomy video (0.13). Median 18.4 h between video and question. Authors: could not study how long it lasts; US-tailored; measured susceptibility, not behaviour.
27. Jigsaw in India: five videos in Bengali, Hindi and Marathi announced Dec 2022, with a short questionnaire (SECOND-HAND, Tribune/ThePrint). No published results found; nothing in Tamil; no Maharashtra evaluation.
28. Maertens et al., Nature Communications 2025, five experiments, N = 11,759: text and video inoculation stayed effective about one month; game-based faded faster; memory boosters restored the effect. READ (abstract page). Earlier JEP:Applied 2021: effect held three months in one experiment but decayed over two months without regular testing (SECOND-HAND). Suggestions that 3-5 boosters could last a year are untested.

## (g) Does it get easier with practice?

29. I found no study that measures verification speed or accuracy improving as a layperson repeats a checking routine, and none that times a lateral-reading check in seconds. Nearest: Amar (14 weeks of practice, effect held at 4 months); Guess 2020 (decay without reinforcement); Lally (automaticity rises with repetition for simple daily acts). "It gets easier" is an inference from habit science, not a measured fact for checking.

## What the evidence supports

Step 1, HEAR. Nothing tests "hear first, do nothing". What supports it indirectly: sharing tracks attention, not belief (Pennycook 2021), and sharing is habitual (Ceylan 2023). The step needs a plain rule "hearing is not believing".

Step 2, UNDERSTAND. The vaguest step. "Understanding" a video is not what any study tests; people cannot reliably see a fake with their own eyes (Köbis 2021, SECOND-HAND). What has evidence is lateral reading (who is behind it, what do others say) and Amar's weeks of source and sharing training. As a gate ("does it matter enough to act on") it is the attention step in (a), unproven in this form.

Step 3, CHECK at the official account. Reasonable and cheap, but no study of it. For: a self-opened account is harder for a stranger to plant a fake in than a forward (inference). Against: accounts are hacked (SEC 2024, Modi 2021), official handles of BJP, Congress and DMK posted AI videos in 2026, TVK released an AI Annadurai in 2025, and platform labels catch about 30% of AI posts. Many claims have no official channel (a leaked call).

Step 4, SHARE only if you trust it. Best supported: accuracy prompts (about 5-10% fewer poor shares), forward limits slow but do not stop spread (Melo 2019), taught sharing criteria raised sharing discernment 0.21 SD in Bihar (Amar 2025). All US or Bihar-school, short horizon, except Amar.

The habit claim. Plausible and partly supported: Lally (median 66 days, range 18-254, 39 of 96, simple daily acts); Amar (14 weeks, held at 4 months, India). Failures fit it too: one hour in Bihar did nothing (2021); accuracy prompts fade (Ceylan). Not supported: that checking becomes "easy" or "masters" deepfakes.

## Wording risks (what NOT to say)

- "This makes you immune", "you will master it", "whatever the amount of deepfakes": no study shows it. The best India study shows +0.26 to +0.32 SD, from rejecting false items. Say "you will be fooled less" or "you will pass on fewer fakes".
- "The official account is always the truth": accounts are hacked and parties post AI. Say "the official account first, and a second place besides".
- "66 days" as a promise: say "for many about two months, for some much less or more" (18 to 254; median 66 for the 39 of 96 whose curves fit; simple daily acts).
- "Scientists proved lateral reading or accuracy prompts work against deepfakes": tested on websites and headlines; effects 5-10%, or "twice as good but half the points".
- "Doubt everything": warnings made people call real videos fake (Ternovski 2022; UK 45%); the liar's dividend (text more than video).
- Kerala or Tamil Nadu school programmes "work": counted by students reached, not outcomes.
- WhatsApp's forward limit "cut misinformation 70%": the company's claim, no method.
- "The platform label will tell you": about 30% labelled in one audit; none of 19 Indian election AI videos labelled.
- "A face or voice proves it": people are near chance by eye in several lab studies.
- Do not present the funnel as researched. Say "a method built from parts tested separately".

## Better ordering and a missing step

1. Put a PAUSE first, before "hear": "about to share or act? if not, let it pass". The robust finding is impulse sharing; this is the accuracy-prompt step, costs nothing, and removes the cost of checking everything.
2. Merge "understand" into "check" in plain words: "who is saying this, and where else is it?" A layperson cannot "understand" a deepfake; they can ask who says it and who else does (lateral reading).
3. Check in TWO places: the person's own account AND one independent place (a news outlet, the TN Fact Check Unit, or the DAU WhatsApp tipline in Tamil for audio and video). Reason: hacks and party AI. Confirm the tipline number on the DAU site before naming it (not verified here).
4. Add "no answer, no share": if you cannot check, do not forward. The one part needing no skill, and what forward limits do.
5. Add the habit mechanism as an if-then plan: "if a video makes me angry or scared and I want to forward it, then I first look at where it came from". Implementation intentions have support (0.65 in 2006, much smaller after bias correction) and fit Lally's cue-linked habits.
6. Add a family-group step: when you checked and it was false, say so in the family WhatsApp group. Amar found parents of trained students improved too (READ); it is the only evidence that the habit spreads past one person, though it was tested as classroom spill-over, not as correcting a group.
7. Honest close: "you will not catch every fake, but you will stop being the one who passes it on".
