TL;DR. The Reuters Institute’s Digital News Report 2026 puts trust in news at 37%, the lowest since 2015, and 62% of respondents worry about what is real online. Repetition alone makes claims feel truer, an effect documented in 181 studies and strongest right after exposure. The defences that work are small and they fade: an accuracy prompt cut intentions to share false headlines by 10% across 20 experiments, but an independent replication found about half the original effect; a 1.5-minute inoculation video lifted technique recognition by d = 0.28 to 0.68 in the lab and about 5% on YouTube; lateral reading shows the largest effect among source-evaluation methods (g = 0.55). The CEO move: give every decision-relevant belief a confidence level and a review date, and never count repetition as evidence. The student move: run a 10-minute daily routine and a 20-minute weekly audit, because every intervention here decays without practice.
Related CEOtudent guides cover filtering information, checking AI output claim by claim and probability scoring. This article adds the evidence on how repetition distorts belief, a side-by-side comparison of the interventions that measurably help, and a routine built on both.
What is epistemic hygiene, and why does it need a routine?
Epistemic hygiene is the set of small, repeated habits that keep what you believe in line with the evidence: noticing where a belief came from, stating how confident you are, checking claims before acting on them, and revising when wrong. In this article the term has a narrow meaning: a routine whose components are each backed by an experimental literature.
The hygiene metaphor fits because the evidence-backed interventions work at the moment of use and lose strength over days and weeks. A one-off course does not produce lasting protection; a habit can.
The Reuters Institute for the Study of Journalism, a University of Oxford research centre, surveys online news users each year. Its 2026 report draws on almost 100,000 respondents in 48 markets, surveyed by YouGov from the middle of January to the end of February 2026.
| Indicator (average across markets) | Digital News Report 2026 | Comparison point |
|---|---|---|
| Trust news most of the time | 37% | 40% for the previous three years (-3 pp) |
| Avoid the news sometimes or often | 42% | 29% in 2017, when first measured (+13 pp) |
| Concerned about what is real and fake online | 62% | Up 4 pp on the previous year |
| Use social media and video networks for online news | 54% | 51% use news organisations’ own sites and apps |
| Use AI chatbots for news | 10% | 7% the previous year; 16% of under-35s |
| Trust answers from AI chatbots | 20% | 37% trust news overall (17 pp gap) |
Source: Reuters Institute for the Study of Journalism, Digital News Report 2026 (Executive Summary and Key Findings). Online survey; the report notes it slightly over-represents online use. Differences in pp computed by CEOtudent (calc.py).
For an individual, two points follow. More news now reaches people through third-party platforms than through news organisations’ own sites, so more of the work of judging reliability falls on the reader. And the common responses, distrusting everything or avoiding the news, are not the same as accuracy: avoidance reduces exposure but does not tell you which of your existing beliefs are wrong.
Why does repetition quietly rewrite what you believe?
The best-documented threat to belief accuracy is the claim you meet several times without noticing. The illusory truth effect is the finding that repeated statements are rated as more true than new ones, although repetition adds no evidence. The first paper appeared in 1977 (Hasher, Goldstein and Toppino). A 2022 systematic map by Henderson, Westwood and Simons identified 181 studies in 93 articles, chapters or theses across five decades and 12 countries. The only meta-analysis at the time, Dechêne and colleagues (2010), combined 51 studies and estimated d = 0.39 within items (the same statement before and after repetition) and d = 0.50 between items (repeated versus new statements). These figures are cited as reported in the map, whose authors note that other bias-correction approaches might estimate a smaller effect.
Knowledge is a weak shield. Henderson and colleagues summarise work showing the effect persists when people are warned about it, when they know the factual answer (Fazio, Brashier, Payne and Marsh, 2015), and when they are told which statements are true.
What changes the effect is time. A registered report by Henderson, Simons and Barr (2021) tested 567 analysed participants at four intervals after exposure.
| Interval between exposure and judgement | Illusory truth effect (repeated minus new, log odds) | Share of the immediate effect remaining |
|---|---|---|
| Immediately | 1.04 | 100% |
| One day | 0.56 | 54% |
| One week | 0.37 | 36% |
| One month | 0.20 | 19% |
Source: Henderson, Simons and Barr (2021), Journal of Cognition, Table 6 (all four contrasts significant after Holm-Bonferroni correction). The right-hand column is a CEOtudent calculation (calc.py).
The moment of greatest vulnerability is right after you have seen a claim several times in one session, which is how a feed, a group chat or a long chatbot exchange works. A belief that feels settled after an evening of scrolling has gained repetition, not evidence. The routine below therefore treats “how many times have I seen this?” as a warning sign.
Which interventions actually improve accuracy, and how big are the effects?
Four families of interventions have experimental support: accuracy prompts, inoculation (prebunking), lateral reading and calibration training. They measure different outcomes, so their effect sizes cannot be ranked directly. What can be compared is time cost, what changed, how the result held up when retested, and how long it lasted.
| Intervention | Key study (sample) | Time cost | What it changed | Replication or field test | Durability evidence |
|---|---|---|---|---|---|
| Accuracy prompt (for example, rate one headline’s accuracy first) | Pennycook and Rand 2022, internal meta-analysis of 20 experiments (N = 26,863) | One rating question | Intentions to share false headlines fell by 10% relative to control | Independent preregistered replication (pooled n = 1,583): effect about 50% of the original | Did not significantly decay over successive trials within a session |
| Inoculation video: learn one manipulation technique | Roozenbeek et al. 2022, six lab studies (n = 6,464) | About 1.5 minutes | Technique recognition d = 0.28 to 0.68 (mean 0.50); sharing discernment mean d = 0.17 | YouTube field study (n = 22,632): recognition up about 5% on average, Cohen’s h = 0.09 | Text and video versions remained effective for one month (Maertens et al. 2025, N = 11,759) |
| Memory booster after inoculation | Maertens et al. 2025, Study 5 | One short booster video | At day 29: d = 0.440 with memory booster vs d = 0.230 without (about 1.9 times) | Five preregistered longitudinal experiments | Memory booster retained 100% of the original effect; no booster retained 78% |
| Lateral reading: leave the site and search what others say about it | Fendt, Muth and Edelsbrunner 2025, meta-analysis of 64 studies | Six 50-minute lessons (300 minutes) in the Stanford field study | Lateral reading g = 0.55 (43 effect sizes) vs g = 0.42 for all source-evaluation approaches | Stanford district-wide, cluster-randomised field study, about 500 students | Meta-analysis finds repeated practice may enhance effects; wide prediction interval |
| Calibration and debiasing training | Chang, Chen, Mellers and Tetlock 2016, four forecasting tournaments | Less than one hour | Brier scores (accuracy of probability forecasts) improved 6 to 11% over control | Effect held across four years of tournaments | Authors call the effects likely lower-bound estimates |
CEOtudent analysis of Pennycook and Rand (2022), Roozenbeek, Freeman and van der Linden (2021), Roozenbeek et al. (2022), Maertens et al. (2025), Fendt et al. (2025), Wineburg et al. (2022, via Stanford GSE), and Chang et al. (2016). Means, ratios and the 300-minute total computed in calc.py. Outcome measures differ across rows, so effect sizes are not directly comparable.
Five patterns stand out.
1. The cheapest interventions are the most fragile. In the independent replication by Roozenbeek, Freeman and van der Linden, truth discernment improved about 1.4 times in the treatment group relative to control, against about 2.8 times in the original study: an effect around half the size, driven mostly by better baseline discernment in the replication’s control group. A tool that halves under replication is still useful, but it is not a cure.
2. Recognising a trick is easier than changing behaviour. In the inoculation lab studies, the mean effect on recognising a technique (d = 0.50) is about three times the mean effect on sharing decisions (d = 0.17, or 33% of it).
3. The real world shrinks lab effects. On YouTube, where viewers could skip the ad, the same videos raised recognition by about 5% (Cohen’s h = 0.09). Over 15 days they were shown to about 5.4 million people at an average $0.05 per 30-second view, so a small effect still matters for platforms. For an individual, a lab effect size describes a quiet room, not a feed.
4. Effects fade unless refreshed. In the 2025 booster experiments, a memory-focused booster given nine days after the first video kept the full original effect at day 29; the no-booster group kept 78%. This is the strongest argument for a routine over a one-off lesson.
5. The most effective habit is the most expensive. Lateral reading showed the largest effect of four source-evaluation approaches (g = 0.55, which is 0.13 above the overall 0.42). In the Stanford field study, students who took six 50-minute lessons did twice as well on a test of spotting dubious websites as before, yet still earned only half the possible points.
What does the replication record say about these headline numbers?
Each of these literatures has a caveat that a careful reader should carry.
- Accuracy prompts. The 10% figure comes from an internal meta-analysis by the group that developed the intervention; it includes all relevant studies regardless of outcome, which limits selective reporting, but it is one group’s programme. The original Nature paper (Pennycook et al., 2021) combined four survey experiments with a Twitter field experiment in which users were asked to rate the accuracy of a single non-political headline. The direction replicates; the size is uncertain.
- Inoculation. The lab effects on recognition are moderate and the effects on sharing are small: the 2022 paper’s preregistered hypotheses were supported for 16 of 20 outcome measures, and two of the four exceptions were sharing measures. The field effect is real but small. Both inoculation papers include co-authors from Jigsaw, a Google unit, which funded the 2022 study.
- Lateral reading. For lateral reading the 2025 meta-analysis gives a 95% prediction interval of -0.34 to 1.44: a new study could find anything from a small negative to a large positive effect. The Stanford study was funded by Google.org, as the university’s release discloses.
- Illusory truth. The 2010 meta-analysis is dated, and its coded data could not be provided to later researchers, according to Henderson and colleagues. The effect replicates; its exact size is less certain than d = 0.39 and 0.50 suggest.
- Calibration training. The 6 to 11% Brier improvement comes from geopolitical forecasting tournaments; transfer to everyday beliefs was not tested in that study. The decision journal guide covers those training results in detail.
None of this argues for doing nothing. It argues for stacking several cheap habits and measuring yourself instead of assuming a habit works.
The CEOtudent daily epistemic hygiene routine
The routine below is a CEOtudent editorial framework. Its structure follows the studies above; its time budgets are design choices, not findings. It has a daily loop of about 10 minutes and a weekly audit of about 20 minutes: 70 minutes a week over five working days.
| Step | When | What you do | Time budget | Evidence it rests on | Stop rule |
|---|---|---|---|---|---|
| 1. Intake: accuracy first | Before opening a feed, inbox or chatbot session | Ask one question: “Is the next thing I read likely to be accurate?” Then, before saving or sharing anything, ask whether it is true, not whether it is interesting | 2 minutes | Accuracy prompts (Pennycook and Rand 2022; replication at about half size) | If you notice you have seen a claim three or more times in one session, mark it “repeated, unverified” and stop treating it as known |
| 2. Label: confidence and origin | When a new claim could affect a decision, a purchase, or what you tell others | Write the claim in one sentence, a confidence level (for example 60%), where you first saw it, and how many times you have seen it | 3 minutes | Illusory truth effect is strongest immediately after repetition (Henderson et al. 2021); calibration training improves probability accuracy (Chang et al. 2016) | Claims with no decision attached do not need a label |
| 3. Check: read laterally | Only for labelled claims that will change an action this week | Leave the source. Search for what independent sources say about the author or outlet, then look for the primary document | 5 minutes | Lateral reading (Fendt et al. 2025 meta-analysis; Stanford field study) | If two independent, credible sources disagree with the claim, lower your confidence and do not act on it yet |
| 4. Audit: score and update | Once a week | Review the belief ledger, mark claims as true, false or unresolved, score your confidence, and refresh one manipulation technique you learned | 20 minutes | Booster sessions preserve inoculation effects (Maertens et al. 2025); calibration requires scoring | Archive resolved entries; carry unresolved ones forward with a new review date |
Three design choices need a word of explanation. Most content is skipped: lateral reading took 300 minutes of instruction to teach and takes effort to run, so checking is reserved for claims that change an action; deciding what not to process at all is covered in the strategic ignorance guide. Repetition is recorded as a warning: the ledger separates “times seen” from “independent sources checked,” and only the second can raise confidence. The audit includes a booster: the longitudinal research found memory, more than motivation, predicted how long protection lasted, so a two-minute weekly refresher on one technique (emotional language, incoherence, false dichotomies, scapegoating, ad hominem attacks) is the cheapest booster available.
Step 3 is not a full verification procedure; for high-stakes AI output, use the 5-step AI fact-check protocol.
The belief ledger: a template you can copy
A decision journal records choices. A belief ledger, a CEOtudent editorial framework, records claims you hold so you can later see how often you were right at each confidence level.
| # | Claim (one sentence) | Date logged | Where first seen | Times seen before logging | Confidence (%) | What would change your mind | Independent sources checked | Review date | Outcome (1 = true, 0 = false, U = unresolved) | Squared error |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | ||||||||||
| 2 | ||||||||||
| 3 |
The last column keeps score. For each resolved claim, subtract the outcome (1 or 0) from your confidence as a decimal and square it. The average is a simple Brier score on a 0 to 1 scale: 0 is perfect, and always saying 50% scores 0.25. (Chang and colleagues use the original formulation, which runs from 0 to 2.)
A worked example, using illustrative assumptions only (four invented ledger entries, not data from any study):
| Entry | Confidence (illustrative assumption) | Outcome (illustrative assumption) | Squared error |
|---|---|---|---|
| A | 90% | True (1) | 0.01 |
| B | 70% | False (0) | 0.49 |
| C | 60% | True (1) | 0.16 |
| D | 80% | True (1) | 0.04 |
| Average | 75% | 75% true | 0.175 |
Illustrative CEOtudent calculation; arithmetic reproduced in calc.py.
Average confidence (75%) matches the share of claims that turned out true (75%), which looks like good calibration. The Brier score of 0.175 beats the 0.25 of always saying 50%, but most of the error comes from one confident miss (entry B). That row is the one to study: where did the claim come from, how often had you seen it, and which check was skipped? The habits of being wrong quickly and cheaply are covered in the intellectual honesty guide.
The weekly epistemic hygiene scorecard
Use this scorecard during the weekly audit. It is a CEOtudent editorial framework, not a validated psychological scale. Score each item 0, 1 or 2; the maximum is 12.
| Habit | 0 points | 1 point | 2 points |
|---|---|---|---|
| Accuracy-first intake | Rarely asked before opening feeds | Asked on some days | Asked before most sessions |
| Repetition flags | Never noticed repeated claims | Noticed but not recorded | Recorded “times seen” and treated it as a warning |
| Labelling | No claims labelled | Some decision-relevant claims labelled | Every decision-relevant claim labelled with a confidence level |
| Lateral checks | Acted on at least one unchecked claim | Checked some claims before acting | Checked every claim that changed an action |
| Updating | No confidence changed this week | Changed confidence, without writing why | Changed confidence and recorded the evidence that moved it |
| Booster | No refresher | Thought about a technique without reviewing it | Reviewed one manipulation technique with an example |
As an editorial rule of thumb (not a validated cut-off), 8 or more for four weeks in a row suggests the routine is becoming a habit. A falling score is more informative than a low one.
How does this fit the CEO and student lens?
The CEO side is ownership of inputs. A chief executive does not let a rumour repeated in three meetings become strategy unchecked. Personal beliefs drive decisions about careers, money and which tools to adopt, so the same standard applies: explicit confidence, and repetition never counted as evidence. When an AI system is the source, this also guards against handing your judgment to the tool, the risk described in the cognitive offloading research summary.
The student side is measured learning. The belief ledger is an answer key you build over time, and the decay findings mean skills for spotting manipulation need spaced review like any other subject.
One finding links the two. Pennycook, Cheyne, Koehler and Fugelsang (2020) measured agreement that beliefs should change with evidence on an 8-item scale (actively open-minded thinking about evidence, AOT-E). In their first study (N = 375), higher scores correlated with lower conspiracy belief (r = -0.36) and lower paranormal belief (r = -0.40). These are correlations, not proof that the attitude changes beliefs, and the scale was more predictive for liberals than conservatives. The routine turns the attitude into behaviour that can be checked.
What this does not mean
- It does not mean distrusting everything. News avoidance is already widespread (42%), and neither avoidance nor blanket distrust tells you which beliefs are wrong. The goal is calibrated trust.
- It does not mean the effects are large. Every intervention here produces small to moderate gains, several shrink outside the lab, and one halved under replication. No single habit is enough.
- It does not mean checking everything. Effort goes only to claims that will change an action.
- It does not replace expertise. On medical, legal or financial questions the check step should lead to qualified, primary sources. This article is general information, not professional advice.
- It does not prove your beliefs are right. A Brier score reflects only the claims you chose to log; logging easy claims flatters you.
Frequently asked questions
What is epistemic hygiene in simple terms?
A set of small daily habits that keep beliefs matched to evidence: noting where a claim came from, stating confidence, checking before acting, and updating when wrong. Here it is four steps (intake, label, check, audit) taking about 70 minutes a week.
Does seeing a claim many times make it more likely to be true?
No. Repetition makes a claim feel truer without adding evidence. The illusory truth effect appears across 181 studies, even when people know the correct answer, and is strongest right after exposure.
Do accuracy nudges really work?
Modestly. Across 20 experiments, prompting people to think about accuracy cut intentions to share false headlines by 10% relative to control; an independent replication found about half the original effect. The prompt is cheap enough to use, but not sufficient alone.
What is prebunking, and how long does it last?
Prebunking (inoculation) teaches a manipulation technique before you meet it. Short videos improved recognition in the lab and by about 5% on YouTube. Text and video versions stayed effective for one month, and a memory booster kept the full effect.
What is lateral reading?
Leaving an unfamiliar website to check what independent sources say about it, as professional fact-checkers do. A 2025 meta-analysis found the largest effects of four source-evaluation approaches (g = 0.55). It takes practice: after six lessons, Stanford study students did twice as well as before but still earned half the possible points.
How is a belief ledger different from a decision journal?
A decision journal records choices, reasoning and outcomes. A belief ledger records claims, your confidence in each, and whether each turned out true, so you can measure calibration.
How can you score yourself without a forecasting platform?
Use the ledger’s squared-error column: subtract the outcome (1 or 0) from your confidence as a decimal, square it, and average across resolved claims. Lower is better; always saying 50% scores 0.25.
Sources
- Reuters Institute for the Study of Journalism (2026). Digital News Report 2026. University of Oxford. Executive Summary and Key Findings; Methodology.
- Henderson, E. L., Westwood, S. J., and Simons, D. J. (2022). A reproducible systematic map of research on the illusory truth effect. Psychonomic Bulletin and Review, 29, 1065-1088.
- Henderson, E. L., Simons, D. J., and Barr, D. J. (2021). The trajectory of truth: A longitudinal study of the illusory truth effect. Journal of Cognition, 4(1), 29.
- Dechêne, A., Stahl, C., Hansen, J., and Wänke, M. (2010). The truth about the truth: A meta-analytic review of the truth effect. Personality and Social Psychology Review, 14(2), 238-257. (Figures cited as reported in source 2; not opened directly.)
- Hasher, L., Goldstein, D., and Toppino, T. (1977). Frequency and the conference of referential validity. Journal of Verbal Learning and Verbal Behavior, 16(1), 107-112. (Cited as the first paper on the effect, per source 2; not opened directly.)
- Fazio, L. K., Brashier, N. M., Payne, B. K., and Marsh, E. J. (2015). Knowledge does not protect against illusory truth. Journal of Experimental Psychology: General, 144(5), 993-1002. (Finding cited as summarised in source 3; not opened directly.)
- Pennycook, G., and Rand, D. G. (2022). Accuracy prompts are a replicable and generalizable approach for reducing the spread of misinformation. Nature Communications, 13, 2333.
- Pennycook, G., Epstein, Z., Mosleh, M., Arechar, A. A., Eckles, D., and Rand, D. G. (2021). Shifting attention to accuracy can reduce misinformation online. Nature, 592, 590-595. (Abstract only.)
- Roozenbeek, J., Freeman, A. L. J., and van der Linden, S. (2021). How accurate are accuracy-nudge interventions? A preregistered direct replication of Pennycook et al. (2020). Psychological Science, 32(7), 1169-1178.
- Roozenbeek, J., van der Linden, S., Goldberg, B., Rathje, S., and Lewandowsky, S. (2022). Psychological inoculation improves resilience against misinformation on social media. Science Advances, 8(34), eabo6254.
- Maertens, R., Roozenbeek, J., Simons, J. S., Lewandowsky, S., Maturo, V., Goldberg, B., Xu, R., and van der Linden, S. (2025). Psychological booster shots targeting memory increase long-term resistance against misinformation. Nature Communications, 16, 2062.
- Fendt, M., Muth, X., and Edelsbrunner, P. A. (2025). Judging a text by its author: A meta-analysis of interventions to foster source credibility assessment. Learning and Individual Differences, 124, 102782.
- Stanford Graduate School of Education (2022, April 19). It doesn’t take long to learn how to spot misinformation online, Stanford study finds. University news release on Wineburg, Breakstone, McGrew, Smith and Ortega, Journal of Educational Psychology (2022).
- Digital Inquiry Group (2022). Lateral Reading on the Open Internet: study summary (Wineburg et al., Journal of Educational Psychology).
- Chang, W., Chen, E., Mellers, B., and Tetlock, P. (2016). Developing expert political judgment: The impact of training and practice on judgmental accuracy in geopolitical forecasting tournaments. Judgment and Decision Making, 11(5), 509-526.
- Pennycook, G., Cheyne, J. A., Koehler, D. J., and Fugelsang, J. A. (2020). On the belief that beliefs should change according to evidence: Implications for conspiratorial, moral, paranormal, political, religious, and science beliefs. Judgment and Decision Making, 15(4), 476-498.
Note on CEOtudent analyses: the intervention comparison table, the “share of the immediate effect remaining” column, all percentage-point differences in the Digital News Report table, and the illustrative Brier example are CEOtudent calculations from the published figures (script retained). The daily routine, the belief ledger and the weekly scorecard are CEOtudent editorial frameworks, not validated instruments; their time budgets are design choices. Figures deliberately not used: the “nearly tripled” truth-discernment claim from the 2020 accuracy-nudge study (superseded by the independent replication, which found about half that effect), the Stanford release’s phrase “twice as likely” to spot questionable websites (the same release also says “twice as well” on the test; the second, more conservative wording is used), and any single headline figure for how much information is produced per day, since no primary statistical source for such a figure was opened.
This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.
This post is also available in:
