# A usable taxonomy of misleading information for Getting It Right

Research memo and editorial proposal, 6 October 2026

## Recommendation

Use a small set of claim-level explanations, not a universal badge system that sorts people or whole topics into misinformation categories. Start with six observable content problems: made-up material, wrong context, missing context, selective evidence, misleading numbers, and unsupported causal claims. Put confirmation bias in a separate learning collection. Keep intent and distribution in separate editorial fields, generally undisplayed unless independently evidenced.

The proposed schema is a design hypothesis, not an experimentally validated taxonomy. Research supports some warning labels, corrections, and teaching interventions. It does not establish that attaching this particular set of tags to long-form educational topic pages improves understanding, durable knowledge, or behavior. A short explanation beside the relevant claim should carry the meaning even if the reader never opens a glossary.

The aim is calibrated understanding: accepting well-supported information, identifying limits, and suspending judgment when warranted. Maximizing distrust, label clicks, or the number of claims classified as misleading would be the wrong objective.

## What existing frameworks do and do not supply

Wardle and Derakhshan's 2017 Council of Europe report distinguishes misinformation, disinformation, and malinformation using falsity and harmful intent, and separately discusses agents, messages, interpreters, and creation/production/distribution. The report reproduces Wardle’s seven-format scheme, including satire/parody, false connection, misleading content, false context, imposter content, manipulated content, and fabricated content. It is a useful conceptual map, not an experimentally validated reader interface. Its harm-based definitions are also not universal: contemporary authors sometimes use misinformation regardless of intent. State the site's convention explicitly rather than presenting one definition as settled. [Report, primary full text](https://rm.coe.int/information-disordertoward-an-interdisciplinary-framework-for-researc/168076277c)

Tandoc, Lim, and Ling (2018) reviewed 34 scholarly articles from 2003–2017 and identified six uses of “fake news,” organized by facticity and deception. This is a taxonomy of how scholars used an unstable term, not six mutually exclusive natural kinds. Satire, advertising, and propaganda are genres or purposes; fabrication and manipulation describe content operations. Do not import that mixture directly into a glossary badge system, or imply that satire or advertising is inherently false. [Institutional publication record and abstract](https://pure.ewha.ac.kr/en/publications/defining-fake-news-a-typology-of-scholarly-definitions/)

Vraga and Bode (2020) emphasize the best available evidence and relevant expertise at a particular time. Their conceptual treatment helps distinguish an inaccurate empirical claim from legitimate uncertainty, evolving findings, or disagreement over values. Editorially, record what evidence was available at publication, use domain-appropriate expertise, and make later revisions visible. Consensus is evidence about the state of a field, not a replacement for explaining that evidence. [Author manuscript, full text](https://emilyk.vraga.org/wp-content/uploads/2020/02/Vraga_Bode_Forum2019_Final.pdf)

These frameworks offer vocabulary and boundary questions. None warrants claiming exhaustive coverage, independent categories, or a validated ordering from mild to severe. The launch schema below is deliberately narrower and operational.

## Four distinctions the site must preserve

1. **Content problem:** What specifically is wrong or misleading in the claim, image, comparison, or reasoning? These are the main candidate explanatory tags.
2. **Distribution mechanism:** How did it travel or gain apparent credibility? Repetition, recirculation, impersonation, coordinated amplification, and source laundering belong here. A distribution mechanism does not establish falsity; accurate material can spread the same way.
3. **Cognitive process:** How might anyone, including editors, select or interpret evidence? Confirmation bias, familiarity, attention, and motivated reasoning are general learning topics. They are not diagnoses of a speaker's mental state or a reader's reasons for disagreement.
4. **Intent:** Was a specific actor knowingly deceiving? Wrongness, repetition, affiliation, emotional language, and financial benefit alone do not prove knowledge or intent. “Lie” should never be an automatic synonym for “false.”

A fifth field is needed for **evidential assessment**: supported, contradicted, insufficient evidence, mixed/conditional, or not an empirical claim. This is separate from all four above. An unsupported causal claim may eventually prove true; misleading context can surround a literally true fact. “Contested” describes disagreement, not its evidential merits. Do not create a score that averages these dimensions.

## What the empirical evidence supports

### Labels can help with the labeled claim

Martel and Rand (2024) analyzed 21 randomized experiments with 14,133 participants. Warning labels reduced belief in and reported sharing of false headlines on average, including among respondents low in trust in fact-checkers, although effects were smaller there. The reported average relative reductions were 27.6% for belief and 24.7% for sharing. These are reductions in experimental outcome measures, not percentage-point drops in population misinformation or observed real-world sharing. High coverage of false items and headline judgments limit transfer to occasional explanatory tags on educational pages. The article's abstract and methodological summary were available; the complete subscription text was not verified here. [Primary article](https://www.nature.com/articles/s41562-024-01973-x)

Pennycook, Bear, Collins, and Rand (2020) found an implied-truth effect: warning some false headlines can make unlabeled ones appear more accurate. A study condition that also marked verified true headlines removed that ambiguity. This supports measuring spillovers and displaying review scope; it does not prove that adding a “not reviewed” notice to this site's pages solves the problem. Do not treat a bare absence of tags as a quality certification. [Primary article and abstract](https://pubsonline.informs.org/doi/abs/10.1287/mnsc.2019.3478)

Epstein and colleagues (2022; 1,473 US participants) tested warnings attributed to a hypothetical hybrid AI/crowd system. Warnings improved sharing discernment; explanations of how the system worked showed only directional additional evidence and did not increase self-reported trust. This is not proof that a glossary explanation of a reasoning problem helps, nor that transparency automatically earns trust. [Primary proceedings article, abstract accessible and PDF linked](https://ojs.aaai.org/index.php/ICWSM/article/view/19283)

### Teach detection carefully; technique recognition is not truth recognition

Roozenbeek and colleagues (2022) tested short videos about five manipulation techniques in six randomized studies (6,464 participants) and a YouTube field study (22,632 responses). The work supports improving recognition of techniques and some related judgment measures. Crucially, the field outcome was identifying a technique in a test headline, not verifying an actual news claim or measuring organic sharing. The pooled field effect (h = 0.09) was below the preregistered smallest effect of interest (h = 0.10), with differences between the two videos. Different outcomes across studies should not be compressed into “prebunking stops misinformation at scale.” The content and findings are relevant precedents for lesson design, not proof for passive glossary badges. [Primary full-text author-institution PDF](https://research-information.bris.ac.uk/ws/portalfiles/portal/338515097/sciadv.abo6254.pdf)

Modirrousta-Galian and Higham (2023) reanalyzed five studies of gamified inoculation using receiver-operating-characteristic methods to separate discrimination from response bias. Their critique cautions that lower credibility ratings can reflect generalized skepticism rather than improved separation of true and false items. This targets particular games, datasets, and measures; it does not invalidate all inoculation interventions. The publisher abstract and author-manuscript location were identified, but the manuscript fetch failed in this research session, so no unverified effect sizes are used. [Primary indexed abstract](https://pubmed.ncbi.nlm.nih.gov/36996156/) and [institutional manuscript](https://eprints.soton.ac.uk/475360/2/2023_JEPG_Accepted_Manuscript.pdf)

Pennycook and colleagues (2024) tested emotional-manipulation inoculation across five studies with 7,286 online participants, using true/false content with emotional manipulation varied independently. Recognizing emotional manipulation improved without improved truth discernment; pairing inoculation with an accuracy prompt improved discernment by increasing belief in true content. This is especially important for this site: emotional rhetoric can accompany true claims. Its absence does not verify a claim. This assessment uses the accessible primary abstract, not the full subscription article. [Primary article](https://www.nature.com/articles/s41562-024-02023-2)

### Avoid educating readers into blanket disbelief

Guess and colleagues (2020) tested brief media-literacy tips in the US and India. Effects varied by sample and follow-up: online samples improved in discernment, while the rural face-to-face India study did not show the same result. Some effects persisted in the US but not India. This discourages assuming a universally effective “spot the trick” checklist. Reading environment and prior experience matter. [Primary open-access article](https://pmc.ncbi.nlm.nih.gov/articles/PMC7355018/)

Hoes and colleagues (2024) conducted three survey experiments across the US, Poland, and Hong Kong (6,127 participants). Interventions reduced belief in false information but also credibility of factual information; proposed alternative framings did not simply eliminate this trade-off. These are short-term experimental spillovers, not proof that every correction permanently erodes institutional trust. The page reports an author correction; use its updated version. [Primary open-access article](https://www.nature.com/articles/s41562-024-01884-x)

Altay, De Angelis, and Hoes (2024; preregistered US experiment, 3,919 participants) compared skepticism-oriented, trust-oriented, and mixed literacy tips. All improved discernment against control; mixed tips improved both true and false judgments, but many between-treatment contrasts were not significant. The results support including positive examples of justified confidence, with caution about declaring one framing superior. Outcomes used selected headlines and immediate responses, not long-term educational attainment. [Primary open-access article](https://www.nature.com/articles/s44271-024-00121-5)

Pfänder and Altay's 2025 meta-analysis covered 67 articles, 194,438 participants, and 40 countries. Participants generally distinguished true from fact-checked false news and leaned toward skepticism. Most participants were from the US or Europe, and results depended on stimulus selection: freshly sampled news differed from familiar fact-checked collections. This is evidence against assuming indiscriminate gullibility, not an estimate of all readers' everyday competence or the prevalence of misinformation. [Primary full text](https://www.nature.com/articles/s41562-024-02086-1)

### Corrections and verification need substance

Ecker and colleagues' 2022 review synthesizes cognitive, social, and affective drivers and correction research. A factual correction should explain the error and provide an intelligible account of what is supported, while acknowledging residual uncertainty. Do not avoid correcting a claim merely because a generalized “backfire effect” is feared; equally, do not promise full or permanent erasure of the original belief. The review's accessible summary was verified, but full text was not retrieved here; this memo does not rely on uninspected numerical claims. [Primary review](https://www.nature.com/articles/s44159-021-00006-y)

An unqualified “do your own research” instruction is insufficient. Aslett and colleagues (2024) found that asking people to search to evaluate misinformation could increase perceived veracity; low-quality search results and information environments matter. This does not show that trained verification or all searching is harmful. Link readers to the underlying evidence and demonstrate independent corroboration rather than simply telling them to search the original claim. [Primary article, methods and findings accessible](https://www.nature.com/articles/s41586-023-06883-y)

## Proposed schema

Store the analysis at the smallest meaningful unit: a quoted claim or a specific chart, image, or paragraph, with its time/place/scope. A topic page can contain several differently assessed claims and should not inherit a single warning status.

Suggested editorial record:

- claim_id, exact_claim, source_url, source_date, evidence_as_of, jurisdiction_or_scope
- assessment: supported | contradicted | insufficient_evidence | mixed_or_conditional | non_empirical
- assessment_summary: one plain-language sentence, with uncertainty explicit
- content_issues: zero or more controlled IDs, each with a claim-specific explanation and evidence links
- distribution: optional observations, each with evidence; distinguish known origin from earliest located copy
- learning_links: cognitive or verification concepts, introduced without claiming causation
- intent: not_assessed | insufficient_evidence | documented; separate actor, behavior, knowledge evidence, and attribution wording
- sources, reviewer, reviewed_at, corrections, next_review_trigger

These are editorial fields, not a required public dashboard. Prefer a compact public presentation: “What the evidence shows,” then a specific explanation such as “The chart compares totals from cities of very different sizes.” A nearby link can read “Why the denominator matters.” A generic chip alone is not an explanation.

Use no more visible labels than needed to understand the passage; a starting interface limit of one primary explanation and one related link is a testable design choice, not a research-established optimum. Support keyboard navigation and screen readers, use text rather than color alone, and avoid hover-only definitions. On mobile, keep the factual answer visible before supplementary taxonomy.

## Editorial gates and difficult cases

Before applying an issue label, the reviewer must write: (a) the exact claim/inference being evaluated; (b) the observable defect; (c) the evidence showing it; (d) the changed conclusion when corrected; and (e) why a reasonable alternate reading does not resolve it. If that cannot be done, publish uncertainty or explanatory context rather than a verdict.

- An omission is not automatically misleading. All writing omits information. The missing item must materially change the specific interpretation at stake.
- A minority conclusion is not false because it is unpopular. Explain evidential weight and methods rather than counting affiliations.
- Predictions should be evaluated against the information and uncertainty available when made. A low-probability event occurring does not alone refute a probabilistic forecast.
- Values and policy preferences can coexist with empirical premises. Assess the premises separately; do not tag the preference as misinformation.
- Satire, fiction, illustrative graphics, and disclosed synthetic images are not intrinsically misleading. Their later presentation as documentation of a real event may be.
- Correlation can support causal inference within an adequate design. “Correlation is not causation” must not become “observational evidence is worthless.”
- Financial interests and affiliations warrant disclosure when relevant but do not substitute for evaluating a claim.
- Do not infer coordinating actors from similar wording alone. Shared sources, copying, or independently convergent views are alternatives.
- Reserve “fabricated” for evidence of invented material. An unverified quote should initially be “unverified,” not “made up.”

For intent, reliable evidence might include authenticated contemporaneous instructions, admissions, or records showing the actor knew the material was false and represented it as true. A court or investigative finding should be attributed with scope and status, rather than converted into an unlimited site judgment. Distinguish creator from republisher. Legal/reputationally sensitive accusations require appropriate review.

## Evaluation plan before expanding tags

**Stage 1: editorial reliability.** Independently double-code a diverse pilot corpus with genuine counterexamples, unresolved claims, accurate emotional content, and overlapping issues. Record disagreement and time per decision. Report agreement by tag, not only pooled agreement; use raw agreement plus an appropriate chance-adjusted statistic, with prevalence caveats. Revise categories readers/editors confuse. Do not adjudicate uncertainty away merely to improve agreement.

**Stage 2: comprehension interviews.** Ask varied readers to think aloud on mobile and desktop. Ask what the label implies about the claim, the whole article, the author's intent, and untagged claims. Test whether “missing context” means “false” to them, whether they mistake confirmation bias for a personal accusation, and whether glossary detours impair understanding. Include readers with different literacy levels and accessibility needs. A small purposive interview sample finds failure modes; it does not estimate population effects.

**Stage 3: preregistered randomized evaluation.** Compare (A) evidence explanation alone, (B) explanation plus named tag and glossary link, and (C) explanation plus a plain-language question/link without the technical label. Keep facts, sources, length, and visual salience as comparable as practical. If resources permit, separately randomize an interactive worked example; do not bundle every desirable feature and call the result a tag effect.

Primary outcomes: comprehension of the corrected inference and calibrated judgments on new, unseen true/false/uncertain examples. Secondary outcomes: delayed retention and transfer after roughly a week, confidence calibration, recognition of uncertainty, source-checking behavior, and time/effort. Treat click-through and self-reported helpfulness as usage measures, not learning. Include true items and measure rejection of them; include unlabeled items to estimate spillover. Use novel test materials and alternate forms to reduce test-retest artifacts. Analyze both discrimination and response tendency. Sharing intention, if measured, must not be reported as observed sharing.

Specify a smallest worthwhile learning benefit and tolerable increase in false rejection before power analysis. Use pilot estimates and uncertainty to determine sample size; do not invent a universal N. Report confidence intervals and item-level variation, account for readers and stimuli, and distinguish preregistered analyses from exploration. Recruit beyond the project's enthusiasts. Subgroup analyses should assess accessibility and generalizability, not enable political targeting.

**Release gate:** retain the tags only if they improve a meaningful comprehension/transfer outcome or navigation without material misinterpretation or reduced acceptance of accurate information. If named tags add no value over concise explanatory links, keep the links and glossary but remove the badge layer. Reassess after content or audience changes.

## Remaining uncertainties

The reviewed literature does not settle the best label wording, number, placement, or emotional tone for this site; whether readers will voluntarily open explanations; whether learning transfers between statistical, historical, visual, and scientific claims; how many overlapping tags users can interpret; or whether “why this can mislead” creates less defensiveness than verdict language. It also does not establish that effects survive months, work equally across languages, or remain stable with generative-media changes. These are explicit empirical questions, not details to fill with confident design folklore.

## Search and access limits

This was a targeted critical research pass, not a systematic review with exhaustive database searching and formal risk-of-bias coding. Searches deliberately included adverse effects, response-bias critiques, mixed outcomes, and recent primary work through 2025; no claim is made to exhaust 2026 publications. Publisher/author/institutional sources were preferred. Several PMC and Nature opens returned access/redirect errors despite searchable content; the memo distinguishes full-text inspection from abstract-level support. Reported publication years come from the records, not the search engine's inconsistent relative-age labels. Samples and outcomes are study-specific. The repo-ready glossary is original proposed copy, not text copied from those taxonomies.