Publishers, regulators and platforms rebuild the newsroom on three fronts at once. Inside the building, editors wire generative systems into daily production and rewrite their own editorial guidelines around them. At the perimeter, publishers test which parts of their archive, their data feeds and their verification work they can sell to AI companies and to enterprise buyers. Outside, synthetic video, coordinated bot networks and platform-mediated summaries fill the information environment and reach audiences before the original article does.
This week’s links are about those three fronts. Copyright litigation and licensing keep defining what training data is worth. Public authorities in Europe and Asia move from principles to procedures, on cybersecurity testing, on takedown rules and on newsroom AI policies. Press agencies and large publishers turn trust itself into a product line, from Reuters on Snowflake to AFP’s factchecking service and Thomson Reuters’ own frontier model.
The pattern is direct. Regulators, audiences and enterprise buyers now ask journalism to function as reliable infrastructure in an information system that industrialises around it. The business question is whether that infrastructure can earn a price before the audience relationship becomes intermediated.
This image proposes a working hypothesis for the whole dispatch. A lighthouse in the middle of the Cyber Territories metropolis, a glass tower where journalists and AI systems work together, casts one defined beam of verified information across a city crowded with deepfakes, bot networks and manipulated imagery. Below the tower, secured data channels reach paying users, businesses and machines through subscriptions, digital products, B2B data services, API access and licensing partnerships. Quality journalism as reliable infrastructure in an industrialised information environment, priced and paid for through a portfolio of revenue lines. (Synthetic Image)
I. On law, licensing and the price of training data
One tension marks this chapter, the gap between how the United States and the European Union treat text and data mining, and how publishers learn to monetise the same corpus on both sides.
26.1 — US copyright fights over AI training remain unresolved, and do not translate to Europe
TechCrunch’s analysis of the ongoing US litigation over training AI models on copyrighted books sets out how uncertain the legal picture remains, even after the first round of district-court decisions. The piece frames the debate around the American fair use doctrine and the question of whether ingestion for training qualifies as transformative enough.
European readers should treat this as a United States conversation. The fair use doctrine does not exist in the European Union; the relevant framework is copyright law, and the moral and related rights that surround them. Presenting US outcomes as a general answer, misreads both the litigation record and the applicable European law.
Even inside the American system, extending fair use to full-corpus ingestion for commercial foundation models remains a considerable stretch. The cases that have gone furthest still involve narrow factual settings, and appellate review has barely begun. European rights holders should keep negotiating licences on European terms and treat US commentary as background rather than as precedent.
Reflections How should European publishers frame licensing offers when their US counterparts point to fair use style arguments as an alternative to paying? What safeguards will keep copyright enforceable when AI companies train models outside the EU and then deploy them inside it?
26.2 — USA Today reformats content for AI licensing, and Reuters joins Snowflake
Two commercial signals emerged in the same week. Digiday reports that USA Today Network reformats its content specifically to attract more AI licensing deals, restructuring articles and metadata so that AI buyers can ingest and cite them more easily. In parallel, Reuters announced that it has joined the Snowflake Marketplace to offer AI-ready, trusted news content to enterprise customers who want licensed, verifiable inputs into their own models and analytics pipelines.
Both moves recognise that the buyer changes. Publishers now package content for machines and for the teams that build with them, alongside the traditional B2C audience. The value proposition rests on provenance, structure and legal cleanliness, exactly the attributes that general web collection cannot guarantee.
Here lawsuits and licensing meet as two halves of one process. Litigation establishes the legal boundary of unlicensed training; marketplace deals establish what a workable price looks like once publishers enforce that boundary. Publishers who invest in structured, machine-readable formats today will hold more leverage in the next round of negotiations.
Reflections Which parts of a news archive create the most licensing value once publishers structure them properly? How can smaller European publishers pool their content into offers that enterprise AI buyers can read at scale, without ceding control over pricing?
26.3 — Thomson Reuters launches its own frontier model, Australia debates platform pay-up rules
Thomson Reuters announced that it leverages its world-class data assets to launch its own frontier model, aimed primarily at legal, tax and professional research use cases. The company bets that curated, high-value proprietary data plus a domain-specific model will beat generic frontier systems on the tasks its clients actually pay for.
At the political end of the same dynamic, AAP reports that Australia’s major parties agree on new pay-up rules for social media platforms, designed to secure payments to publishers whose content circulates on those services. This iteration extends the model that began with the News Media Bargaining Code.
Together these moves describe the same strategic logic from two directions. Owners of scarce, verified content move up the stack, either by building their own AI products or by requiring platforms to pay for the material that already fuels theirs. Quality journalism functions as a specialised input that carries a real cost of production, and European publishers should treat it that way.
Reflections When a large publisher operates its own frontier model, how should it price access for smaller newsrooms that could become both customers and competitors? Which regulatory design keeps platform pay-up rules from turning into a permanent subsidy that dulls the incentive to innovate on the publisher side?
II. On sovereignty, security and adversarial information
The state moves back into the AI and information stack, in Europe as well as in Washington and Beijing. Security testing, model oversight and influence operations converge into one file.
26.4 — France uses AI to test its own cybersecurity, and Brussels probes the labs
Reuters reports that France will use AI tools to test its cybersecurity after a hack at the tax agency, a project framed explicitly around national resilience and the country’s ambition to keep sovereign AI capacity, including Mistral, in the loop. In the same week, Euractiv obtained the letter in which the European Commission ordered leading AI labs to detail their security practices, covering red-teaming, model evaluations and incident reporting.
Together these two moves show a European stack forming around AI sovereignty. The debate now extends beyond what the United States and China do, with the EU asserting jurisdiction over how frontier models get secured, tested and audited on European infrastructure.
For newsrooms, sovereignty carries operational weight. It determines which models editors can integrate into editorial workflows without breaching data-protection or contractual obligations, and which vendors will certify that their systems meet the emerging European baseline.
Reflections How should a European newsroom weigh the operational advantages of frontier US models against the compliance and sovereignty case for European alternatives? Which security disclosures from AI labs genuinely help editors and CISOs, and which risk becoming compliance theatre?
26.5 — OpenAI disrupts a Russian ChatGPT campaign, and Yonhap sets the standard for how to report a platform-led disclosure
CNBC reports that OpenAI disrupted a Russian influence operation that used ChatGPT to generate propaganda content, part of the company’s periodic threat reports on state-linked misuse. Almost simultaneously, X said it identified a Chinese bot farm targeting the debate around AI data centres. The underlying story comes from Yonhap, the South Korean press agency, and the choice of source matters. Yonhap does exactly what a press agency should do with a platform-issued announcement of this kind: it reports the disclosure as a claim by X, attributes each element to the platform, adds independent context from South Korean cybersecurity researchers and public-sector sources, and refrains from treating the platform’s own communication as a verified finding. That editorial posture, precise attribution and independent framing, is the template other newsrooms should follow when they cover the same file.
Two things stand out in the underlying substance. First, the adversarial use of AI to generate content and to amplify it has become routine, and the actors behind it match those long identified in threat-intelligence work. Second, the disclosures come from the platforms themselves, which puts newsrooms in the position of reporting on evidence they cannot fully verify without cooperation from those same platforms.
Coverage should reflect that dependency, as Yonhap’s does. When a platform publishes a takedown announcement, editors should report it as a claim, seek independent confirmation where possible, and place it in the wider pattern of state-linked influence operations that intelligence services have documented for years.
Reflections What editorial standards should govern the reporting of platform-issued takedown claims, so that they inform readers without becoming free public relations? How can independent researchers get sustained access to the data needed to verify or challenge these announcements over time?
III. On verification, trust and editorial guidelines
Trust rebuilds through specific practices, most of them unglamorous. Factchecking desks, visual verification protocols and formal AI use policies do the day-to-day work of that reconstruction.
26.6 — AFP’s factchecking service and Reuters on visual verification
AFP’s dedicated factchecking desk on AI has grown into one of the most consistent public references on synthetic media, coordinated influence and manipulated imagery. A recent example illustrates the method. AFP debunked an AI-generated video that circulated as real footage of the Nepal-Tibet floods, ran it through Google’s SynthID detector, and documented the visual anomalies typical of AI generation. AFP’s own factchecking manual codifies the underlying approach: transparency about sources, prioritisation of harmful and viral claims, and detailed public explanations of the method behind each verdict. In parallel, Reuters published a piece explaining what visual verification is and why it practises it, describing the systematic checks its picture and video desks apply before publication.
These pieces also position both agencies commercially. A press agency that can demonstrate a documented, repeatable verification process turns that process into a market asset. Enterprise clients, platforms and other newsrooms have a reason to buy or licence access to it, precisely because independent verification at scale carries a high cost of production.
The wider pattern shows verification becoming a specialised industrial function, tightly coupled to press agencies as critical information infrastructure. Public broadcasters and quality news publishers can rely on it, integrate it and cite it, provided the underlying business model of the factchecking desks remains viable.
Reflections Which combination of subscription, licensing and public funding will keep large-scale factchecking desks sustainable over the next decade? How should a newsroom credit and cite external factchecks in its own reporting, so that readers understand where the verification actually happened?
26.7 — Editorial AI guidelines from Amman, EFE and CMN, and NYT tests AI search summaries
Three newsrooms formalised their AI policies in the same window. The Community Media Network in Amman amended its editorial policy to establish professional guidelines on AI use. Spain’s Agencia EFE published a public guide to AI use inside the agency. And Vietnam has officially adopted a formal process for handling fake and false news, which gives the government a codified procedure that will shape how domestic media operate.
On the product side, Semafor reports that The New York Times tests AI-generated search summaries on its own site, an experiment that continues alongside its ongoing litigation against OpenAI and Microsoft.
The common thread is procedural. Whether the driver is professional ethics in Amman and Madrid, state control in Hanoi, or product experimentation in New York, newsrooms and regulators move from principles to written procedures. That formalisation makes AI use auditable, both internally and for the audience.
Reflections Where should the line settle between disclosing AI use to readers in detail and overwhelming them with technical caveats on every article? How can a written newsroom AI policy stay useful when the underlying tools change every few months?
IV. On the shape of the newsroom, and its audience
The final chapter examines the newsroom itself, its people, its automation choices and the channels through which its work reaches readers.
26.8 — Automation inside the newsroom, from AP school boards to Sport in Spain, Nine and LatAm
Four pieces describe the operational reality. The Associated Press launched AP School Board Search, a tool built to transform local coverage of thousands of school boards by making local records searchable and reportable at scale. In Spain, WAN-IFRA describes how the sports daily Sport explores AI for its digital newsroom, integrating generative tools into production while keeping editorial control. In Australia, The Guardian reports that Nine Entertainment signs AI deals and cuts costs across the network. And Latam Journalism Review covers Google and Futura Lab launching an AI programme for newsrooms in Chile, Ecuador, Paraguay, Peru and Uruguay, with parallel growth support for eight South African news creators through Google’s Growth Lab and the AI for Media Network hackathon in 2026.
A separate signal completes the picture. The Blaze profiles a startup that automates investigative journalism and asks whether that automation serves the public interest at all. The answer depends on whether the automation extends what human investigators can do, or replaces the judgement calls that make an investigation defensible in court and in public.
Two patterns emerge. First, newsrooms deploy generative tools to unlock coverage that previously remained uneconomic, from thousands of school boards in the United States to sport verticals in Madrid. Second, a substantial share of the training, funding and tooling for that transition comes from a single dominant platform, which raises a question about the resilience of any newsroom that depends on it.
Reflections Which automation choices genuinely expand a newsroom’s editorial capacity, and which merely reduce headcount without adding new coverage? How should quality newsrooms structure their partnerships with dominant platforms so that they gain capability without ceding strategic autonomy?
26.9 — AI reshapes distribution, from Google’s AI search to Parag Agrawal’s future web, AFP–Dalet, and the newsletter question
TechCrunch, citing new data, reports that Google’s AI search rapidly becomes the default, which continues the shift of referral traffic away from ranked links and towards summarised answers. Parag Agrawal offers a longer view on AI search and the future web, describing a web that reorganises itself around retrieval by machines as much as by humans. AFP and Dalet have announced an integration that brings trusted content directly into newsroom production tools, a business-to-business distribution move rather than a consumer one. INMA asks whether newsletters will survive the AI assault on email, as inbox summarisation and AI-native readers start to intermediate the last direct channel publishers own. FT Strategies interviews Angela Mackay on what comes beyond reader revenue, pointing towards a mix of subscriptions, B2B products and licensing.
The distribution question has become structural. AI summaries reshape consumer traffic through general search; B2B channels and content marketplaces expand enterprise distribution; and AI assistants themselves mediate the direct channel of email.
Newsroom leaders should plan for a portfolio of distribution surfaces at once, general search, AI answer engines, licensed enterprise feeds, direct subscriptions, newsletters and audio, on the understanding that no single surface will remain stable enough on its own.
Reflections If a majority of consumer readers first encounter your journalism inside an AI answer rather than on your site, which parts of your business model still work? How should a quality newsroom decide which distribution surfaces to invest in, and which to accept as commodity channels?
Conclusion
Dispatch 26 traces the outline of journalism as infrastructure. Publishers, regulators and platforms rebuild the newsroom into something that resembles a public utility for verified information, with private capital, public interest and platform power pulling on it from three sides at once.
The legal front now has visible contours. Europe operates on copyright and related rights, with licensing agreements as the working currency. The United States operates on fair use, with unresolved litigation and settlements that have not yet established a stable price. European publishers should keep the two regimes distinct and negotiate on their own legal ground.
The security and sovereignty front moves fast. France testing its own systems with AI, the Commission auditing frontier labs, OpenAI disrupting Russian influence campaigns, X flagging a Chinese bot farm and Yonhap covering it with the professional caution the file deserves: these pieces belong to the same picture, in which governments and the industry treat the information layer as critical infrastructure.
The editorial front reaches the reader directly. AFP’s factchecking desk and Reuters’ visual verification protocols show what industrial-grade trust looks like when agencies build it properly. AI use policies at CMN, EFE and inside Vietnam’s official framework show how quickly the practices become codified. The New York Times testing AI summaries on its own pages shows that a single organisation can litigate against generative players and integrate their techniques into its own product at the same time.
The strategic question for the next quarter concerns which of those revenue lines a given newsroom can defend, and which it will need to build from scratch. The answer will differ by market, by size and by the strength of the underlying archive. What all cases share is that the audience relationship, the licensing relationship and the platform relationship now belong on one portfolio and on one strategic map.