Dispatch 25
Seventy years after a handful of European news agencies signed a founding text in Strasbourg, their successors meet again as EANA, the European Alliance of News Agencies representing 33 member organisations across 33 countries and more than 670 million people. The anniversary statement names the tensions that now define our week: copyright and content provenance, AI-driven disinformation, and threats to the safety of journalists.
Around that anchor, this week’s material sketches a wider picture. AI systems keep expanding into the writing, distribution, and verification of news, while regulators, publishers, unions, and readers each search for a viable equilibrium.
Three threads run through the signals. First, the industrial production of AI text and imagery is now measurable, and the tools to detect and label it are catching up. Second, the economic relationship between publishers and AI platforms is moving from anecdote to data, with early evidence that licensing deals shape citation shares. Third, the human premium, in journalism as elsewhere, is becoming a strategic argument, and one that press agencies as critical information infrastructure are well placed to make.
The Strasbourg text of 1956 spoke of cooperation between independent agencies. In 2026 that same principle, applied to AI governance and to the defence of verified information, is what gives the anniversary its practical meaning.
I. On press agencies and the anniversary that frames the week
One event sets the register for everything that follows.
25.1 — EANA marks seventy years and names the agenda for the next decade
The European Alliance of News Agencies celebrated its 70th anniversary on 21 August 2026, recalling its founding in Strasbourg in August 1956 and taking stock of a network that now spans 33 member organisations serving more than 670 million people. President Stefano De Alessandri stated that “verified information serves the public interest, regardless of national borders” and that cooperation is “the foundation on which independent journalism survives”. Secretary General Alexandru Ion Giboi added that “our members are what makes our organisation strong”, while AGERPRES Director General Claudia Nicolae described the Alliance’s legacy as one of “solidarity, cooperation, and a common voice”.
The statement identifies three pressures: copyright and content provenance, AI-driven disinformation, and threats to journalist safety. It also confirms EANA’s position on generative AI, summarised in the principle of “no consent, no content” for the use of news agency work in training. This is a clear European frame for the licensing debate, and it belongs alongside the copyright directive as part of the equilibrium that publishers, agencies, and AI developers are now negotiating.
For CEOs and policymakers, the anniversary is more than a commemoration. It positions news agencies as critical information infrastructure, distinct from news publishers and public broadcasters, and it names cooperation, innovation, and high journalistic standards, points highlighted by board member Aimilios Perdikaris of ANA-MPA as the conditions for economically sustainable journalism.
Reflections If “no consent, no content” becomes the European default for AI training on agency material, which licensing models can carry that principle into workable prices? How can a seventy-year-old alliance translate its heritage of cooperation into governance capacity for AI, provenance, and safety in the next decade?
II. On the industrialisation of AI text and the search for provenance
Detection, labelling, and measurement of machine writing all moved forward this week.
25.2 — Anthropic watermarks Claude, and the debate over what writing is
Anthropic has begun adding an invisible watermark to text generated by Claude, explained on Nieman Lab as a system that deliberately selects slightly less-likely words in detectable patterns, for example “guava” instead of “mango”. The Guardian reports that the change responds to an EU regulation requiring watermarking of AI-generated text from December, with the pattern detectable by Anthropic and by holders of a decoding key. Critics such as John Gruber argue that the method is “a perverse adulteration of what it means to write”, because it constrains the model’s choice of the most precise word.
The technical detail matters for editorial policy. Steven Murdoch of UCL considers the perceived quality loss likely to be small, since language models already rely on controlled randomness to avoid repetition and loops. At the same time, watermarking is one of the few tools that can address model collapse and provide a signal for detection at scale.
For newsrooms, the practical question is whether watermarking becomes a standard input for provenance workflows alongside SynthID and C2PA, or whether it remains a compliance layer visible only to the model provider. The European regulator has set a direction; publishers and agencies now need to decide how detection results are used inside editorial checks, and how they are disclosed to readers.
Reflections If watermarks become mandatory for large models operating in the EU, what verification duties should editors assume before they publish or quote AI-assisted text? How should a publisher weigh a marginal loss in prose quality against a measurable gain in traceability for the information ecosystem?
25.3 — Pew measures how much of the web is now written with AI
A Pew Research Center analysis of almost half a million English-language webpages estimates that one in ten pages collected in July 2026 show significant signs of AI authorship, rising to more than one third among pages published after ChatGPT’s release. The share reaches roughly one in ten on .com pages, against 4.6% on .org and around one percent on .edu and .gov. Compared with 2023, em dashes appear about twice as often, Oxford commas increase by 63%, AI-associated words more than double, and negative parallelism nearly triples.
The figures give a first stable baseline for a phenomenon that has so far been discussed in impressions. They also confirm the stylistic signatures that professional editors already recognise, and they suggest that institutional and academic domains have adopted AI writing more slowly than commercial ones.
The New Yorker’s essay on “the prehistory of AI slop” places the current wave in a longer history of mechanised writing, from eighteenth-century letter-writing manuals to Wycliffe A. Hill’s 1931 “Plot Robot”. Read together, the two pieces suggest that formulaic writing has a long history, and that today’s change lies in scale and cost, which have shifted by orders of magnitude and turned the current cycle into a governance problem for information ecosystems.
Reflections When one in three post-2022 web pages carries signs of AI authorship, what standard of disclosure should quality newsrooms adopt to remain distinguishable? How can historical perspective on formulaic writing help editors set proportionate rules for AI use, rather than reactive ones?
25.4 — When AI writes shorter, not just faster: the FT column that crossed a line
Poynter reports that the Financial Times appended a note to a column by Harvard professor Ricardo Hausmann, disclosing that AI had been used to condense a longer draft before submission, in breach of the FT’s editorial code, which prohibits AI in the writing process. The FT’s principles allow AI for research, transcription, and data analysis with human oversight, while insisting that journalism must be “reported, written and created by our journalists and editors”.
The case sharpens the question posed by the FT itself: when does an editing tool become a writing tool? Condensation reshapes emphasis and argument, and it is difficult to argue that a machine-condensed column is still fully authored by the byline it carries.
For editors, the practical takeaway is that AI policies need to reach outside contributors, including columnists and guest authors, and to specify which editorial acts remain reserved to humans. The line is not obvious, and every organisation will draw it in a slightly different place; making the rule explicit is the minimum condition for trust.
Reflections Which editorial acts, from copy-editing to structural condensation, should remain reserved to human authors even when disclosed? How should a newsroom respond to a contributor who has crossed that line, once disclosure has been made?
III. On synthetic content, information disorder, and elections
Where AI content is unlabelled or manipulated, the effects on public information are already visible.
25.5 — Deepfakes are the headline; election infrastructure is the risk
Rest of World argues that the real danger of AI in elections lies in the systems embedded in voter registers, biometric verification, and results software, rather than in the deepfakes that dominate coverage. The piece describes India’s 2015 programme linking election records to Aadhaar, in which about 5.5 million voters were removed in two states before the Supreme Court intervened, with a later algorithmic failure rate reported at 93%. It also flags growing reliance on private vendors across Africa, Asia, and Latin America for election technology, raising questions about foreign data transfers, unauditable proprietary systems, and chatbot misinformation from consumer AI products.
Poynter’s account of Median Strategies’ fake polls, which briefly influenced coverage of Los Angeles and Wisconsin races before real results contradicted them, makes the same point at a smaller scale. Poynter’s Kelly McBride calls reliance on an unknown pollster without published methodology “inexcusable” and reminds editors that polls are estimates, not predictions.
For democratic institutions and news organisations, the combined message is that AI-related electoral risk is a supply-chain problem. Auditability of vendors, transparency of methods, and disciplined use of poll data belong on the same governance list as deepfake response protocols.
Reflections Which parts of election infrastructure should be legally required to disclose the AI components they use, and to whom? How should newsrooms verify pollsters and datasets before amplifying figures that can affect turnout and expectations?
25.6 — Fake newsrooms, fake pitches, and the effort to re-humanise the process
Interlochen Public Radio reports that Prism News has shut down its network of more than 200 AI-generated local news and hobby sites, including titles that posed as local newsrooms under fabricated bylines such as “Dr. Elena Rodriguez”. The Israel-based startup had presented itself as an answer to news deserts but produced no original coverage, aggregating and slightly altering material from real outlets. In parallel, Nieman Lab describes The Food Section’s decision to accept story pitches only by phone from September, after founder Hanna Raskin estimated that a quarter of recent pitches were “wholly AI”.
Two responses, at opposite ends of the market, converge on the same insight: when synthetic content becomes cheap enough to industrialise, verification and human contact become scarce goods. Closing a 200-site network and opening a pitch hotline are complementary moves, one enforcing legality and reputation, the other rebuilding editorial trust at the front door.
The Verge’s report on Reddit’s experiment turning text posts into AI-voiced short videos, labelled “Real conversation voiced by AI”, shows the same logic on the distribution side. As synthetic formats proliferate, clarity about what is machine-produced and what is not becomes an editorial obligation, and a competitive asset.
Reflections What legal and platform standards should govern networks of AI-generated sites that impersonate local journalism? How can quality newsrooms turn “human by design” workflows, such as phone-only pitches, into a durable brand signal?
IV. On tools, skills, and the human premium
New tools are helping newsrooms verify, and new skill requirements are reshaping the labour market.
25.7 — Google’s Backstory and the shape of AI-assisted verification
Nieman Lab describes Google DeepMind’s Backstory, an experimental tool that helps fact-checkers determine whether images are AI-generated or manipulated and where they have appeared online. Built on Gemini models and AI agents, it checks SynthID watermarks and C2PA credentials, produces a cited report, and logs its steps. India Today team leader Bal Krishna says the tool has “no doubt” sped up daily work by combining tasks that once required “five different tools, five different logins”.
The design point is important: Backstory is presented as a first pass, not a source of citation. India Today does not cite it directly and requires reporters to verify findings manually. Digital literacy expert Mike Caulfield estimates it can reduce a 50-minute verification process to three minutes for indexed material, while noting that images from TikTok, WhatsApp, or Signal remain outside its reach.
For editors, the implication is a workflow question. Verification will increasingly rely on assisted tools whose outputs must themselves be validated, and provenance signals such as SynthID and C2PA will only be useful if editorial systems can read and record them at scale.
Reflections Which verification steps should remain manual as assisted tools such as Backstory become standard first-pass instruments? How should newsrooms disclose to readers the role of AI in their fact-checking, without weakening confidence in the finding?
25.8 — Recruiters raise the AI skills bar, and the “human premium” grows
Journalism.co.uk reports that the NCTJ has found news employers increasingly demanding advanced AI skills from journalism graduates, including critical prompting, fact-checking, context awareness, ethical judgement, and understanding of bias and hallucination. Its research covered 14 UK news organisations and shows that every one has adopted AI to some degree. In parallel, Forbes summarises reports from LinkedIn, Visier, and the World Economic Forum arguing that AI is creating a “human premium” around judgement, communication, curiosity, critical thinking, and interpersonal intelligence.
Read together, the two pieces describe a bifurcated skills market. Technical AI literacy is now baseline, while the returns to distinctively human capabilities, applied to editorial judgement and audience relationships, are increasing.
For news agencies and publishers, this argues for two coordinated investments: structured AI training for staff at every level, and explicit recognition of the editorial and interpersonal work that AI cannot substitute. Boyd Taylor Coolman’s essay in America Magazine on Pope Leo XIV’s warning, in “Magnifica Humanitas”, that AI can “extinguish the desire to ask questions” offers a philosophical counterpart. Coolman’s recovery of a medieval account of intellectual virtue is a reminder that reading, thinking, and questioning are worthwhile in themselves, a point with practical consequences for how newsrooms design work.
Reflections Which AI skills should be included in the core curriculum for every journalist, and which should remain specialist tracks? How should organisations recognise and reward the human capabilities that AI cannot replace, without turning them into a nostalgic argument?
25.9 — Cooperation and skills at continental scale
Two international items complete the skills picture. Africanews reports on the 7th Forum on China-Africa Media Cooperation in Beijing, where African Union of Broadcasting CEO Gregoire Ndjaka stressed that Africa must “develop the skills to understand, adapt and use” AI technologies rather than simply consume them. Mediabites summarises WAN-IFRA CEO Stig Kirk Ørskov’s first-summer assessment of global journalism, in which he warns that AI-powered search and chatbots increasingly answer directly and remove publishers from the referral chain, calling the current platform arrangement “not a sustainable foundation for a healthy information ecosystem”.
The two views reinforce each other. Skills, cooperation, and the defence of professional journalism are converging into a single agenda, and it is one that international alliances, from EANA to WAN-IFRA to the AUB, are increasingly expected to lead.
For strategy leaders, the practical question is how to align internal training programmes, external partnerships, and public advocacy so that the “human premium” is not only an HR argument but a distinctive market position.
Reflections Which cooperative structures are best placed to build AI skills across small and mid-sized newsrooms in Europe and beyond? How should industry alliances translate skills investment into pricing power when negotiating with AI platforms?
V. On law, licensing, and the economics of AI in news
Rules and money are moving together, and the shape of a workable equilibrium is starting to emerge.
25.10 — The EU AI Act as an editorial reference beyond Europe
ISS Africa argues that South African newsrooms and legacy media should treat the EU AI Act as a reference for transparency, accountability, and information integrity in an African context. Konrad-Adenauer-Stiftung Media Programme Director Hendrik Sittig is quoted as saying that the Act permits newsroom AI use while requiring “greater transparency, clear labelling and accountable editorial oversight”. A CINIA study cited in the piece finds that South African AI adoption often depends on individual employee interest, while few newsrooms have clear policies and some journalists reportedly use AI “secretly”.
The article documents the risks that result, including fake citations in a draft National AI Policy and alleged non-existent case law in a Johannesburg acting judge’s ruling. It also references the Press Council of South Africa’s guidance that AI-generated material be checked by “human eyes and hands”, fact-checked, and transparently identified.
For European agencies and publishers, the wider point is that the AI Act is becoming an editorial and reputational reference well beyond the EU’s jurisdiction. Consistent labelling and oversight practices can travel across markets, and the CDSM copyright framework offers a companion baseline for the licensing conversation.
Reflections How can European newsrooms help partner organisations elsewhere adopt AI Act–compatible transparency standards, without imposing them? Which elements of the AI Act should the industry itself codify into operational rules before regulators need to enforce them?
25.11 — Musicians, artists, and the limits of licensing without consent
Bloomberg reports that musicians are pushing back against studios using their songs to train AI, noting that record labels can license their catalogues but that AI companies cannot manipulate specific tracks without artist consent. The distinction between rights held by labels and rights that remain with performers is central to the current dispute, and it mirrors the position taken by news agencies on their own material.
MIT News describes CSAIL research identifying “attribution decay”, a phenomenon in which large training datasets make individual images increasingly difficult to trace inside a generative model’s output. Zheng Dai and Professor David Gifford show that at sufficient scale, removing a single image, an artist’s entire body of work, or every photograph of a person can leave outputs unchanged; Gifford argues that the findings raise questions about copyrightability, author compensation, and whether models should be revised to produce outputs guaranteed to be unattributable to individual creators.
Read together, the two pieces sketch a legal frontier. Consent and licensing remain the working framework in Europe, and the CDSM’s opt-out mechanism gives a first structure, while technical work on attribution and provenance is likely to shape how far claims can be brought and priced. This is law in search of equilibrium, in the strict sense: claims will establish the boundaries, and negotiations will set the workable prices.
Reflections How should European copyright practice adapt when attribution to individual works becomes technically difficult at scale? What licensing structures can protect performers and journalists whose consent has not been secured by the entities that hold their catalogue or archive?
25.12 — Licensing shows up in citation shares, and Google offers publishers a new button
A Press Ranger and OtterlyAI study of 129.3 million AI citations across seven platforms finds that publishers with OpenAI licensing deals receive 48% more citations per page on ChatGPT than unlicensed publishers, and 46% more on average across all platforms. News accounted for 7.2% of AI citations, with Future plc, Forbes, People Inc., Condé Nast, and Hearst capturing 69% of citations to licensed publishers. OpenAI is the only licensor showing a clear home-platform advantage; Google-licensed publishers see slightly lower citation rates on Google AI Overviews, and Perplexity-licensed publishers reach parity on Perplexity.
TechCrunch reports that Google is giving publishers a “Preferred Sources” button they can embed on their sites, allowing readers to mark them as favourites across Search, Discover, and Google News. Google cites earlier studies showing readers are twice as likely to click through when a preferred source is available; more than 345,000 unique sources had already been selected by users as of May.
For strategy leaders, the numbers matter more than the labels. Citation share and referral traffic are becoming the currency of the AI-mediated web, and licensing agreements now have measurable effects on both. The Columbia Journalism Review’s account of the Axios–OpenAI three-year deal, funding thirteen local newsletters and training OpenAI’s models on Axios’s free content, and the NewsGuild’s coordinated “No Slop in Our Shops” protest at USA Today Co. against a Palantir partnership show the political economy of these arrangements from two sides. Deals are becoming a business tool; they are also becoming a governance and labour question.
The Editor and Publisher piece on the agentic-interface era, and the argument that the filter bubble is morphing into something new as agents mediate access to information, completes the picture. As agents choose which sources to consult, the value of being a preferred, licensed, and verifiable source will compound.
Reflections If licensing deals now show a measurable citation premium, how should mid-sized publishers evaluate the trade-off between short-term revenue and long-term dependency? Which governance mechanisms should be built into publisher–platform deals to protect editorial independence and staff conditions?
VI. On press freedom and the fragility of public information
Two items remind us that the political economy of information rests on the safety of journalists and the credibility of investigation.
25.13 — Bangladesh criminalises press-conference questions
The Committee to Protect Journalists reports that Bangladesh’s International Crimes Tribunal has issued fresh warrants for journalists Shyamal Dutta, Mozammel Babu, and Farzana Rupa, alleging that questions they asked at a July 2024 press conference incited a crackdown on demonstrators. All three have been detained since 2024, and, if convicted, could face the death penalty because the Tribunal has no sentencing guidelines for specific offences. CPJ Asia-Pacific Program Coordinator Kunal Majumder called the case “outrageous and deeply troubling” and said editorial decisions “cannot be criminalized”.
This is a European story too. Safety of journalists is one of the three challenges named in the EANA anniversary statement, and international solidarity through alliances such as EANA and WAN-IFRA is one of the few instruments that can raise the political cost of such prosecutions.
For readers of this dispatch, the case is a reminder that AI governance and press freedom belong on the same agenda. A weakened press in one region reduces the supply of verified information for everyone.
Reflections Which international mechanisms are most effective when the act of asking a question is treated as a criminal offence? How should European agencies and publishers structure their solidarity work when press freedom risks converge with AI governance risks?
25.14 — Investigative journalism and the ground it stands on
Craig Unger’s Substack essay argues that investigative journalism has lost much of its impact because Americans no longer share a common factual reality, while social-media algorithms, silos, and weakened institutional checks prevent revelations from producing accountability. He contrasts the effects of Seymour Hersh’s My Lai reporting, the New York Times’ publication of the Pentagon Papers, and Woodward and Bernstein’s Watergate coverage with the political response to more recent scandals, and points to media consolidation, billionaire ownership, and platform capture as structural changes.
Whatever one makes of the argument as political analysis, its editorial implication is important. When information ecosystems fragment, the marginal value of verified, professionally produced journalism increases, and so does the responsibility of institutions that supply it. The Audiencers’ article on synthetic audiences, and its rule to use them “to narrow, never to decide”, makes the same point at the level of research method: AI can help newsrooms sharpen questions, and human evidence must confirm the answers.
For CEOs and editors, this argues for renewed investment in investigation as an infrastructure, not a project, and for a public case that ties AI, licensing, and press freedom into one coherent story.
Reflections If revelations no longer move the political needle, what new formats and partnerships can rebuild the link between investigation and accountability? How should synthetic-audience tools be integrated into editorial planning without displacing the human research that anchors quality journalism?
Conclusion
The Strasbourg text of 1956 was written by agencies that had lived through the collapse and rebuilding of European information. Their successors marked the anniversary this week by naming the tensions of 2026: copyright and provenance, AI-driven disinformation, and the safety of journalists. The choice of these three, rather than a longer list, is itself a statement of priorities.
Around that anchor, the week’s material describes an ecosystem moving from anecdote to measurement. Pew tells us how much of the web is now written with AI. Press Ranger and OtterlyAI tell us how much a licensing deal is worth in citation share. The EU AI Act tells us how transparency will be enforced from December, and the CDSM tells us how consent and remuneration should be framed in copyright. Backstory and SynthID tell us how provenance will be checked. Each of these is a piece of infrastructure that publishers and agencies will need to master.
Two things follow. First, the “no consent, no content” principle set out by EANA is more than a slogan; it is the European answer to a global question about training data, and it belongs at the heart of the licensing conversation. Second, the human premium documented by employers and researchers is not only an HR argument; it is a strategic position for the quality newsroom and for press agencies as critical information infrastructure.
The founding text of 1956 was signed in a specific room in Strasbourg. Seventy years on, the practical work of that same cooperation runs through workflow decisions, licensing negotiations, transparency reports, and training programmes. The anniversary matters because it names the direction; the next decade will be measured in how consistently the direction is held.
For CEOs, policymakers, and newsroom leaders reading this dispatch, the immediate question is straightforward. Which of the mechanisms surfaced this week, transparency, licensing, verification, skills, and cooperation, do our organisations already run well, and which need to be built before December, before the next election cycle, and before the next anniversary?