On 31 August the European Alliance of News Agencies joined the SPUR Coalition as an affiliate member, bringing 33 European agencies into a body that defines technical standards and commercial infrastructure for the use of news content by generative AI systems. One membership decision rarely changes a market. This one is worth to look at more closely, because it marks the moment when the agency layer of the European information system prepares to negotiate through shared standards rather than one contract at a time.
This week therefore we will draw a picture of what the members of that coalition and there media clients are actually doing. The Danish agency Ritzau put official emergency messages directly onto citizens’ phones. The Taiwanese press agency begins selling machine access to thirty years of archive by monthly subscription. The American agency AP moves video production to India. A Swiss publisher signed a group-wide enterprise agreement with OpenAI while avoiding to license its journalism. A Regional Court in Munich decided that an AI model had memorised protected works, and two American newspapers filed a fresh copyright claim in New York.
These signals are the signs of an economy under construction. Content is becoming a licensed, measured and machine-readable, distribution is becoming a question of who controls the access layer, and the value of a newsroom is defined by what it can prove about provenance, rights and accuracy.
Markus Knall of Ippen.Media brought back a useful term from an AI and media exchange hosted by Mediahuis in Brussels, the barbell model: AI splits the media market into boutique journalism at one end and large efficient services at the other, while the thin middle integrated both extremes as a lean operational core. This dispatch uses that hypothesis as a reading frame, and treats the sector news from EANA and SPUR members as a barometer of where the weight is settling.
I. On the sector organising itself
One decision that changes the level at which the industry negotiates.
27.1 — EANA takes the agency case into a standards coalition
EANA, founded in Strasbourg in 1956 and headquartered in Bern, has become an affiliate member of the SPUR Coalition, whose members include the BBC, the Financial Times, Sky News, Guardian Media Group, Telegraph Media Group, Mediahuis and the Associated Press. The News Media Coalition confirms that the arrangement gives EANA’s 33 agencies reduced-rate SPUR membership and a seat in standard-setting on content licensing and usage transparency, and the move was also picked up by the Slovak agency TASR. Secretary General Alexandru Giboi framed the affiliation as bringing the perspective of Europe’s leading agencies into what he called one of the most consequential debates of today.
The strategic content of this decision lies in the word standards. Licensing negotiations between a single agency and an AI model do little more than setting a price. A shared reporting standard for how generative applications use news content produces something more durable: a measurable unit of consumption, an audit trail, and a basis on which many rights holders can price the same behaviour consistently. For agencies, which supply media, companies, institutions and intermediaries rather than consumers, that measurement layer is the commercial foundation of the entire business.
There is a second effect worth naming. Agencies and publishers have overlapping interests and genuinely different positions, since an agency sells rights into the market that publishers themselves resell downstream in altered form. Putting both inside one standards body creates options for those positions to be reconciled in technical definitions instead of in litigations and opportunism. That is how infrastructure industries have historically settled: the specification comes first, and the tariff follows.
Reflections
Which decisions should a standards coalition be allowed to take on behalf of its members, and which must remain with each members own board?
How can we know, twelve months from now, whether shared measurement has improved the price of news content or merely standardised it?
II. On press agencies as critical information infrastructure
Three signals in which agencies define the layer through which official, verified and machine-readable information reaches the market.
27.2 — Ritzau places official messages directly in citizens’ hands
On 2 September, Ritzau launched DK Update, an app carrying operational messages from all Danish police districts, the Prosecution Service and Emergency Services Copenhagen, available in the App Store and Google Play. The service extends an existing Ritzau product that already distributes authority messages to virtually every Danish newsroom, and the agency states the ambition of including as many authorities as possible so that citizens can consult one place. The product manager describes the launch as a response to demand from media and journalists for direct notification when new information becomes available.
This is a business-to-consumer product built on a business-to-business asset, and that combination is the disrupting part. The value in DK Update is the pre-existing authenticated pipe between public authorities and the agency, with its verification discipline, its editorial standards and its uptime obligations. Agencies such as Belga operate that pipe exclusively towards media, companies, institutions and intermediaries; Ritzau decides to use the same infrastructure to serve the public directly, concluding that reliability, and not reach, is the scarce good.
The broader pattern concerns the definition of a press agency in a period of synthetic content. An institution that can guarantee the origin of an official message performs a function closer to a utility than to a publisher. Governments and regulators tend to notice such functions late, and to regulate them only after a failure.
Reflections
When an agency distributes official messages straight to citizens, what editorial responsibility does it take on beyond accurate transmission?
Where should the line run between a public authority’s own communication channel and an independent agency’s infrastructure?
27.3 — Agencies start selling access to machines, by subscription and by integration
Taiwan’s Central News Agency launched a subscription-based Model Context Protocol service on 31 August, letting applications such as ChatGPT, Claude and Grok query nearly five million news stories from more than thirty years, 3.5 million photographs spanning a century and public data from around 150 government agencies, with source citation built in, at NT$200 a month. AFP took a different route, announcing with Dalet an integrated newsroom experience in which AFP video arrives inside Dalet Flex with its scene-by-scene editorial metadata, shot composition, locations, soundbites and linked articles preserved. Belga News Agency, where the author of this newsletter serves as CEO, announced a partnership with FirstHour.ai covering the crisis communication chain from incident detection and AI-assisted drafting through approvals to distribution via Belga.share and post-crisis analytics, with NIS2 and DORA documentation as the reference frame.
Three products, one commercial logic: the customer is a system , and the deliverable is structured, rights-cleared, attributable content with a meter attached. CNA prices machine access explicitly and states a public-interest reason for it, namely reducing inaccurate information about Taiwan by supplying verified local reporting and Traditional Chinese terminology. AFP prices the preservation of editorial context, which is what makes archive material usable at speed. Each approach converts metadata discipline, long treated as a cost centre, into the reason a customer pays.
Ana Jakimovska of Mediahuis captured the shift in Knall’s account of the Brussels session mentioned earlier: B2C becomes B2A2C, and B2B becomes thus B2A2B, with content read first by machines and only afterwards by people. Agencies have an advantage here, because they have priced structured delivery to third parties for a century. The scarce asset is a catalogue whose provenance can be demonstrated line by line, and there is no such thing as a free lunch in producing one.
Reflections
What is the right unit of pricing when the customer is an agent: the query, the token, the citation, or the commercial outcome it supports?
How much of a newsroom’s metadata discipline should be sold as a product, and how much retained as a competitive advantage?
27.4 — Agencies rebuild production and commercial bases
The Associated Press is laying off 20 United States video production staff and moving the work to a centralised hub in Noida, with a further 10 American photo editors and 10 London video staff identified, following the elimination of roughly 8% of its global workforce in 2024; executive editor Julie Pace says the hub will scale video and photo capabilities while improving speed and quality, and the News Media Guild questions how well staff outside the country will know American topics and regions. NTB has appointed Lars Bratsberg to lead commercial growth, a hire from Spond with nine years at Google behind him, as the agency sells news, communication services and technology products into 22 countries and reorganises its picture agency of more than 500 million assets. Alongside these, dpa is continuing its Super Searchers information-literacy programme with BAGSO after training more than a thousand mostly library staff, Kazinform and the Albanian Telegraphic Agency signed a content and training memorandum on 1 September, and Ukrinform is publishing a book drawn from a year of commander interviews for the Lviv Publishers’ Forum.
Visual output represents more than 80% of AP’s production, which makes the cost per finished asset the decisive variable in the agency model; centralising that work reduces unit cost and transfers local knowledge risk to the editing layer. NTB’s hire points the other way, towards revenue: an agency that sells software into 22 countries needs commercial leadership from technology markets rather than from print sales. Both are the same strategy at different ends of the profit and loss account.
The pattern to watch is specialisation. dpa monetises trust through media literacy, Ukrinform monetises a documentary archive through publishing, Kazinform buys reach through bilateral agreements. Agencies that identify which scarce capability they own tend to leverage it; those that continue to create undifferentiated volume compete against the marginal cost of generation, which is heading towards zero.
Reflections
How far can visual production be centralised before the local knowledge that made it credible stops arriving with the footage?
Which capability inside your own organisation is genuinely scarce, and what evidence would convince a client to pay a premium for it?
III. On governance at the base of the system
Two agencies and one publishers’ association show how quickly institutional independence becomes an operational question.
27.5 — Independence is tested at the level of contracts and appointments
The interim leadership of Hungary’s public media fund MTVA announced it will file a criminal complaint after an internal review concluded that the organisation had been directed outside its formal governance, with a professional consultant, contracted as an adviser since 2019 allegedly deciding personnel and operational matters; staff reportedly referred to an invisible fifth floor in a four-storey building. In Bulgaria, BTA director general Kiril Valchev is taking unpaid leave to run as a vice-presidential candidate, saying he had not sought the nomination and citing more than twenty years of legal practice. In Hungary, the publishers’ association MLE announced two new members while acting as a co-regulatory participant in the committee assessing candidates for the presidency of the Media Council and nominating a member to the Independent Public Media Board. Ukrinform, meanwhile, hosts a round table on 11 September on the ethical use of AI in Ukrainian newsrooms, building on a policy template already adopted by seven outlets.
Editorial independence is usually discussed as a principle and is always at peril to be lost as a procedure. The MTVA allegations describe informal authority operating through advisory contracts and message traffic, which is precisely the space that governance codes tend to leave undefined. The BTA case raises the mirror-image question of what an agency chief executive’s public candidacy does to the perceived neutrality of the wire, leading to unpaid leave as an answer.
The Ukrainian round table shows the constructive version of the same challenge. Written newsroom AI policies, adopted outlet by outlet and tested against legal and ethical requirements, convert a contested principle into an auditable procedure. Institutions that document who decided what, and when, are considerably harder to capture.
Reflections
What safeguards would make advisory contracts inside a public media organisation transparent enough to be verifiable from outside?
Which political roles should an agency’s chief executive be able to seek with unpaid leave as available remedy?
IV. On publishers testing the barbell
Signals from SPUR members and the wider publishing market, read against the Barbell hypothesis.
27.6 — Ringier separates AI tooling from content licensing
Ringier and OpenAI have launched a multi-year collaboration providing long-term access to OpenAI’s models and a phased rollout of ChatGPT Enterprise across the group under European data storage, covering employee enablement, software development, marketplaces, editorial work and corporate functions. The agreement explicitly excludes any licence of Ringier’s content, and chief executive Marc Walder states that technology supports people and does not replace journalistic responsibility. SRF’s analysis examines the tension with Walder’s earlier description of unauthorised use of journalistic content by AI as the greatest theft in the history of journalism, and points out that Switzerland has no binding rules on declaring AI use or its extent in editorial products.
The construction here is more precise than the commentary suggests. Ringier is buying enterprise capability while withholding the asset it intends to price separately, which is a rational sequencing decision for any rights holder that expects a licensing market to mature. The exposure lies in the operational detail: enterprise deployments involve prompts, documents and workflows that carry editorial knowledge, and the contractual boundary around that material determines whether the separation holds in practice.
For the sector, this is the template most large European groups will examine. Buy tooling under European data terms, keep the content licence as a distinct negotiation with its own price and its own machine-readable reservation of rights, and publish the internal rules on declaration before a regulator writes them.
Reflections
Which contractual guarantees would satisfy you that enterprise AI deployment leaves your content licensing position intact?
What should readers be told about AI use in an editorial product, and who is best placed to set that standard?
27.7 — The middle of the market is where the pressure lands
Sipa Ouest-France has identified 12 million euro of payroll savings to avoid a redundancy plan at France’s most widely circulated daily, with the sale of its Paris headquarters and radio stations expected to raise more than 14 million euro by the end of 2026. Bonnier News is consolidating 35 salespeople into a single hub for print advertising across Dagens Nyheter, Expressen and Mitt i. Sanoma Media Finland announced on 4 September that refusing certain cookies will require a paid subscription from October, with the fee set against the advertising revenue foregone, citing an August precedent from Finland’s Supreme Administrative Court. CMA Media launched the free RMC+ platform on 1 September with nearly 10,000 hours of content, while 93% of BFMTV and RMC advertising revenue still comes from linear television.
Each of these is a decision about where a publisher intends to sit on the barbell. Ouest-France is defending scale in a regional franchise by cutting fixed cost. Sanoma is pricing consent, which converts a compliance obligation into a revenue line and tests how much a Finnish reader will pay to refuse tracking. CMA Media is buying optionality in streaming while acknowledging that its cash still comes from broadcast advertising.
The barbell hypothesis predicts that the thin part of the bar is the general-interest, mid-scale, advertising-funded product with no distinctive rights position. What these five signals show is that publishers are aware of it and are choosing an end deliberately. The decisive test will be whether the products at each end can be financed from the same balance sheet during the transition.
Reflections
If a reader can pay to refuse tracking, what does that price reveal about the real value of consent in your market?
Which position on the barbell is your organisation actually financing today, judged by where new investment goes?
V. On rights, regulation and case law
Three signals in which courts, competition authorities and legislators define the terms on which content and synthetic output circulate.
27.8 — Munich holds that memorisation by a model is reproduction
On 31 July the 42nd Civil Chamber of the Regional Court of Munich I ruled largely in favour of GEMA against Suno in case 42 O 763/25, granting injunctive relief, information and damages claims concerning six works including “Atemlos durch die Nacht”, “Forever Young” and “Daddy Cool”. The court found that the v3.5 and v4 models had memorised reproducible parts of works obtained from YouTube through stream-ripping that circumvented its rolling cipher, held that the American training reproductions were outside the exception invoked in defence under 17 U.S.C. §107, and held that the reproductions inside the models and in their outputs in Germany fell outside the text-and-data-mining exception in §44b UrhG. GEMA, which describes the judgment as its second court victory against an AI provider after the November 2025 OpenAI ruling now on appeal, states that licences must be acquired; the judgment is open to appeal. The French-language analysis by Marie-Avril Roux Steinkuehler stresses the same distinction the court drew, between analysing data and storing recognisable versions of protected works.
Two American filings frame the contrast. The United States government filed a brief on 1 September supporting OpenAI in the New York Times litigation, arguing that AI training is extraordinarily transformative and that the country has a strong interest in rejecting the newspapers’ position; a brief carries advisory weight, the first two judges to address the question reached diverging conclusions, and the outcome remains open. Four days later, the Seattle Times and Newsday sued OpenAI and Microsoft in the Southern District of New York, alleging that paywalled journalism was copied into training datasets and seeking destruction of the datasets and models involved. Stibbe counts more than 70 active or recently resolved AI copyright actions, notes that Article 4 CDSM allows a machine-readable reservation of rights, and flags the Advocate General’s opinion in Like Company v Google, C-250/25, expected this month.
The European position is being established through copyright reasoning rather than through imported doctrine, and the Munich chamber has located the decisive question with precision: the boundary is crossed where analysis becomes storage of recognisable protected material. Claims of this kind establish the legal limits; licensing negotiations establish the workable price; a business model emerges from the combination of the two. Both halves are necessary, and neither substitutes for the other.
Reflections
How should a rights holder express a machine-readable reservation of rights so that it is enforceable and auditable rather than merely declared?
What would a licence look like that priced memorisation separately from retrieval, and who would verify the difference?
27.9 — Regulators move from principles to instruments
EU antitrust regulators sent publishers a questionnaire in July, with an 28 August deadline, on Google’s proposal to let them opt out of AI search without losing traditional search ranking; the opt-out was announced in June, completed globally on 31 August after a UK order, and the investigation could lead to a further fine on top of more than 10 billion euro already imposed. On 1 September the European Commission designated ChatGPT as a Very Large Online Search Engine under the Digital Services Act, alongside Reddit and Roblox as platforms, giving them until January 2027 to assess and mitigate systemic risks. South Korea’s media commission has requested 34.6 billion won for 2027, including 10 billion won to train AI on broadcasters’ stock footage and funding for a fact-checking transparency centre.
Each instrument reveals a different signal about the same market. The competition file asks whether an opt-out can be genuine when the same company controls the ranking that publishers depend on. The DSA designation of a conversational service as a search engine confirms that answer interfaces have become the gateway, with the systemic-risk obligations that follow.
Taken together, this is regulation shifting from statements of principle to countable instruments: a machine-readable opt-out, and a designated gateway with audit duties, Instruments create evidence, and evidence is what makes a licensing market function.
Reflections
Under what conditions would you use an AI search opt-out, and what would you need to see in the traffic data before trusting it?
What systemic risks can be created by a chatbot that qualifies as a search engine?
27.10 — Rules for synthetic content and information conduct take shape
Brazil’s top electoral court defined a deepfake, by a 5-2 vote, as synthetic content produced or manipulated through AI or equivalent technology that is realistic enough to create, reproduce or alter the image, voice or speech, setting new standards for identifying AI-generated deepfakes in political ads, ahead of a first presidential round on 4 October. In China, regulations on multichannel distribution services took effect on 1 September, requiring MCN agencies to register, appoint a content-management director, maintain a permanent monitoring team and terminate creators who breach the rules, in a sector of more than 40,000 agencies and roughly 100,000 signed influencers; the reporting by the Croatian agency HINA, notes an estimated 100 billion yuan market. The same authorities reported on 3 September the removal of more than 5.61 million items and action against over 49,000 accounts in a campaign against AI misuse, building on labelling rules in force since September 2025. Mexico’s draft guidelines on audience rights would require broadcasters to separate news from opinion and appoint ombudspersons, with fines up to 1% of annual revenue, and have drawn criticism from CIRT and from Reporters Without Borders over ambiguity and anonymous complaints.
These four measures are attempts to regulate the new reality through different constitutional tradition based on the reliability of what an audience sees. A court definition of synthetic content is narrow and enforceable in an electoral context. An administrative registration duty for intermediaries is broad and gives the state leverage over the entire creator economy. A rights-of-reply regime for broadcasters addresses accountability and invites the objection that the regulator ends up assessing editorial judgement.
The comparison matters for European practice because the technology moves faster than the safeguards. A workable definition of manipulated media, tied to a specific harm and a specific procedure, is more useful to newsrooms than a general prohibition, and considerably harder to write.
Reflections
Which legal definition of synthetic content could a newsroom apply in the hour before publication rather than in litigation afterwards?
Where does an audience-rights regime stop protecting audiences and start shaping editorial judgement?
VI. On the economics of attention and the cost of verification
Two signals on how the open web now generates revenue, and on what it costs to keep the record accurate.
27.11 — Machine visits grow while human sessions decline
Digiday reports that the open web is separating into human audiences and AI crawlers, citing People Inc.’s 22% year-on-year fall in core sessions and 40% decline in Google search traffic against flat session-based revenue, 16% growth in non-session revenue and a rise from 39% to 43% of digital revenue; chief innovation officer Jon Roberts states that the open web is getting smaller, while Forbes describes AI licensing models as nascent. Press Gazette reports 51Degrees analysis of three billion visits showing that AmazonBot accounted for more than 4% of all visits and Meta-External-Agent almost 3%, that AI bots reached 13% of web sessions in May 2026 against 8% a year earlier, and that publishers block only around 21% of AI scrapers in robots.txt. The Journalism Relay Project and the Media Leadership Think Tank found in their study of 263 South African news sites that only 30.4% block at least one AI crawler, that leading sites lost about a fifth of daily page views in a year, and that chatbots referred under two million visits, less than 1% of 256 million chatbot visits recorded.
Machine consumption is rising towards a significant share of infrastructure load while contributing almost nothing in referral traffic, which means the cost of serving content is migrating away from the revenue it once produced. The South African figures show how this creates discrimination: blocking requires engineering capacity, so smaller publishers remain the most accessible and the least compensated. Michael Markovitz of the Media Leadership Think Tank describes robots.txt accurately as a yes-or-no gate rather than an enforcement tool.
This is why the SPUR standards work in signal 27.1 has commercial substance. A gate answers one question; a meter answers the questions that produce an invoice, namely which content, how often, for what purpose and at what value.
Reflections
What share of your infrastructure cost is now attributable to machine consumption, and who inside the organisation owns that number?
What answers does the standard meter has to come up with to mesure value creation?
27.12 — Verification becomes the visible product
Reuters has published an account of how its global visual verification team reviews hundreds of images daily and publishes perhaps a dozen, interviewing original photographers, examining metadata, comparing shadow direction with satellite, archive and street-view imagery, and scanning material with several detection tools that the agency states are imperfect. The Society of Professional Journalists is revising its Code of Ethics for the first time in twelve years, with a 967-word draft going to a vote in October that addresses AI, verification, fabricated images and creator journalists, and with vice chairman Chris Roberts stating that journalists remain ethically responsible for their work regardless of the tools used. The demand side of verification is documented in Meta’s disclosure that it disrupted an Iran-based network of 35 accounts using AI-generated posts followed by about 79,400 Instagram accounts, in NPR and NewsGuard’s finding that chatbots debunked false narratives about three-quarters of the time while AI summaries performed worse, and in Queensland University of Technology research showing that 12.5% of analysed AI Overviews on 15-minute cities presented false claims as one side of a legitimate debate.
The failures now reach the institutions that consume journalism. Guardian Australia found at least 39 parliamentary submissions containing apparently fabricated references, with more than 100 papers carrying ChatGPT URL tags, including misattributed work by named academics. Press Gazette established that four prolific technology and finance freelancers could not be verified as real people, leading Hacker Noon and Information Age to remove articles, in a pattern linked to promotional campaigns for cryptocurrency projects. Bernama’s chief executive used the Asia-Pacific Media Forum in Shenzhen to call for cross-border fact-checking and shared principles for responsible AI use, and a research consortium is surveying journalists in the United Kingdom, Brazil, Canada, the Netherlands and South Africa on the professional qualities that define the craft.
Verification has moved from an internal cost to an externally priced capability. Reuters publishes its method because the method is the product, and an agency that can demonstrate provenance holds an asset that scales with the volume of synthetic material in circulation. That is the strength of the quality newsroom as an institution: a repeatable procedure, documented and auditable, rather than the reputation of any individual byline.
Reflections
What would you charge for verification if it were sold as a service rather than absorbed as a cost of publication?
Which parts of your verification method can you publish without weakening it, and what does publishing them achieve?
Conclusion
The barbell hypothesis explains most of what this week’s signals contain, and it leaves one question open. If the market divides into boutique journalism and large efficient services, the position worth holding is the bar itself: the layer through which verified content, its rights and its usage records pass between producers and machines. Whoever specifies that layer determines how value is distributed across everything attached to it.
This is why the EANA decision to join a standards coalition belongs at the front of this edition. Agencies occupy that middle position by construction, since they sell structured rights and verified material into every other part of the system. CNA pricing machine access, AFP preserving editorial metadata inside a production system, Ritzau guaranteeing the origin of an official message and Reuters publishing its verification method are all versions of the same commercial claim: the scarce good is proven provenance, and it costs money to produce.
The legal picture supports that claim from two directions. Munich has established that storing recognisable protected works inside a model is reproduction under German and American analysis alike, while Brussels has designated a conversational service as a gateway with systemic-risk duties. Claims define the boundary, negotiations set the price, and a functioning market requires both. The American filings this week are a reminder that the boundary is contested rather than settled, and that European rights holders should build their position on copyright reasoning and on machine-readable reservations they can actually enforce.
For the operators, the practical agenda is narrow enough to act on. Know what proportion of your infrastructure serves machines and what that costs. Express your rights in a form a system can read, and your terms in a form an agent can transact against. Decide which capability in your organisation is genuinely scarce, and price it before someone else standardises it for you.
Markus Knall’s closing observation deserves the last word, because it points past the tooling to the demand side. If every process and every piece of content becomes AI-optimised, the remaining reason for a reader to choose your journalism is the confidence that what you publish is true and that something new has been discovered. That confidence is the one asset in this market that cannot be generated, only earned and then documented.