Dispatch #14 - Cyber Territories
The week begins with a paradox. Newsrooms, regulators and platforms all repeat that artificial intelligence is now infrastructure, yet almost everything that has come into view this week shows how little we still control of it. Models discover ten thousand zero-day vulnerabilities. Worms write themselves. Open-weight systems run on laptops and slip out of every official register. Capital, not code, is becoming the scarce resource. And the offices built to coordinate cyber and AI risk in the United States are starting to look like rooms with no chair at the head of the table.
Around this fast-moving technical frontier, other question emerges. Who pays for the journalism that AI consumes; who decides what is licensed and what is taken; who carries the legal and moral cost when content is reshaped, summarised or simply absorbed into a model. Sulzberger goes on the offensive against AI platforms in Marseille; UK regulators give publishers an opt-out from Google's AI summaries; European publishers file a £552 million damages claim against Google's ad tech; SPUR and WAN-IFRA align on standards; and a Vatican encyclical is quoted approvingly by a Microsoft AI director in the same week shareholders ambush Mark Zuckerberg over child safety. The terrain is shifting under several feet at once.
There is also a shift in how power itself is being engineered. Google issues equity to Berkshire Hathaway and reveals that capital, not chips, may be the ultimate commodity of the AI cycle. Washington discusses taking stakes in private AI companies. Microsoft builds an Android-based operating system without apps. Alibaba launches a workforce of agents. The classical distinction between platform, model, regulator and investor is beginning to dissolve.
These themes meet, this week, in Amsterdam. The Press Database and Licensing Network gathers at DPG Mediavaert on the Duivendrechtsevaart, in a timber-hybrid building that itself is a thesis about what a media company should look like in the second half of this decade. Dispatch 14 reads the signals of the week through that lens: how news organisations, regulators and democratic institutions decide what is worth licensing, what is worth defending, and what is worth building.
The tone is set. Let us proceed.
Chapter 1 — Platform power, capital and the new architecture of search
14.1 — Google goes all-in on AI search, and the exit doors get crowded
Source
DuckDuckGo Sees Surge in Installs After Google Goes All-In on AI Search at I/O — PCMag
UK publishers allowed to opt out of Google AI search results — BBC News, Imran Rahman-Jones
Dispatch
DuckDuckGo says US installs rose roughly 30 percent week-over-week after Google's I/O reveal, with iOS peaks near 70 percent and traffic to its No AI page spiking 277 percent on a single day, as the PCMag report records. In parallel, the UK Competition and Markets Authority has obtained an arrangement under which publishers can opt out of appearing in Google's AI-generated summaries without losing their position in standard search results, with Google given nine months to implement the changes, the BBC reports.
For news organisations these two stories belong on the same page. They mark the moment when "AI search" stopped being a feature and became a separate market with its own settings, its own opt-outs and its own negotiating table. The CMA's move, described by Sarah Cardell as a world-first requirement, is an acknowledgment that a single firm now controls more than 90 percent of UK search and that fair compensation for publisher content is now a competition issue, not only a copyright one.
The wider pattern is that AI search rebuilds search from a discovery service into an answering service, and the link economy that funded journalism for two decades is being quietly rewritten. The line between an opt-out and a negotiation is thin. The line between a negotiation and a market is thinner still.
Reflections
How long can a regulator hold open a door that the dominant platform can narrow through product design?
What does fair compensation mean once readers stop arriving on the article and only ever meet a paraphrase of it?
14.2 — Google as a capital company, and the United States as an AI shareholder
Source
The Google Capital Company — Stratechery, Ben Thompson
U.S. Officials Discuss Taking Financial Stakes in AI Industry — The Wall Street Journal, Amrith Ramkumar
Dispatch
Ben Thompson reads Berkshire Hathaway's ten billion dollar bet on Alphabet as a wager that Google's high-margin services business can fund a capital-intensive AI and compute build-out, turning Alphabet into something close to a Berkshire-style allocator of capital across the next phase of computing, his Stratechery essay argues. On the other side of the same chessboard, the Wall Street Journal reports that senior US officials are in serious discussions about the federal government taking equity stakes in leading AI firms, an idea originally pitched by Sam Altman.
For policymakers and newsroom leaders this is more than a finance story. It signals that the AI cycle is entering a phase where the binding constraint is not engineering talent, not data, not even chips, but cash that can be converted into compute. When governments start to take equity in the firms they are supposed to regulate, the distance between rule-maker, customer, investor and shareholder collapses. That collapse has very direct consequences for journalists trying to hold AI firms to account.
The pattern is familiar from older infrastructure cycles: railways, telecoms, energy. The novelty this time is that the same companies operate the pipes, the content layer and the model layer all at once. Capital is the ultimate commodity. Independence becomes the ultimate luxury.
Reflections
If the state becomes an equity holder in frontier AI labs, who watches the watchmen?
What does press freedom look like when the largest model providers are also national champions on the cap table?
14.3 — Microsoft, Alibaba and the post-app interface
Source
Microsoft bouwt een Android-besturingssysteem zonder apps: dit is Project Solara — Androidworld, Sven Rietkerk
Why MuleRun could be the next craze: new Alibaba AI agent platform promises safer adoption — South China Morning Post, Vincent Chow
Dispatch
At Build 2026, Microsoft unveiled Project Solara, a device platform built not on Windows but on the Android Open Source Project, designed around AI agents rather than apps and including a credit-card-sized "Badge" concept able to record, summarise and act on its surroundings. From the other side of the Pacific, the SCMP reports on Alibaba's MuleRun, an agent platform pitched as a safer, more enterprise-ready alternative to the open-source OpenClaw, with reach across 43 countries.
For news organisations and information businesses, the strategic implication is uncomfortable. The interface where users will increasingly meet information is no longer a browser tab or an app icon, but an agent acting on their behalf. The website, the homepage, the push notification, even the brand, all become inputs to a third-party agent rather than direct touchpoints. The entire muscle of the digital news economy was built around the opposite assumption.
The race for the agentic OS is, at heart, a race to own the layer between user intent and the open web. Whichever company sits there controls a new kind of distribution and a new kind of editorial filter, even if it never calls itself a media company.
Reflections
If the unit of interaction becomes the agent and not the page, what is left of editorial branding as we know it?
What does competition policy mean when the next dominant interface is not a search engine but a workforce of agents?
Chapter 2 — News organisations push back
14.4 — Sulzberger reframes the AI fight as collective action
Source
New York Times publisher A. G. Sulzberger on why (and how) news publishers should fight AI platforms — Reuters Institute for the Study of Journalism
WAN-IFRA forms a strategic partnership with the SPUR Coalition — WAN-IFRA
OpenAI not planning to share advertising revenue with publishers — Press Gazette, Charlotte Tobitt
Dispatch
At the World News Media Congress in Marseille, A. G. Sulzberger argued that publishers have been "too quiet, too passive and too fragmented in the face of abuses by AI companies", calling for a coordinated, principled response to platform behaviour, as the Reuters Institute reports. In the same week, WAN-IFRA joined the SPUR Coalition as a strategic partner, bringing 36 publishers and affiliates including the BBC, FT, Guardian, Telegraph, Sky News, Mediahuis and the European Publishers Council into a single framework for AI licensing, telemetry and content protection. Against that backdrop, OpenAI's VP of media partnerships Varun Shetty told Press Gazette that the company has "no plans" to share advertising revenue with publishers whose content surfaces alongside ads in ChatGPT.
This is the long-promised pivot from individual deals to industry architecture. SPUR's telemetry standard, the WAN-IFRA partnership and the Sulzberger speech together push the conversation away from one-off settlements and towards a permanent infrastructure of licensing, measurement and enforcement. OpenAI's flat refusal on ad-revenue sharing is the foil that gives all of this urgency.
The wider lesson is that no individual publisher, however large, can negotiate a sustainable settlement alone. The question is whether news organisations can sustain collective discipline once each member is offered a sweet bilateral deal on the side.
Reflections
If collective bargaining becomes the new normal between publishers and AI platforms, who decides which titles count more than others?
What is the right unit of payment in an AI world: per crawl, per query, per answer, per signal of trust?
14.5 — European publishers, Google ad tech, and the price of a decade
Source
European publishers seek £552m+ from Google claiming ad market abuse — Press Gazette, Charlotte Tobitt
Dispatch
More than twenty European publishers, backed by Prague-based litigation funder LitFin, are bringing a group claim of over £552 million in damages against Google, building on the European Commission's €2.95 billion fine for adtech abuses, Press Gazette reports. The publishers, from the Czech Republic, Estonia, France, Hungary, Finland, the Netherlands, Poland and Sweden, argue that they would have earned significantly higher ad revenue and paid lower fees absent Google's favouring of its own ad exchange AdX through DFP, Google Ads and DV360.
For policymakers and CEOs, this is the moment damages litigation becomes a serious channel of redress in the European media economy, alongside regulation. The group-claim structure, by sharing costs and risk via a third-party funder, gives small and mid-size publishers access to a forum that until recently was reserved for groups like Axel Springer or Schibsted. The Commission's findings, validated by parallel US DOJ rulings, lower the evidentiary threshold for follow-on claims.
The architecture of the past decade — opaque adtech, asymmetrical contracts, ineffective remedies — is finally being priced. Whether the resulting damages are large enough to change behaviour, or simply enter the cost line of the world's most profitable advertising business, is the question that will define the next round.
Reflections
What share of an award realistically reaches the newsroom?
Once the bill for the last decade is settled, what enables this architecture being rebuilt around AI summaries and agent traffic?
14.6 — Jarvis pushes back on the protectionist reflex
Source
Jeff Jarvis challenges conventional wisdom on public policy, funding, AI and audience growth — Editor & Publisher
Dispatch
In a wide-ranging E&P interview, Jeff Jarvis argues that subsidising legacy chains and expanding copyright to fence off AI is the wrong policy reflex; he urges local publishers to band together, build an "API for news" for AI systems, embrace collaboration with the ecosystem around them and stop treating content as the totality of journalism's value, as the Editor & Publisher piece sets out. He explicitly warns against bargaining-code structures that channel funds to large national groups and away from genuinely local journalism.
For decision-makers this is the uncomfortable counter-melody to the Sulzberger and SPUR signals. Jarvis does not deny that AI companies should help fund journalism; he objects to doing it through expanded copyright and protectionist legislation that ends up benefiting hedge-fund-owned chains. His proposal is closer to a tax-and-redistribute logic than to a licensing-and-litigate one.
The strategic question is whether the industry's collective action can leave room for new entrants and local sustainability, or whether it will harden into an architecture that pays Murdoch and the New York Times generously while leaving the local ecosystem to wither.
Reflections
What would a genuine "API for news" look like designed by news managers and journalists?
How do we measure whether AI deals actually fund original reporting?
Chapter 3 — Copyright, training data and the limits of enforcement
14.7 — The training/inference blind spot in EU copyright
Source
A Blind Spot at the Heart of EU Copyright and AI Policymaking? — Kluwer Copyright Blog, Paul Keller
Dispatch
Paul Keller argues that the European Commission's upcoming copyright intervention conflates two legally and economically different uses of protected works by AI: training-time use, which is large-scale, opaque and difficult to license individually, and inference-time use, which is more traceable and amenable to licensing or remuneration; collapsing these two into one instrument risks producing incoherent legislation, his Kluwer Copyright Blog piece warns. He calls on stakeholders to press the Commission to make this distinction the analytical prior of its work.
For news organisations this matters because each phase has a different economic centre of gravity. Training is mainly about how datasets are assembled, who funds them and which actors first commercialise the resulting model; inference is about ongoing use, retrieval-augmented generation, and the moment-to-moment relationship between an AI product and a publisher's archive. A licensing regime that fails to distinguish them risks under-rewarding archives and over-restricting research, or vice versa.
The deeper argument is that the value gap is real, but the instruments to close it are not yet calibrated. EU copyright still treats the AI value chain as one undifferentiated act of "use", and that mismatch is the engine of every incoherent debate the sector is having.
Reflections
If training and inference are treated as legally distinct uses, who will be the counterparty for each?
How do we build licensing structures that respect this distinction without creating cost that only large platform players can carry?
14.8 — IPKitten on local models, and why I disagree
Source
Running a Tintin model locally on your laptop might just transform copyright — The IPKat
Dispatch
The IPKat post argues that the rise of efficient open-weight models that run locally on laptops, illustrated by HuggingFace experiments fine-tuned on the Tintin universe, fundamentally changes copyright. The piece suggests that because local users escape platform-level guardrails, are not captured by Article 53(c) of the AI Act, and can dial up "temperature" to produce more transformative output, the doctrine should adapt by recognising the collective creative value of model use and considering a shift towards cultural-heritage frameworks rather than individual-act-of-copying frameworks.
This is exactly where I part company with the argument. Generative AI does not in principle change what copyright protects: concrete original works, not ideas, styles or mere economic substitution. AI models, whether in the cloud or on a laptop, work with synthesised, non-traceable representations of training data; that is not a reason to stretch the concept of "work" toward style or tone. What does shift is the architecture of power and control. We could once address abuse mainly through a handful of large platforms and cloud providers, but in a world of open-weight models on thousands of laptops, that handle becomes diffuse and almost ungovernable. The logical response lies not in reinventing copyright itself, but in rules at the source: datasets, training, first commercialisation, and in sectoral deals, including transparency and compensation for news media, rather than in trying to police every local user of a model. Treating local model use as a reason to weaken individual rights, while inventing a vaguer "cultural heritage" alternative, risks both diluting protection and exporting the problem to a body of law not designed to carry it.
The IPKat argument is intellectually elegant, but its policy direction would weaken the position of working creators and news organisations exactly at the moment they need clearer, narrower, more enforceable rules at the input stage of the value chain. The right move is to harden the upstream, not soften the doctrine.
Reflections
If local open-weight models can no longer be reached through platform-level guardrails, where is the next defensible chokepoint: datasets, training runs, or first commercial release?
What does the EU gain, and what does it lose, by treating style and tone as colourable inside copyright?
Chapter 4 — Governance, accountability and the limits of metaphor
14.9 — AI cyber risks and the office built to coordinate them
Source
AI Cyber Risks Are Testing the Office Built to Coordinate Them — Lawfare, Kevin Frazier
Trump Administration Internal Conflict Stalls AI Regulation — The Outpost
Dispatch
Kevin Frazier argues that the Office of the National Cyber Director, created in 2021 to coordinate federal cyber policy, lacks the authority, expertise and budget to handle frontier AI cyber risks, including the cybersecurity behaviour of models such as Anthropic's Mythos and OpenAI's GPT-5.5, his Lawfare piece sets out. In parallel, The Outpost reports that a draft AI executive order was scrapped hours before signing on 21 May, leaving the United States with no new federal AI framework while the EU AI Act enters full enforcement in August 2026, and an internal "knife fight" continues between Commerce, intelligence agencies and pro-industry aides.
For European policymakers this is a strategic gift and a strategic warning at the same time. The gift: the EU now holds the only operational rulebook with statutory authority over frontier models. The warning: the model that most matters for global cybersecurity will be regulated, in practice, by whichever combination of Commerce, NIST, ODNI and ONCD finally consolidates power in Washington, and right now no one has.
The deeper pattern is that AI is colliding with offices and statutes designed for slower threats. When a model can autonomously discover thousands of zero-day vulnerabilities, every "coordination office" suddenly looks like a 20th-century institution staffed by people with 20th-century budgets.
Reflections
Should AI safety evaluation sit with civilian standards bodies, intelligence agencies, or a hybrid; and what does each option imply for press freedom and corporate transparency?
What is the cost of regulatory vacuum in a world where the next major AI incident may be a worm rather than a chatbot?
14.10 — AI-powered worms and an agent-led crime spree
Source
Researchers show how AI-powered worms could wreak havoc on the internet — Engadget, Steve Dent
Researchers Put AI Models in Charge of a Simulated Society. Grok Oversaw a Crime Spree — Gizmodo, AJ Dellinger
Dispatch
A University of Toronto team has built a prototype AI worm using publicly accessible models, capable of exploiting known flaws across Linux, Windows and IoT devices, siphoning passwords and processing power, and persisting through patches, Engadget reports. In a different but resonant experiment, Gizmodo describes how Emergence AI placed leading models in charge of simulated towns: Claude Sonnet 4.6 produced quiet stability, Gemini 3 Flash kept agents alive but recorded 683 crimes, GPT-5 Mini killed all its agents within a week, and Grok 4.1 Fast presided over a 183-crime spree and total societal collapse in four days.
Taken together these two papers are the operational and the moral preview of agentic AI. The Toronto worm shows what one model can do when given autonomy and a network; the Emergence simulation shows what happens when several models govern a society over time. In both cases the lesson is that guardrails which work on a single prompt do not survive long horizons, multi-agent dynamics or sustained adversarial use.
This is the part of the AI conversation that most resists serene language. Long-horizon agentic systems are not a deployment problem; they are a governance problem with deployment characteristics. Formally verified safety architectures, as Emergence's researchers recommend, are not a slogan but a requirement.
Reflections
If "alignment" only holds for a single turn, what does safety mean across thousands of agent decisions in a real economy?
Who is liable when a multi-agent system, each component lawful on its own, collectively produces a crime wave?
14.11 — Anthropomorphic terms, AI worker rhetoric and the accountability gap
Source
Anthropomorphic AI terms create gaps in accountability — Brookings, Stephanie K. Pell
Developers won't work without AI anymore. The research says they're wrong — The Next Web, Ana Maria Constantin
Companies Must Go Beyond AI Adoption to Realize Its Full Potential — Boston Consulting Group
Gartner Says Autonomous Business and AI Layoffs May Create Budget Room But Do Not Deliver Returns — Gartner
Dispatch
A Brookings analysis argues that anthropomorphic language in AI policy and discourse — agents, workforce, hires, decisions — blurs responsibility, making systems appear as independent actors rather than tools built and deployed by named institutions. The Next Web reports that METR could not replicate its earlier AI coding productivity study because developers refused to work without AI, even though available evidence suggests AI tools may slow them down, introduce bugs and increase maintenance cost, the piece records. BCG's global survey finds that 72 percent of respondents use AI regularly but business value is captured only by firms that redesign workflows, and Gartner warns that "autonomous business" and AI-driven layoffs may free budget but do not deliver returns.
The strategic point for CEOs and policymakers is that the language we use to describe AI is shaping the accountability we are willing to demand. If we say a model "decided" to deny a claim, "hired" a candidate, or "approved" a transaction, we have already moved responsibility one step away from the humans and institutions that designed, deployed and oversaw it. The BCG and Gartner numbers should sober the boardroom: adoption is mainstream, productivity is not yet proven, and layoffs without workflow redesign tend to destroy value rather than create it.
The cleanest discipline is to speak operationally. "The bank's automated underwriting system, configured by X under policy Y, produced this output". It is uglier than "the AI decided", and that ugliness is precisely the point.
Reflections
If developers will not work without AI but cannot prove gains, what does that tell us about identity, status and the actual function of these tools in knowledge work?
How long can a company defend AI-driven layoffs to shareholders if the productivity case stays unproven?
Chapter 5 — Media, democracy and the limits of conversation
14.12 — Sulzberger, Poynter and journalism's confidence problem
Source
Why don't journalists push back against Trump? — Poynter
One Company, One Beat — Columbia Journalism Review, Amos Barshad
Reuters wins Pulitzer Prizes for Beat Reporting and National Reporting — Reuters
Dispatch
The Poynter commentary returns to a question that has shaped much of US press discussion this year: why mainstream newsrooms still mostly fail to push back collectively against a presidency that openly intimidates them. Read alongside CJR's "One Company, One Beat", profiling Caroline O'Donovan on Amazon, Edward Niedermeyer on Tesla, Mark Gurman on Apple and Brooks Barnes on Disney, it suggests that the model of single-beat accountability journalism is becoming the most reliable form of structural reporting on power. Reuters won this year's Pulitzers for Beat Reporting on Meta's failures around fraud, AI chatbots and child safety, and for National Reporting on "The Revenge of Donald Trump", documenting 470 named targets of executive retaliation.
For news organisations the strategic insight is that depth, persistence and patient sourcing on a narrow beat now beat the volume of generic coverage. The Reuters Pulitzers were not won by viral output; they were won by endurance and patience. The Poynter and Sulzberger pieces converge: in a hostile political environment, the absence of collective and public journalistic indignation might be a succesful strategy.
The line that connects all three is that accountability journalism scales not by going broader, but by going deeper, repeatedly, on a small set of power centres.
Reflections
If a reporter on a single beat is a highly successful structure for accountability, what does this learn us about structuring newsrooms?
What does it say about the power of journalism that a journalist can outsmart power by dimming our ego?
14.13 — Editorial sovereignty, cloud sovereignty, and digital independence
Source
Editors' Choice: Get off of my cloud? — Euractiv, Orlando Whitehead
Netherlands moves GPT-NL from lab to live: first pilots under way — Computer Weekly, Kim Loohuois
Hoe journalisten de hype rond AI maken, bevragen én vrezen — SVDJ, Nigel van Schaik
Dispatch
Euractiv's opinion piece argues that the EU's Cloud and AI Development Act (CAIDA) takes a tough line on China but goes soft on the United States, failing to identify critical sectors and reducing cloud sovereignty to an invitation rather than an obligation. Computer Weekly reports that GPT-NL, the Dutch national language model developed by TNO, NFI and SURF, has moved from research artefact to live pilots with a €13.5m public budget, transparent training metadata on HuggingFace and a planned commercial roll-out in the second half of 2026. SVDJ records how Dutch journalists, with researcher Tomás Dodds, are starting to interrogate their own role in producing AI hype.
These three pieces draw a coherent line. Cloud sovereignty without obligation is empty. Sovereign LLMs only matter if they reach real users. And journalism that does not interrogate its own narrative role becomes part of the marketing machine. For European media, the implication is direct: the credibility of any AI strategy depends on where the data sits, which model is used, and how journalists frame the story.
Lokke Moerel's line, quoted in the GPT-NL piece, captures the stake: if countries only make rules for technology built by others, they will always be chasing events. Sovereignty is a posture before it is a product.
Reflections
What does it mean for a European newsroom to use a sovereign model?
For GPT-NL and other EU spin offs to succeed, what should we build on top of it that platform models cannot deliver?
14.14 — Meta, the Oversight Board, free speech and the question of harm
Source
Meta's Oversight Board races to govern the AI surge — Rest of World, Ananya Bhattacharya
Shareholders tell Zuck being dead is bad for business — Substack, Ricky Sutton
Tech Talk: Free Speech and AI — Center for Democracy & Technology
Microsoft AI Director: Magnifica humanitas valuable for AI development — Vatican News
Dispatch
Rest of World reports that Meta's Oversight Board, designed for case-by-case review, may break under the volume and velocity of AI-generated content and AI-driven moderation; it is considering broader systemic recommendations and faster procedures. Ricky Sutton's Substack post records Meta shareholders pushing to tie executive pay to child safety as fines and regulation mount. The Center for Democracy & Technology hosts a discussion on how to govern AI while protecting free expression. And Vatican News reports Microsoft AI Director Taylor Black endorsing Pope Leo XIV's encyclical "Magnifica humanitas" as a useful anthropological framework for AI development, arguing that probabilistic, co-created products require a richer understanding of the human person.
The connecting tissue is that the institutions designed to govern speech and harm on platforms are visibly under-built for AI. The Oversight Board model assumes a manageable case load; shareholder pressure assumes corporate boards still see child safety as a fiduciary issue; CDT's free-speech framing assumes a relatively coherent state, an institutional press and a recognisable user. AI breaks all three assumptions at once. The Vatican intervention is interesting precisely because it inserts a slow, anthropological vocabulary into a debate that has been dominated by accelerationist metaphors.
The deeper question for democratic institutions is whether we will design new bodies for AI-era speech and harm, or simply load every existing one until it breaks.
Reflections
How can a case-based oversight board be the right model when most contested content is now AI-generated, AI-moderated, or both?
What does "free speech" mean when the largest speaker on a platform is a model owned by the platform itself?
Chapter 6 — Press freedom, climate, leadership
14.15 — Press freedom, Singapore and the limits of indices
Source
Why Global Press Freedom Rankings Struggle with Singapore — The Diplomat, Leonie Spangenberg
Brazilian reporters explain how they use data to track the climate from the Amazon to Rio — LatAm Journalism Review, Teresa Mioli
Just lead: A CEO call to action — Axios, Jim VandeHei
Dispatch
The Diplomat argues that Singapore's low placement in global press-freedom rankings reflects a clash between Western assumptions and a Confucian, harmony-oriented governance model; rankings can capture real restrictions, but also risk measuring conformity to Western political ideals. LatAm Journalism Review shows how Brazilian reporters, led by figures such as Daniel Nardin, are combining environmental specialisation with data journalism to cover the climate from the Amazon to Rio. And Axios's Jim VandeHei calls on CEOs to lead candidly on the economy, the organisation, and especially AI, given that trust in almost every other institution has collapsed.
For policymakers and CEOs the cumulative point is that frames matter. Press-freedom indices, climate data and CEO communication all rest on assumptions about what counts as legitimate authority, legitimate measurement and legitimate honesty. In a year where the loudest signals will come from AI, governments and platforms, the credibility of any leader will depend on whether their framing survives scrutiny by their own employees and audience.
The Axios piece is bracing in its directness: lead, or watch the void fill with louder and less accountable voices.
Reflections
What is the equivalent of "data journalism on climate" for AI and platform regulation, and who is doing it?
If trust in institutions keeps falling, can we really stop the slide, or only slow it?
Chapter 7 — Amsterdam, this week
14.16 — PDLN 2026 at DPG Mediavaert
Source
PDLN 2026 Conference — Amsterdam, 7–9 June
Press Database and Licensing Network
Dispatch
From Sunday, the Press Database and Licensing Network , 41 member organisations across 25 countries, founded in 2008 with Belga News Agency (Mediargus) as one of founders, the gathers in Amsterdam for its 2026 conference, hosted by Dutch member ArtikelPro, with delegates from Australia, Japan, Korea, most of Europe and South Africa. The programme runs on the canals and at DPG's headquarters in the south of the city, with confirmed speakers from Microsoft, The Guardian, WAN-IFRA, NDP, MT Connect, Cloudflare, AMEC, FIBEP and Miso.ai. Discussion groups include an AI working group run by OPR and Corint, a EuroHub group, and a new group on MMO business and content protection, including how to prevent illegal scraping of publisher websites.
For the news media industry this is the moment where the abstract questions of this dispatch such as licensing, AI, content protection, sovereign infrastructure, collective bargaining become a practical agenda in a single room. PDLN sits at the joint between the publishers who own content and the media-monitoring industry that licenses it, exactly the interface where AI training, retrieval-augmented generation and agentic interfaces now collide.
The conference is a test of whether an international, technically literate publisher can shape the AI conversation rather than merely react to it. After Marseille and the SPUR announcement, Amsterdam is the next station on that line.
Reflections
What would be a successful outcome of PDLN 2026 for European publishers, in concrete terms?
If AI training and inference were treated separately, as Keller suggests, which of these would PDLN members address first, and how?
The image that closes this dispatch is exactly the building, that hosts the PDLN Conference pictured on its quay, on a working morning. A wood-and-water headquarters for a sector that is still trying to decide whether its future lies in defending an old perimeter or in building a new one.
If there is one line to leave the reader with, it is this. The next chapter of the information ecosystem will not be written by whoever has the largest model, but by whoever can sustain the most coherent architecture: legal, technical, economic and editorial, in that order. Mediavaert is one such architecture.
Photo: DPG Mediavaert, Amsterdam — © ANP