Dispatch #19 - Cyber Territories

The open web took shape on a simple bargain. Publishers allowed machines to read their pages, and in exchange those machines sent readers back. This week the terms of that bargain came up for renewal across three continents, and the people who create information began preparing for the possibility that the exchange no longer holds.

When a crawler takes thousands of pages for every visitor it returns, the arrangement stops being reciprocal and becomes extraction. The links this week describe the question asked in many jurisdictions and many registers.

Who pays for the knowledge that machines now consume, and what happens to the commons when the people who fund it decide to close the gate? Regulators in Paris, Copenhagen, Abuja and Dublin are testing legal answers. Publishers in the United States and the United Kingdom are testing technical ones. Researchers in Oxford are measuring how the same systems reshape opinion while the argument over payment continues.

These threads converge on one structural choice. If the automated reading of the web can no longer be assumed, the architecture of access itself becomes the battleground. Publishers are now willing to build a wall between human readers, who remain welcome, and machine readers, who must ask permission, accept limits and pay. This dispatch follows the themes of licensing, trust, governance and platform power towards that wall, and closes on the image it produces.

Chapter 1 — Licensing and the Search for Equilibrium

19.1 — Getty shows that litigation and licensing are one process
Source
• Getty Images Fought AI In The Courts, Lost, Then Lucratively Licensed. The New Publishing Standard.

Dispatch
Getty Images sued Stability AI in the United Kingdom in 2023, then saw most of its claims fall away, with the High Court rejecting the secondary copyright argument in November 2025 on the ground that model weights did not themselves store copies of its images. Rather than treat that ruling as an ending, Getty separated two distinct products, the right to train a model and the right to display an image, and licensed only the second. It signed a display agreement with Perplexity in October 2025 and a comparable one with OpenAI in June 2026, both conditioned on attribution and linking, and neither granting training rights.

For decision-makers the lesson concerns product design rather than victory in court. Training rights and display rights carry different risks and command different prices, and a publisher can decline the first while selling the second. Getty attached attribution as a contractual term, which turns provenance into a paid feature rather than a favour. The market read this as value creation, with reported share-price increases ranging from roughly ninety to over two hundred per cent across outlets, against a Q1 2026 net loss of 4.4 million dollars, down sharply from a year earlier.
This is what equilibrium looks like when it forms. Claims in court establish the legal boundaries, and negotiated licences establish workable prices, and from that combination a sustainable business model emerges. The News Corp arrangement with OpenAI, reported above 250 million dollars over five years, and comparable deals at Axel Springer, Le Monde and Dotdash Meredith, suggest that the firms doing the rights and metadata work early are negotiating from strength.

Reflections
• How should a news organisation price display access separately from training access without undermining either?
• What does a firm need to prepare in its archive and metadata before it can negotiate from strength rather than from grievance?

19.2 — Denmark and France put a European price on press content
Source
• Denmark intervenes in EU court case over publishers' rights. Reuters.
• French antitrust watchdog orders Meta to resume talks with media groups over publishing fees. The Star / Reuters.
• French watchdog orders Meta back to press payment talks after copyright deals expire. Euronews.

Dispatch
Two European institutions moved in the same week to give press publishers' rights concrete force. Denmark filed a written intervention in the European Court of Justice Streamz case, supporting Belgium against Google, Meta, Spotify, Streamz and Sony, and arguing that a ruling for the platforms would dilute the press publishers' right under Article 15 of the CDSM directive.

In Paris the competition authority found that Meta had likely abused its dominant position by imposing its own method for calculating fees while withholding the information publishers needed to assess them, and ordered Meta to present a payment plan within fifteen days.

For strategists the significance lies in where the pressure originates. The French decision rests on competition law rather than copyright alone, which means the remedy addresses conduct, the refusal to share information and the imposition of terms, rather than the abstract question of whether content was used. The author says the authority deliberately declined to name a provisional sum, to avoid creating a focal point for the negotiation, which leaves the price to the parties while the state enforces the duty to bargain in good faith.

The pattern across Copenhagen and Paris is a European insistence that content
19.4 — Newsrooms write the rules for AI they can defend
Source
• ABC unveils new AI policies. ABC (Australia).
• A new study looks at the skills journalists are losing and gaining because of AI tools. Nieman Lab.

Dispatch
Australia's public broadcaster set out formal rules for the use of AI in its journalism, joining a growing group of newsrooms that treat internal governance as a condition of public trust rather than an afterthought. Alongside this, new research summarised by Nieman Lab examines which skills journalists gain and which they lose as they adopt AI tools, from faster research and transcription to a measurable risk of atrophy in verification and source work.

For newsroom leaders the two items describe a single management problem. A public AI policy protects the institution externally, and a skills strategy protects it internally, since a newsroom that automates its craft without maintaining its judgement erodes the very expertise that makes its output worth licensing. The commercial case for quality journalism and the training case for journalists are the same case.
The broader pattern connects standards to scarcity. Verified, accountable reporting commands a price precisely because it is costly to produce and hard to replicate, and a newsroom that lets its verification skills weaken lowers the value of what it sells. Governance and craft together defend the asset.

Reflections
• Which editorial skills must a newsroom actively preserve as AI absorbs routine tasks, and how would a leader measure their decline?
• How does a published AI policy change the trust relationship with audiences, funders and licensing partners?

19.5 — Authenticity inverts when synthetic feels more real
Source
• AI footage of politicians seen as 'more authentic' than the real thing. The Star.
• Lily Jay faces claims of AI-generated charity videos. Let's Data Science.
• Bombay High Court grants interim relief to Preity Zinta against deepfakes. NDTV.

Dispatch
Three items describe the same disturbance from different angles. Research reported by The Star found that audiences sometimes judge AI-generated footage of politicians to be more authentic than genuine recordings, a charity faced allegations that its emotional appeal videos were synthetic, and the Bombay High Court granted the actor Preity Zinta interim relief against deepfakes using her likeness. Perception, fundraising and personality rights are each affected.

For decision-makers the finding on perceived authenticity carries the sharpest warning. When a synthetic image satisfies an audience's expectation of what truth should look like more fully than reality does, the traditional defence of showing the real recording weakens. Courts are responding with likeness protection, as in the Zinta ruling, and charities and campaigns face a reputational exposure that arrives faster than verification can.
The pattern points back to provenance. Authenticity can no longer be assumed from the look of an image, so it must be established through attribution, chain of custody and verifiable sourcing. This is the same asset that Getty priced in the first chapter, trust made legible, and it explains why display deals condition payment on credit and linking.

Reflections
• When audiences find synthetic footage more convincing than real footage, how does a newsroom prove that authentic material is authentic?
• What responsibility does a charity or campaign carry to disclose AI-generated emotional appeals to its donors?

19.6 — Detection struggles even at the company that built the image
Source
• Meta AI image detector fails to identify some of its own cropped AI images. Reuters.
• How one fake photo let Russian propaganda cast doubt on the Kyiv Lavra strike. Kyiv Independent.
• AI fakes about the Lavra, fake Euronews reports and manipulations about Crimea. Detector Media.

Reuters reported that Meta's own AI image detector failed to flag some AI images it had generated once they were cropped, which shows how easily provenance signals break under ordinary editing. In Ukraine, a single fabricated photograph allowed Russian propaganda to cast doubt on the Lavra strike, and Detector Media catalogued a wider series of AI fakes and impersonated Euronews reports over one week in June. Fabrication and doubt operate together.

For editors and platform operators the failure of first-party detection matters because so much verification policy assumes that machine-generated content can be reliably labelled. When a detector misses cropped images from its own model, watermarking and automated flags become one input among several rather than a settled solution. According to the Ukrainian reporting, a Meta spokesperson attributed erroneous fact-check labels on posts about the Lavra attack to a technical fault, which itself became material for the disinformation campaign.

The broader pattern is that doubt is now a weapon in its own right. A propagandist does not need every fake to be believed, only enough uncertainty that audiences distrust genuine reporting. That raises the value of institutions whose verification can be traced and defended, and it explains why provenance is becoming a paid and contractual feature rather than a technical afterthought.

Reflections
• When a company's own detector cannot reliably identify its own edited outputs, how much weight should regulators place on automated labelling?
• How does a newsroom counter a disinformation strategy whose goal is doubt rather than belief?

Chapter 3 — Platform Power, Governance and Public Discourse

19.7 — AI-mediated communication can steer opinion at scale
Source
• AI-powered social media can subtly manipulate opinion at scale, new study finds. Oxford Internet Institute.
• Misinformation inciting harm to refugees, UN. The Straits Times.

Dispatch
Researchers at the Oxford Internet Institute and the Hasso Plattner Institute found that large language models systematically shifted the position expressed in social media posts on contested topics, even when instructed to preserve the original meaning, and that these small changes could accumulate across millions of interactions to influence public opinion. In one test of X's "Explain this post" feature, the imbalance traced back to a single instruction telling the model to challenge mainstream narratives. Separately, the UN warned that misinformation is inciting real-world harm against refugees.

For policymakers the important finding concerns implementation rather than the models alone. The direction and size of the effect depended on choices made by the platform, which means accountability attaches to product decisions that are documented and auditable. Senior author Sandra Wachter noted that AI-mediated communication is a subtle form of influence that the law has yet to address, and the study observes that neither the EU AI Act nor the Digital Services Act directly covers it.

The pattern links quiet editorial nudges to visible harm. When a rewriting feature tilts contested debate and when misinformation drives violence against vulnerable groups, the governance question moves from content moderation to the design of the tools that shape expression. The largest platforms carry a corresponding responsibility to make those design choices transparent.

Reflections
• When a single platform instruction can tilt opinion on contested topics, what disclosure should regulators require about such prompts?
• Where does responsibility rest when harm follows from a platform's implementation choices rather than from the underlying model?

19.8 — Platform gatekeeping reaches creators and definitions
Source
• YouTube warns UK creators over government's new content control rules. The News.
• The Guardian on Meta's algorithm and digital publishers such as LadBible. The Guardian.
• Committee to Protect Journalists in turmoil after 17-1 vote on definition of journalist. Washington Free Beacon.

Dispatch
Three items show gatekeeping expressed through rules, algorithms and definitions. YouTube warned United Kingdom creators about the effect of new government content-control rules on their reach, The Guardian reported how changes to Meta's algorithm reshaped the fortunes of digital publishers such as LadBible, and the Committee to Protect Journalists faced internal conflict after a 17 to 1 vote over how to define a journalist for its Gaza casualty list. Distribution, reach and professional identity are each subject to a gate.

For strategists the common thread is dependence on a definition or a setting controlled by someone else. A publisher whose audience depends on an algorithm can lose reach without notice, a creator whose livelihood depends on platform rules is exposed to policy shifts, and a profession whose boundaries are contested struggles to defend its members and its standards. Control over the criteria is control over the outcome.

The pattern connects platform power to institutional authority. When the largest gateways set the terms of visibility and when professional bodies divide over who counts as a journalist, the infrastructure of public information becomes a matter of governance rather than of neutral plumbing. The organisations that own the gate carry responsibility for how it is used.

Reflections
• How should a publisher structure its audience relationships to reduce exposure to a single platform's algorithm?
• Who should define who counts as a journalist, and what protections and duties should follow from that definition?

Chapter 4 — Towards the Wall Between Humans and Machines

19.9 — AI Overviews and the measured cost to publishers
Source
• AI Overviews cut publisher clicks 39.8% in first randomized study. PPC Land.
• Journalism job cuts in 2026. Press Gazette.

Dispatch
The first randomised field experiment on Google's AI Overviews, run across 1,065 United States participants and more than 68,000 searches, found that the summaries cut outbound publisher clicks by 39.8 per cent and raised zero-click searches by 34.5 per cent. Because participants were randomly assigned to see or hide the overview, the study provides a causal estimate rather than a correlation, and it observed that more than 87 per cent of overviews sat above every organic result. Press Gazette continued to track journalism job cuts through 2026.

For news organisations this connects a mechanism to a payroll. When a summary answers the query on the results page, the visit that once funded the reporting does not occur, and the causal design here strengthens the case that the loss is real rather than seasonal. The proximity of measured traffic decline and continuing newsroom cuts describes a direct pressure on the economics of production.

The pattern establishes the premise for everything that follows in this chapter. Information is scarce and costly to produce, and a distribution model that consumes it without returning readers removes the revenue that pays for it. Once publishers can measure that loss precisely, the question of whether to keep the gate open becomes a financial calculation rather than an article of faith.

Reflections
• When outbound clicks fall by nearly forty per cent, which revenue models can still sustain original reporting at scale?
• What evidence would persuade a board that opting out of a dominant search engine protects rather than harms the business?

19.10 — Adaptation without walls, from Reuters to Yahoo
Source
• Reuters launches Model Context Protocol server to bring trusted news into AI workflows. Reuters / Editor and Publisher.
• Yahoo Scout shows AI search doesn't have to cannibalize publisher traffic. Playwire.
• Digital News Report 2026, episode 4, how people are using AI chatbots for news. Reuters Institute.

Dispatch
Several publishers are choosing engagement over exclusion. Reuters launched a Model Context Protocol server that lets customers' AI agents access licensed Reuters content programmatically, turning machine access into a paid, controlled product rather than an open crawl. Playwire reported that Yahoo's Scout demonstrates AI search can send traffic back to publishers, and the Reuters Institute's Digital News Report examined how audiences actually use AI chatbots for news.

For strategists these items describe the constructive alternative to closing the gate. Reuters is selling machine access on defined terms, with authentication and subscription rather than free scraping, which converts an AI reader into a customer. Yahoo Scout suggests that a referral-preserving design is technically possible, which matters because it shows the extractive pattern is a choice rather than a law of the medium.

The pattern is the same equilibrium seen in the licensing chapter, now expressed in architecture. A publisher can let machines in through a metered, paid door that returns value, and the MCP model shows what that door looks like in practice. Adaptation and enclosure are two responses to one problem, and the sector is testing both at once.

Reflections
• What commercial terms turn machine access to content from a cost into a durable revenue line?
• If referral-preserving AI search is technically possible, why would a dominant platform choose a design that removes traffic?

19.11 — The nuclear option, and the wall with two gates
Source
• Once Unimaginable, Publishers Are Preparing to Opt Out of Google Search. Adweek.
• Cloudflare's latest AI rankings expose the web's biggest free rider. Business Insider.

Dispatch
Adweek reported that publishers are preparing what it calls the nuclear option, removing themselves from Google Search by blocking its crawler, because Google uses a single crawler to both index sites and train its AI models and will not license content.

From September 15, Cloudflare will set new sign-ups and free-tier customers to block multi-purpose crawlers on any page carrying ads, and USA Today Inc. said it is prepared to delist from Google within six to twelve months, having built licensing deals with Meta, Microsoft and Amazon instead. Business Insider, citing Cloudflare's crawl-to-refer data, reported that Anthropic's bots crawled roughly 2,800 pages for every referral in early July, improved from about 8,800 to one in April, while DuckDuckGo returned close to one referral for every three crawls.

For decision-makers this is the moment the bargain is renegotiated at the level of infrastructure. Stephanie Cohen of Cloudflare framed the aim precisely, a technical solution that lets a publisher remain discoverable without giving content away for free, which reframes crawling as a permissioned transaction rather than an open default. The crawl-to-refer ratios give publishers a number to point to, since a reader who never returns is a cost, and a business cannot subsidise a counterparty that takes thousands of times more than it gives.

The pattern completes the week. When the reciprocal exchange fails, the people who fund information build a boundary, and that boundary separates two kinds of visitor. Human readers remain welcome, and machine readers must identify themselves, accept rate limits and quotas, request permission and, where appropriate, pay. This is enclosure as a rational response to a broken bargain, and it is the direct answer to the measured traffic loss in the previous signals.
Reflections
• When a crawler returns one reader for every few thousand pages taken, at what ratio does open access stop being rational?
• How should a publisher weigh the reach of remaining in a dominant index against the revenue lost to zero-click answers?

Conclusion
The signals this week describe one structural shift. The open web was built on reciprocity, machines could read freely because they returned readers, and that reciprocity is now measured, priced and, in places, withdrawn. The randomised evidence on AI Overviews, the crawl-to-refer ratios from Cloudflare, and the licensing settlements from Getty to Reuters all point to the same recognition, that information is scarce and costly and must be paid for if it is to keep being produced.

The central question running through the week is who pays for knowledge once machines can read it without a reader ever arriving. Europe is answering through law, with Denmark and France insisting that content on a platform generates a payable claim, and the answer is spreading to Nigeria and to collective bodies such as SPUR. Publishers are answering through architecture, with Reuters selling metered machine access and others preparing to close the crawler entirely. Litigation, licensing and innovation are moving the sector towards an equilibrium in which access to quality journalism is a transaction rather than a gift.

That equilibrium produces a distinct image. The web becomes a walled library, welcoming to people and guarded against machines that will not pay. It has two gates. One is marked for human access and stands open, because readers are the point of journalism and always were. The other is marked as a machine gate, an API gate, where authorisation is required and rate limits, quotas and permissions apply, because a machine that consumes without returning value is a cost the commons cannot carry.

This wall is a warning rather than an accusation. It shows what follows when the largest gateways decline to share the value they capture, and it invites the firms that operate at that scale to set a constructive example by paying for what they use. Quality journalism is scarce and expensive, and it deserves proper remuneration, and the combination of claims, licences and new products can still produce a sustainable settlement in which both gates stay open on fair terms.

The illustration for this dispatch makes the settlement visible, a great wall drawn around the library of human knowledge, with a lit human gate that welcomes readers and a separate machine gate where access is authorised, metered and paid, a modern Chinese wall that keeps the commons open to people while asking every machine to identify itself, respect the limits and honour the value it takes.

(Synthetic Image)