Dispatch #23 - Cyber Territories
Dispatch 23
The Machine as Reader
Cyber Territories — Dispatch from the Frontiers
The machine as reader. (Synthetic Image)
For most of the history of journalism, the reader was the point. This week’s signals describe a system in which the first reader is a machine, and in which the physical record itself is being converted into material for that machine to read.
Two events set the frame. Google withdrew a feature that let anyone generate plausible satellite imagery of real places, which is the record of the physical world. Anthropic has been buying and destroying printed books after scanning them, which is the record of human thought. Both turn something durable into something fluid, while the institutions that verify durable things are asked to keep up.
Around them, the commercial layer is settling. Australia has finalised a charge on platform advertising revenue, Getty has licensed display rights to OpenAI, and UK publishers are signing confidential agreements with Google while criticising the terms in public. Underneath sits the audience: Australians now reach news through social platforms more often than directly, The New York Times reports exposure to declining search traffic, and the largest American newspaper group has hired Palantir to turn anonymous visits into known relationships.
The theme is intertwining. Physical and digital record, human and synthetic reader, editorial and commercial decision are now bound together. The professional task is deciding which of those bindings a news organisation accepts.
I. On the Record of the World
Three signals about what remains verifiable when the underlying record becomes generative.
23.1 — Google withdraws synthetic imagery from Google Earth
Google removed an image-generation feature from Google Earth about a day after launch, saying people were “sharing screenshots of generated imagery that appear to violate our policies” and that it would roll the feature back “while we work on implementing stronger guardrails”, as The Independent reported. Open-source researcher Henk van Ess told NPR he had generated refugees at the Mexican border, a nuclear plant in Iran and a hospital with a bomb crater in Gaza, and that “nothing was refused”.
Satellite imagery occupies a specific place in verification work, treated by investigative desks and press agencies as close to independent evidence precisely because it is expensive to produce and hard to fake at scale. A consumer feature that generates convincing variations changes the cost of forgery for an entire category of reporting, and watermarking protects the file rather than the screenshot that circulates.
The speed of the withdrawal is the useful part. Platform product cycles now move faster than the verification community can assess them, and the correction depended on public criticism arriving within hours. A guardrail applied after global release amounts to testing on the information ecosystem.
Reflections When the imagery used to confirm events becomes generative by default, what evidence does a newsroom accept as sufficient? Which verification capabilities should press agencies build in-house rather than source from the companies shipping these features?
23.2 — India presses platforms and rebuilds fact-checking around AI
India’s technology ministry has asked Meta for a compliance roadmap on moderation, data localisation and faster removal of flagged material, including algorithmic changes to slow manipulated content and deepfakes, Mathrubhumi reported. In parallel, the Press Information Bureau’s Fact-Check Unit lost its proposed statutory status after the Bombay High Court struck down the relevant IT Rules provision, and the government is now building an AI chatbot for real-time claim verification, an archive of past checks and a detection system for manipulated content, according to India Today.
The combination is instructive. A court removed the legal basis for a state fact-checking authority, and the executive responded with a technical capability that requires no statute. The unit published more than 2,900 fact-checks by April, and accreditation rules already allow debarment of journalists for two to five years, so the capacity being automated is substantial.
For European readers the comparison is procedural. The AI Act and the DSA place obligations on providers through defined processes with review; an administrative verification capability built into government systems concentrates the same function without them. The question is which institution may declare something false, and how that declaration can be contested.
Reflections Where should the authority to label a claim false sit when courts limit statutory fact-checking bodies? How does a news organisation cooperate with state verification systems while preserving independent editorial judgement?
23.3 — Anthropic converts printed books into training material
Anthropic has been buying used and rare books in bulk, cutting off their spines, scanning the loose pages and sending the remains for recycling, a practice examined critically in The Spectator. The article links the destruction to last year’s $1.5 billion copyright settlement, in which a United States court accepted training on text while finding that Anthropic had acquired titles without authorisation, and notes that destroying the physical copy was treated as a way to avoid retaining a duplicate.
That reasoning belongs to a United States framework and should be read carefully in Europe. Under droit d’auteur and the CDSM Directive the questions are lawfulness of the source, applicability of text and data mining exceptions and effectiveness of reservation of rights, and none of them turn on whether the physical object survives the scan. Lawful acquisition of a copy and lawful use of the expression it contains remain separate matters.
The cultural point is the stronger one. A library holds a book so a later reader can return to it; a training pipeline holds statistical traces so a model can approximate it. Only one of those allows a future disagreement about what the text actually says.
Reflections What obligations should a company accept toward the physical cultural record it consumes for training? If out-of-print works survive only as model weights, who is responsible for the loss?
II. On the Price of Access
Three signals about law and licensing working as two halves of the same process.
23.4 — Australia finalises its charge on platform advertising revenue
Labor has finalised the News Bargaining Incentive and narrowed its base to Australian digital advertising revenue, reported by the Australian Financial Review, which also noted that the carve-out for professional networking sites has been reversed, bringing LinkedIn within scope. Reuters set out the mechanics: 2.5%, up from 2.25%, on platforms with more than A$250 million in local revenue, offset by qualifying agreements with at least six publishers rather than four, with larger offsets for regional outlets and all proceeds directed to the news sector. Standalone AI services remain outside the scheme.
This is the clearest recent example of law and negotiation operating together. The charge establishes the boundary, the offsets establish the price, and the design makes a commercial agreement cheaper than the statutory alternative, with Treasury expecting roughly A$250 million a year to move toward journalism.
Two features deserve attention in Brussels. Including LinkedIn treats professional networks carrying news as distribution; excluding standalone AI services treats generative systems as a separate regulatory problem. European legislators considering remuneration mechanisms face both decisions in a market that already has a neighbouring right.
Reflections If a statutory charge sets the effective floor for licensing, what remains for publishers to negotiate? Should generative AI services be brought inside a bargaining mechanism, or governed under a separate instrument?
23.5 — UK publishers weigh confidential Google terms while search changes shape
Press Gazette describes a prisoner’s dilemma among UK publishers, with major titles signing confidential two-year agreements covering AI training, grounding and AI Mode on take-it-or-leave-it terms including non-disclosure and no-sue clauses, earning single-figure millions per year. One source called them “no-sue deals” that amount to “renting peace”; Jason Kint of Digital Content Next observed that Google seeks what it needs “without putting a direct monetary value to the content licensing”, while Madhav Chinnappa argued for a collective access model built on APIs and subscriptions.
The product side moved in the same week. Google has begun placing a Top Stories carousel inside AI Overviews for news queries, appearing in roughly 15% to 17% of trending news searches in the US and UK according to Newzdash analysis reported by Press Gazette, with publishers using the Search Console generative-AI exclusion likely losing eligibility for that placement. Newzdash founder John Shehata called it “a double-edged sword for publishers”.
Together these describe one mechanism. The CMA ruling gave publishers a right to opt out of generative uses without harming conventional rankings; product design attaches a traffic cost to exercising it and commercial terms attach a cash cost to litigating it. Rights that are expensive to exercise are exercised rarely, which is the test any remedy has to pass.
Reflections What conditions would make a collective licensing model more attractive than an individual confidential agreement? How should a regulator measure whether an opt-out right is genuinely available in commercial practice?
23.6 — Getty licenses display rights and finds a new revenue line
Getty Images and OpenAI announced a multi-year, non-exclusive display partnership in late June, allowing OpenAI to show licensed Getty pictures to ChatGPT users with attribution while granting no model-training rights, according to a MarketBeat analysis carried by The Globe and Mail. The structure mirrors Getty’s October 2025 agreement with Perplexity. The context matters: around $2.0 billion of total debt, roughly $194 million of expected cash interest for 2026, shares at $0.58 in June against an active NYSE compliance notice, and a pending $3.7 billion merger with Shutterstock conditioned by the CMA on divesting the UK and global editorial business.
The separation of display rights from training rights is the transferable idea. It creates recurring, auditable revenue without transferring the asset that generates future value, and it prices distribution rather than ingestion. For press agencies, whose archives combine editorial value with indemnified provenance, that distinction applies directly.
The financial detail carries a caution. Getty negotiated under balance sheet pressure, and terms agreed under pressure set reference points for everyone who negotiates later. Scarcity supports price only when the seller can afford to wait.
Reflections Which rights in an archive should be licensed for display, and which should remain unavailable at any price? How do rights holders avoid setting sector-wide reference prices during periods of individual financial strain?
III. On the Audience That Moves
Three signals about where readers arrive from, and what publishers are building in response.
23.7 — Social access overtakes direct visits in Australia
The twelfth Digital News Report: Australia found social media has passed direct visits to news sites and apps for the first time, with television at 57%, social at 56% and direct access at 52%, reported by AdNews. Among 18 to 24 year olds, 60% have never read a printed newspaper and TikTok use for news reached 48%, up twelve points. Nine percent now use chatbots for news, while 49% distrust news from chatbots and 51% distrust news from social media.
The payment data complicates the decline narrative. Australia records the highest share of consumers paying for digital-only news brands at 35%, and trust in self-selected sources rose five points to 54% against 21% for social and 19% for chatbots. Readers distrust the intermediary while using it, and pay for brands they have chosen deliberately.
That defines the commercial opportunity. Distribution has moved to surfaces where trust is low, and value accrues to organisations that convert an untrusted arrival into a chosen relationship. Creators now serve as a news source for 43% of Australians, which is the competitive set for attention, though the standards set is a different one.
Reflections What does it take to convert an arrival from a low-trust surface into a deliberate subscription decision? Which audience segments justify continued investment in platform distribution, and which do not?
23.8 — The Times leans into video while a large chain turns to data infrastructure
The New York Times added 280,000 digital-only subscribers in the second quarter, down from 310,000, reaching 13.3 million subscribers and $538 million in subscription revenue, with CEO Meredith Kopit Levien saying the company “isn’t immune” to an ecosystem shaped by a small number of large technology companies, according to Nieman Lab. The response is video: eight video journalists hired in January, twelve more roles open, a new Shows tab, and a 10% rise in adjusted operating costs. The Times also disclosed $32.9 million in generative-AI litigation costs since early 2024.
USA Today Co., owner of more than 200 local titles, has partnered with Palantir to analyse and monetise audience behaviour, with CEO Mike Reed describing the goal as converting anonymous interactions into “known, orchestrated first-party relationships”, also reported by Nieman Lab. Unique visitors fell from 180 million to 158 million in a quarter, attributed to lower search referrals rather than lower demand, and Reed said he would prefer a fair licensing deal with Google but would block access if necessary.
The two responses illustrate the choice available to publishers of scale. One invests in a format intermediaries cannot easily summarise; the other invests in knowing its audience more precisely, accepting a supplier whose government work makes this a governance decision at board level rather than a procurement decision at management level.
Reflections Which audience data capabilities should a news organisation own outright rather than source from a vendor? How much of a shift toward video is defensible on editorial grounds when the driver is algorithmic visibility?
23.9 — Building on owned ground while the optimisation industry reorganises
Puneet Gupt, COO of Times Internet, told WAN-IFRA’s Digital Media India event that publishers spent years treating the crawler as their most important reader and “lost track of who our loyal readers were”, summarising his message as “don’t build on borrowed land”. He called this “an ecosystem change” rather than an algorithm change, said “the runway is short”, and set out three requirements: AI-native operations designed as a complete redesign, direct reader relationships built on consent-based data, and trust-led revenue, since “trust is a conversion multiplier in whatever product you create”.
The optimisation industry has already reorganised around the new surfaces. The Verge examined how the SEO sector is attempting to influence AI answers through generative engine optimisation, self-serving listicles and, in cases documented by Microsoft, hidden instructions behind summarise buttons. Google updated its search spam policies in May 2026 to cover manipulation of generative responses, which confirms the practice is widespread enough to require enforcement.
The economics are clear. When visibility inside an answer becomes the scarce good, an industry forms to purchase it, and the cost of maintaining a trustworthy answer layer rises for whoever operates it. Publishers holding a direct relationship with readers hold the one asset this market cannot manufacture.
Reflections Which parts of a publisher’s business currently depend on ground it does not own? How should a newsroom respond when competitors invest in influencing AI answers about their own coverage?
IV. On the Craft That Holds
Three signals about newsroom practice where AI has become routine.
23.10 — Reuters sets speed against editorial responsibility
An analysis of Reuters’ newsroom AI describes assisted reporting in which technology surfaces material while “a journalist must establish what it means”, framing the conflict as automated speed against accountable editorial judgement. The pillars remain experimentation, journalist approval on the principle that “a Reuters story remains a Reuters story”, disclosure of material AI use, and verification, with generative systems prohibited from creating or enhancing news imagery. FactGenie reached 150 journalists and halved the average time to send non-corporate alerts, though error and correction rates remain unpublished.
The industry data gives context. A 2026 Reuters Institute survey found 64% of media leaders consider back-end automation important while 13% describe current initiatives as transformational, and 67% report that AI has reduced no roles. A study of 1,004 UK journalists found 60% reporting some integration, 44% with rules for human oversight and 27% with guidelines on bias and fairness.
The governance insight belongs in every management team. Human review functions as a safeguard only when the reviewer has time, access to the source, an understanding of how the system fails and the authority to reject its output. Sequential automation moves an early error through every later stage before anyone sees it, which is the specific risk agentic workflows introduce at scale.
Reflections What would a meaningful disclosure standard look like when AI use is described as material? Which stages of a reporting chain must always show their sources before a human approval carries weight?
23.11 — African newsrooms and a Korean regulator build AI around the desk
Briefly News, part of Legit, described the development of Nova at INMA’s Africa Newsroom & Tech Summit, where 25 to 40 writers producing up to six stories a day are supported by one or two copy editors. An early custom GPT produced hallucinated corrections, so writers spent “more time looking for the error that wasn’t there than actually fixing real errors”, leading to a rebuilt Chrome-based Moffield Copilot beside the CMS with media-law and policy compliance checks. Editors doubled the stories they could review, and readers remained the constraint: “they can tell when it is AI-created”.
Punch Newspapers set a stricter boundary at the same summit, reported by Công Luận: AI may support ideation, research, summarisation, transcription and proofreading and may generate SEO metadata after publication, while writing articles and editing news photographs remain prohibited. Editor Oyetunji Abioye put it plainly: AI assists the journalist, and every published article has a journalist who answers for it.
Korea’s Internet Newspaper Ethics Committee has taken the same logic into self-regulation, deploying an AI keyword-monitoring system in August that collects articles from 850 participating internet newspapers every two hours and flags profanity, discriminatory expressions, unclear sourcing and crime or suicide reporting. Review-office head Park Young-rye said the purpose is to support accumulated standards and specialist judgement rather than replace human review; detection triggers a human reading, with decisions taken in weekly and monthly meetings.
Reflections Which editorial tasks should remain off limits to automation regardless of measured efficiency gains? When a self-regulatory body automates detection, how does it prevent the tool’s categories from becoming the standard itself?
23.12 — A new book argues AI is accelerating a system already under strain
Journalism in the Age of AI argues that current developments are “largely accelerating a broken system” that has weakened journalists and public trust for decades, tracing AI across ideation, sourcing, verification, storytelling and distribution in an excerpt published by Nieman Lab. CalMatters uses AI to track state legislators and generate tips that Sisi Wei says “would have taken a data-savvy political reporter weeks, if not months, to find”, while Full Fact, Maldita.es and the European Fact-Checking Standards Network run a multilingual system that extracts claims from short-form video.
The failure cases are equally specific. Die Aktuelle published a fabricated Michael Schumacher interview and settled with his family, CNET corrected machine-generated stories, and in November 2025 reporters at Suncoast Searchlight found undisclosed AI editing had inserted hallucinated quotations into their drafts. A study of 22 fact-checkers found available tools often failed to fit workflows and increased workloads.
The strategic reading is that AI amplifies whatever structure a newsroom already has. Organisations with defined standards and clear accountability gain capacity; organisations without them acquire faster ways to publish unverified material. The differentiator is institutional, which is the asset a quality newsroom holds and a model cannot reproduce.
Reflections Which existing weaknesses in your editorial process would AI adoption make more visible within a year? What consent should journalists have over AI systems that alter their drafts before publication?
V. On the Machine as Reader
Three signals about what happens when machines read, speak and are believed.
23.13 — Readers rate AI writing highly and increasingly listen to synthetic voices
A study led by Villanova University researchers and published by Cambridge University Press found that participants aged 18 to 81 could not reliably distinguish AI-generated short stories from those by published authors, as the BBC reported. The highest ratings went to stories written by AI and described as human. Senior author Dr Deena Weisberg said this “reveals a bias towards narratives written by real people”, that “public assumptions about AI’s capabilities are increasingly out of date”, and that AI writing “tends to be clearer, more direct and easier to process”. Greater AI literacy improved detection; literary expertise did not.
The audio layer is scaling in parallel. The Columbia Journalism Review examined synthetic narration across the Times, the Post, the Journal, The New Yorker and The Atlantic, noting that ElevenLabs has passed $500 million in annual recurring revenue with early investor talks at a $22 billion valuation. Around 20% of New Yorker subscribers listen to narrated stories, and the Journal’s Read to Me feature was used roughly 5 million times in a year. The New Yorker declines to narrate fiction synthetically, treating it as performance.
Both findings point at the same commercial fact. Clarity and convenience convert, and a machine supplies both at marginal cost. What it supplies less well is the register that carries irony, restraint and judgement, which is much of what a reader pays a brand for.
Reflections If audiences rate machine-written text highly when they believe a person wrote it, what does authorship promise them? Where does synthetic narration serve accessibility, and where does it change the work itself?
23.14 — Chatbots outrank elected representatives on trust
A Bloomberg newsletter reported survey findings that people now regard AI chatbots as more trustworthy than politicians. The underlying work comes from the Collective Intelligence Project’s Global Dialogues, whose published snapshot draws on roughly 4,000 participants across four rounds and finds respondents consistently rating AI as capable of better decisions on their behalf than their government representatives, with the share rising each round. Only public research institutions and family doctors are trusted more than chatbots.
Read alongside the Australian data, where 49% distrust news from chatbots, this produces a specific picture. People assign high trust to a system they use for personal reasoning and low trust to the same system as a news source. Trust is being allocated by function, and the function of interpretation is moving toward an interface owned by a few companies.
For democratic institutions this is a measurable signal well before it becomes an electoral one. Representative bodies now compete for credibility with a product that answers immediately, never appears evasive and has no constituency to manage. The comparison is unfair to the institutions and flattering to the product, which is why it holds.
Reflections What should elected institutions learn from an interface that outperforms them on perceived honesty? How do news organisations report on public trust in AI systems without either amplifying or dismissing the finding?
23.15 — DeepMind names recursive self-improvement as the investment thesis
Jasjeet Sekhon, chief strategy officer at Google DeepMind, told an audience at UC Berkeley that recursive self-improvement, meaning AI systems that build better versions of themselves, is the core investment thesis behind the sector’s capital expenditure, as first reported by The Information. He acknowledged that AI revenues “don’t sustain the capital expenditures we’re making so far” and offered the analogy that steam engines were used to build the next steam engine, against annual Google capital expenditure running at roughly $200 billion.
This is unusually direct and deserves to be taken at face value by anyone negotiating with these companies. The economics of the current buildout are justified by a capability that does not yet exist, so near-term revenue matters mainly as evidence that the thesis remains credible to investors. Content licensing sits inside that calculation as a cost to be minimised, which explains a great deal about the terms on offer.
There is no free lunch in this arithmetic. Capital committed today must be recovered from advertising, subscription, enterprise contracts or the value extracted from material others produced. Publishers negotiating access rights in 2026 are negotiating against a balance sheet that requires the cheapest possible input.
Reflections How should a rights holder price access to a counterparty whose business case depends on a future capability? If the recursive self-improvement thesis proves slower than expected, which of today’s licensing terms will look mispriced?
Conclusion
The reader has changed identity. A crawler reads to index, a model reads to train, an answer engine reads to summarise, a synthetic voice reads aloud, and an optimisation industry writes for the machine that reads first. At the end of that chain a person still decides whether to believe what they are told, and every institution in this dispatch is arguing about who gets paid, credited or held responsible along the way.
Two costs are becoming visible together. One is producing verified information, always high and now competing with material generated at almost no cost. The other is maintaining the record itself, whether satellite imagery reliable enough to confirm an event or a printed book a future reader can consult. Both fall on institutions that were already carrying them, while the value created afterwards accrues elsewhere.
The regulatory responses are converging on a workable shape. Australia has priced access through a charge with commercial offsets, the CMA has established an opt-out right whose practical availability is now being tested by product design, and Kenya proposes effects-based jurisdiction with a risk framework and a public register for high-risk systems. Each seeks an equilibrium between claims and contracts, and none will work if the right it creates is more expensive to use than to give away.
For news organisations the conclusion is narrow and demanding. Own the relationship with the reader, own the verification capability, and treat every agreement with an intermediary as a decision about which part of the business sits on ground belonging to someone else. Press agencies carry an additional obligation, since their material becomes the reference point other newsrooms and, increasingly, machines rely upon.
The robot on the cover of this dispatch is reading a newspaper. Someone still had to write it, verify it, print it and stand behind it, and that remains the part nobody has automated. The institutions that keep doing that work hold the last position in the chain where a claim can be checked against something real.
Patrick Lacroix writes Cyber Territories in a personal capacity. AI tools are used for research and drafting; the author retains full editorial responsibility.