|
|
Was this forwarded to you? Get The Fourth Estate AI Brief in your inbox →
|
|
|
|
View this issue in your browser
Tuesday Brief
Two Courts, Two Verdicts: The AI Copyright Split Widens
The interesting thing this week with AI and publishing — we got a split decision. A U.S. judge approved Anthropic's $1.5 billion copyright settlement while an Indian court ruled in OpenAI's favor against a news agency. Same week, opposite outcomes. For community publishers watching from the sidelines, this is the clearest signal yet that where you're located — and who your lawyers are — matters more than who's right.
|
|
TEXAS NEWSROOMS
The Rundown: A peer-level look at who's actually using what.
The details:
- The Chronicle surveyed AI adoption at papers in Houston, Dallas, Austin, and smaller Texas markets.
- Transcription and content summarization are the most common AI applications in local newsrooms.
- Adoption rates vary significantly between metro dailies and community weeklies.
- Concerns about accuracy, editorial oversight, and reader transparency remain consistent across outlets.
Why it matters for us: This is a benchmark piece. If you're a 10,000-circ paper wondering whether you're behind, the Chronicle just gave you a comparison set. Worth 10 minutes to see where your competitors landed.
Read at chron.com →
|
|
|
THE $1.5 BILLION BOOK SETTLEMENT
The Rundown: A federal court gave final approval to Anthropic's $1.5 billion payout — the largest copyright settlement in U.S. history. But read the fine print: this was about books, and about piracy. Anthropic got caught downloading millions of pirated books to train Claude, and the money goes to those book authors and publishers — not to newspapers.
The details:
- The math: about $3,000 for each of roughly 500,000 pirated books, split 50/50 between authors and their publishers by default. Final approval landed July 20, 2026.
- The money was for piracy, not training — Anthropic had downloaded more than 7 million books from pirate “shadow libraries” like LibGen to build its training set.
- A judge had already ruled that training AI on legally purchased books is “spectacularly transformative,” and legal. Buying the books was fine; stealing them wasn't.
- It is not binding precedent: Anthropic settled instead of appealing, so no higher court ruled — other judges don't have to follow it.
Why it matters for us: Don't read this as “my archive is worth $3,000 a story.” The payout punished piracy, not training — and the court signaled that training on content an AI company got legally can be fair use. The real lesson for a publisher: your strongest hand is against outright theft, and a licensing deal beats waiting for a lawsuit that may never set a rule.
Read at npr.org →
|
|
|
INDIA: OPENAI VS. ANI
The Rundown: Right after that Anthropic settlement, here's the flip side. In India, OpenAI just won — for now. The Delhi High Court refused to stop OpenAI from training ChatGPT on news agency ANI's articles, calling it “fair dealing.” It's only an interim ruling, but it fits a global pattern that publishers won't love.
The details:
- On July 24, the Delhi High Court refused ANI's bid to block OpenAI, ruling that training ChatGPT on ANI's articles is — for now — “fair dealing” under India's copyright law.
- The judge's reasoning: ANI couldn't show ChatGPT spits its articles back word-for-word. It hands users summaries, not ANI's exact copy, so it isn't a substitute for the wire service.
- This is only an interim ruling, not a final verdict — and ANI's separate claim, that ChatGPT invented stories and wrongly credited them to ANI, still heads to trial.
- It's part of a trend: the court leaned on the U.S. Anthropic and Meta cases, where training was also called fair use. One outlet called it the weakest spot publishers have been in yet.
Why it matters for us: Notice the direction across today’s cases: courts keep calling AI training itself legal — in the U.S., where Anthropic’s $1.5B was for piracy, not training, and now in India. One caveat: the Delhi court ruled under India’s own “fair dealing” law, not America’s “fair use,” so it doesn’t bind U.S. courts — but the pattern does. Suing to stop training is looking like a losing hand. Publishers’ strongest ground now is narrower: outright piracy, word-for-word copying, and AI putting words in your mouth (ANI’s surviving “you faked our byline” claim is the one to watch). The safer play is still a licensing deal, not a lawsuit.
Read at ndtv.com →
|
|
|
EU AI ACT UPDATE
The Rundown: The EU's regulatory framework just handed publishers a research tool.
The details:
- AI providers must now publish summaries of training data sources, including top domain names used.
- Deepfakes and AI-generated public-interest text require clear visible labeling starting August 2026.
- Nine AI practices are banned outright, including AI-generated non-consensual intimate content.
- SMEs get simplified compliance paths and expanded sandbox access — relevant if you're building internal tools.
- General-purpose AI model rules took effect in August 2025; high-risk system rules phase in through 2028.
Why it matters for us: When those public summaries drop, you can actually check whether your archives were scraped. For a county weekly that's been online since 2005, this is free due diligence — no lawyers required to start.
Read at digital-strategy.ec.europa.eu →
|
|
|
THE TRAFFIC QUESTION
The Rundown: A Malaysian outlet asked the question every publisher is quietly dreading.
The details:
- AI assistants increasingly provide direct answers, reducing the need for users to visit source websites.
- The traditional SEO and traffic model breaks when ChatGPT or Google AI Overview summarizes your story.
- Publishers face a fundamental challenge: their content gets consumed without driving clicks.
- The shift raises sustainability questions for ad-supported journalism specifically.
Why it matters for us: Start tracking how much of your traffic comes from AI referrals now — most analytics tools can show this. If you're a free weekly dependent on web ads, the answer changes your 2027 budget math.
Read at thesun.my →
|
|
|
THE UPDATE
Quick Hits From the Week
The Rundown: Smaller items that crossed the wire.
The details:
- The EU's AI Omnibus package added a ban on AI-generated CSAM, effective December 2026 — another compliance date to calendar if you're using any AI image tools.
- High-risk AI rules for product-integrated systems don't hit until August 2028, giving newsroom tool vendors time to adapt.
- Regulatory sandboxes are expanding for SMEs under the AI Act — if you're experimenting with custom AI tools, there may be a testing pathway with less liability.
|
|
|
What is?
AI Model
What it is: The actual engine behind an AI tool — a massive system trained on enormous piles of text (plus images, code, and audio) until it gets good at predicting what comes next. ChatGPT, Claude, and Gemini are the products people use; the model is the brain running inside each one. Companies ship them in different sizes and generations, which is why you'll hear names like Opus, Sonnet, or GPT-5.
Why publishers care: Here's the practical part: not all models are equal. A big, top-tier model writes sharper, more nuanced copy but costs more and runs slower; a smaller, faster one is cheap and quick but less polished. The skill isn't picking “the best AI” — it's matching the model to the job: a heavyweight for a delicate editorial, a lightweight for routine cleanup. That's where the real savings and speed come from.
|
|
|
|
Two courts, two outcomes. The pattern I keep seeing: publishers who wait for legal clarity will wait forever. The ones making deals now — even small ones — are building bargaining power they'll use later. Next week I'll dig into what those deals actually look like for papers under 25,000 circ. Until then, check your analytics for AI referral traffic. You might be surprised what you find.
Trevor Slette runs Quadd.ai — AI tools built for publishers.
|
|
|
|
|
|
AP style without the AI mush.
Quadd's AP-Style Proofreading flags grammar, style, and headline issues with accept-or-reject suggestions. It tightens copy without inventing facts, flattening your voice, or turning every story into AI mush. The accept-or-reject part matters: you stay the editor. The AI just catches what a tired second pair of eyes would catch — if you had a second pair of eyes.
See how proofreading works →
|
|
|
