In this issue: The case for using chatbots to surface facts and arguments, and against outsourcing judgment, taste, and every word under your name. Evan Hirsch from OpenAI on keeping humans in the loop. Plus: Google’s trusted sources, AI ads, and Thomson Reuters’ models.
Hi, I'm Ole Reissmann, a journalist who builds things. I'm the first Director of AI at SPIEGEL. Before that: podcasts, news product development, platform strategy. I write about AI and journalism and send a newsletter you might enjoy.
Google launches “Expert Intelligence”, an initiative to bring trusted sources into Gemini. First: Play Books in Gemini Notebook, with 100,000+ titles from Penguin Random House, Macmillan, O’Reilly and others. Coming soon: “Third-party subscriptions, textbooks, and broader surface availability like AI Mode in Search and the Gemini App.”
ChatGPT Ads has reached a revenue pace equivalent to $1 billion a year, OpenAI says, in less than 200 days. The release is careful to say ads are kept separate from answers, and then promises to make them more native.
Can data make a newsroom more diverse in its choice of sources? In 2023, a Canadian newsroom let researchers demo an AI tool that reads published articles and sorts every quote: politicians, authorities, experts, organizations, celebrities, media, and unaffiliated. After seeing results, the newsroom decided against it. (Sage)
Thomson Reuters has built its own frontier models, aimed at legal, tax, and journalism. The open-weight Thomson-1.0-Small is based on Qwen3.6-35B-A3B and is on HuggingFace for academic and non-commercial use. Another version stays in-house.
Time, a publisher of listicles, has named the 100 most influential people in AI. The usual suspects, plus nine from media. Eleven if you count Wikimedia and TikTok. Two of them are suing AI companies: A.G. Sulzberger is going after OpenAI, and Andrea Bartz after Anthropic.
How does ChatGPT search the web? For free users in instant mode, barely at all. The answer is grounded in page titles and cached 200-character snippets from OpenAI’s own index, with no page opened 93% of the time, an SEO firm reports. Paid thinking mode adds scraped Google results and real page fetches.
The New York Times is testing AI-generated search summaries without humans in the loop. (Max Tani, Semafor)
Firmware hacking used to be state-actor territory. Now it’s a hobby project. A security researcher used Opus 5 to own his peripherals: “Network-connected devices seem near universally fucked at this point? … The next few years are going to be extremely interesting.” (Chaz Schlarp)
Anthropic has drafted a “German Professional Secrecy Addendum” for lawyers using Claude. EU data residency is coming next. (Dennis Fordan, LinkedIn)
How to follow an impeachment trial that runs 120 days? Jaemark Tordecilla’s session-by-session reference for the Sara Duterte trial turns hours of Philippine Senate video into timed transcripts and source-linked digests.
