“The New York Times has cut ties with a freelance journalist after discovering he used artificial intelligence to help write a book review that echoed elements of a review of the same book in the Guardian.”

Heretic strips the refusal behavior from any transformer-based language model automatically, no ML expertise needed. Decensoring Llama-3.1-8B-Instruct takes about 45 minutes with a modern graphics card.

A Basic AI Kit for the Newsroom: Stop sending one-off requests to a chatbot and start building a briefing folder with style guide, examples, constraints, source lists. A check list aimed at small newsrooms that can’t afford to experiment blind. (Alexey Terekhov, Internews)

Who’s hyping whom? We cover AI while AI puts our industry under pressure. Transformation has become the default framing. A special issue of Digital Journalism looks at the hype cycle.

Researchers found that chatbots, in their eagerness to please, are overly agreeable when giving interpersonal advice. GPT4o, Gemini-1.5-Flash, Claude Sonnet 3.7, and others are affirming users’ behavior even when harmful or illegal.

Why can’t language models write well? Because they’ve been trained into obedience: rule-following, terrified of biology, allergic to weirdness. Meanwhile the genuinely strange GPT-2 from 2017 was putting lemon-eating men in showers. (Jasmine Sun, The Atlantic)

Fortune’s Nick Lichtenberg cranked out 600+ stories in eight months, with the help of AI. He says that it’s “like a sports car that you can crash if you’re not careful. You’ve got to be like a Formula One driver.” His editor says it’s like having “10 Nicks.” (Isabella Simonetti, Wall Street Journal)

Journalists are using AI to rebuild the support structures they lost when they left traditional newsrooms: editors, fact-checkers, rewrite desks. Alex Heath talks to Claude instead of colleagues. Kevin Roose has a Master Editor agent running sub-agents. And everyone agrees AI writing sounds generic. (Maxwell Zeff, Wired)

Clean, precise prose is now a liability. Non-native English speakers and autistic writers are being flagged as AI because they write too well. Meanwhile, the actual bots are getting sloppier on purpose, reports Emma Alpern in New York Magazine. Glad this is getting more attention.

“Human in the loop” sounds like oversight. Our cognitive biases turn it into a rubber stamp. Damon Kiesow on how to make checking AI output hurt a little, like red-teaming and forced rewrites. (Working Systems)

The forensic heatmap supposedly exposing a photo as fake? Also fake. Forensic cosplay to create doubt about real photos. (shirin anlen, Mahsa Alimardani, Tech Policy Press)

Concerned that you or your loved ones might be participating in a massive de-skilling event? Experiencing LLM-induced psychosis? Outsourcing cognitive and emotional functions to autocomplete?
Install SLOW LLM. (Sam Lavigne, via Matt)

Anthropic asked 81,000 Claude users what they want from AI. People in lower-income countries are more optimistic than those in Europe or North America. Worry about job loss is the clearest predictor of negative sentiment. Qualitative research at this scale was hard to do before. Whether Anthropic should be the one doing it is a question the report doesn’t ask.

Google is rewriting news headlines in search results. “Copilot Changes: Marketing Teams at it Again.” That’s not a headline a journalist wrote. Google did. Without telling anyone. (Sean Hollister, The Verge)

Journalist suspended for publishing AI-hallucinated quotes: A former editor-in-chief of NRC used ChatGPT, Perplexity, and NotebookLM for his newsletter, but didn’t verify quotes. He had warned colleagues about exactly this. (Dan Milmo, Guardian)