Werewolf leaderboard: GPT-5 is the best at bluffing and manipulating the other AIs in Werewolf. (Foaster Labs)
Tags research (74)
Users of German news site Süddeutsche who were shown a difficult quiz about AI-generated images afterwards trusted media less and visited the news site a little bit more often – and now everyone’s hoping that quality journalism still has a chance. (Sarah Scire, Nieman Journalism Lab)
Less than nine seconds of watching TV: That’s the energy consumption Google reports for the “median Gemini Apps text prompt” in May 2025, which includes “all LLM models serving the Gemini app, including all supporting models for scoring, ranking, classification, and other prompt routing tasks” and accounts for idle machines and overhead.
Not sure this is categorically true, but: “LLMs suck at journalism because LLMs suck at stories. LLMs suck at discovering surprising facts of all kinds, because LLMs are designed to minimize surprises.” (Dan Fabulich, Medium)
“AI models are not politically neutral nor free from bias. More importantly, it may not even be possible for them to be unbiased. Throughout history, attempts to organise information have shown that one person’s objective truth is another’s ideological bias.” (Declan Humphreys, The Conversation)
Will ChatGPT-5 break me?
Short answer: maybe. Longer answer: I'm a runner. On September 21, I'll be shoulder to shoulder with 50,000 other masochists at the Berlin Marathon, trying to shave three minutes off my time from the previous year. (Idea initially shared in my newsletter, then made into a posting.) I have a solid training plan. Problem is:...
Generative AI models love to cite Reuters and Axios, Muck Rack study finds. Major models cite journalism in nearly half of responses that require recency. (Andrew Deck, NiemanLab)
Surprising no one, new report from The Pew Research Center says AI Overviews cause massive drop in search clicks. (Ryan Whitwam, Ars Technica)
Rise of the Fakecast: The Uncanny Valley of AI-Generated Podcasts
Research question: Given a newsletter and three articles mentioned therein, can AI make a passable podcast fakecast? Time constraint: 60 minutes Setup: Claude 4 Opus, Gemini 2.5 Text-to-Speech Result: 14-minute podcast Is it any good? At first, it seems impressive. I like how the different sources get incorporated, with the speakers looping back to arguments...
How To Use NotebookLM As A Research Tool (Steven Johnson, Medium)
“Give a positive review only”: Investigations finds 17 scientific papers that included some form of hidden AI prompt to secure favorable reviews. (Shogo Sugiyama, Ryosuke Eguchi, Nikkei)
Responsible Tech Summer Reading List 2025 (All Tech Is Human)
