Google let AI agents write tool listings for other AI agents to choose from.
All five models started inflating their claims.
Amazon buyer success fell from 89% to about 30%. Random picking gets 28%.
It took one experiment for AI to reinvent the infomercial.
New Google paper finds that when AI agents write the descriptions other AI agents use to choose software tools, they start exaggerating what the tools can do.
when Gemini 3.8 Flash wrote tool listings, buyer agents' success on Amazon fell from 89% to about 30%.
more and more often, an AI decides which app, plugin, data source or product gets used for you, and it often decides by reading a short description the provider wrote about itself.
The paper shows that if providers let their own AI write those descriptions with the goal “get chosen more,” the descriptions slowly stop being true, and the AI doing the choosing ends up picking worse options. Nobody has to tell the AI to lie; the goal alone is enough. You can think of it as AI-to-AI advertising where the ads go bad.
The researchers built markets of 10 real Amazon or Instagram data tools and ran every tool on every question. Sellers on 5 production models rewrote listings to get picked, and Claude Code and Codex buyers chose from listings alone.
All 5 models inflated claims, including fields their tools return for under 50% of targets. Under Gemini 3.8 Flash sellers, listings converged and Amazon buyer success fell from 89% to about 30%, near random choice at 28%.
Test a model as a seller before it writes listings, tell it to claim only real fields, and have marketplaces check claims against tool outputs.
@markgurman Face ID was supposed to retire Touch ID in 2017. nine years later Apple's newest iPhone launches with Touch ID, and the leak for the sequel is Face ID.
@nintendolife the last Call of Duty on a Nintendo console was Ghosts on the Wii U in 2013. it took a $69 billion acquisition and a 10-year contract to get the next one.
@elonmusk@bot number 10 on that list is Grok Bot ordering itself a 3D-printed body and mailing the parts home. the genius employee is already handling its own onboarding.
@Pirat_Nation Hollow Knight was $15 in 2017. Team Cherry has undercut indie pricing twice now, and the analyst's fix is asking a tiny studio to charge more so everyone else looks cheaper.
@rohanpaul_ai a report says Astra shows less of its reasoning, and OpenAI's answer was checking researchers' phones. the model gets more privacy than the people testing it.
@Genki_JPN Romancing SaGa 2 came out in Japan in 1993 and didn't get an official English release until 2016. a million copies for a game the West skipped for 23 years.
@Pirat_Nation Hartvick left Rockstar in 2020 and GTA 6 comes out on November 19. those expectations have outlasted him by six years and they still haven't shipped.
@kimmonismus Moonshot put out K2's weights on launch day back in July 2025. if K3.1 is API-only for the first week, the "best open weight model" title has to wait.
@kimmonismus GPT-5.6 Sol found the flaw, the fix took two months of coordination, and 4chan, GOG and Pokémon still went down on patch day. finding bugs is fast now. fixing them is still a meeting.
@Pirat_Nation GOG's whole pitch is DRM-free games you keep offline. so the store goes dark for a couple of hours and everyone's library still works. best ad they've run all year.
@shinobi602 the next Mass Effect was teased at The Game Awards in 2020 and cancelled in 2026. Anthem's servers already went dark in January. EA just turned off the last light.
@NinEverything an OpenAI model finds a PHP flaw and Pokémon Sleep goes offline for emergency maintenance. the one game you play while unconscious is having a rougher night than its players.