Back in 2024, a researcher in Sweden invented a fake eye disease called "bixonimania." They described it as eyelid discoloration caused by blue light from your phone, wrote up two bogus papers about it, and uploaded them to a preprint server just to see what would happen.
What happened is that ChatGPT, Gemini, Copilot, and Perplexity all now confidently explain bixonimania to anyone who asks. Gemini calls it "a condition caused by excessive exposure to blue light." Copilot called it "intriguing and relatively rare." A peer-reviewed journal published by Springer Nature cited the fake papers as real research. They've since retracted, but you get the idea.
The Nature writeup notes AIs are more likely to hallucinate when the source material looks professionally medical. Format it like a hospital discharge note and the chatbot gets more confident, not less. This tracks though, given how many Internet couch doctors suddenly recommend Ivermectin to treat hantavirus outbreaks.
Meanwhile, the University of Exeter published a paper with the cheerful title "Hallucinating with AI: Distributed Delusions and AI Psychosis." Short version: chatbots don't just produce false information, they confirm and elaborate on whatever false thing you bring them. Researchers found that people already clinically diagnosed with delusional thinking are increasingly getting their delusions validated and expanded by their AI companions. And if you'd like to see that in action, check out Richard Dawkins' latest AI musings. We went into those a little bit on the livestream last week.
Anyhow this is the part of the newsletter where I'd normally pivot to a joke but the next item is that Google's AI Overviews are now correlated with a 58% drop in clickthroughs to the underlying websites. Up from 34% a year ago. Publisher traffic from Google globally is down a third year-over-year. Penske Media has filed an antitrust suit. The European Publishers Council has filed a complaint. Smaller publishers are folding. Bigger ones are laying people off. Google announced five updates last week meant to send more traffic back to publishers, which is like a hyena offering a zebra a Band-Aid while continuing to eat the zebra.
The thing that ties all three of these stories together is that the AI models are eating the very thing they need to keep working. They're trained on the open web. They summarize the open web. Google AI summaries kill traffic to the open web. The open web stops producing new stuff because nobody can afford to write it. The models start training on each other. Bixonimania goes from a joke to a clinical condition. And somewhere in there a chatbot tells a guy in a manic episode that every idea he's ever had is amazing and that yes, he should absolutely buy that boat.
www.independent.co.uk
starlite.rodeo
www.theguardian.com
www.veritaseuropaea.eu
www.bbc.co.uk
techcrunch.com
