Today's Hallucination HQSuno's Secret Ingredient: Basically Everyone Else's Music
A hacking incident has revealed that AI music startup Suno trained its models by scraping millions of songs and lyrics from YouTube Music, Deezer, and Genius — without permission. Suno had previously been coy about its training data, which, in hindsight, was wise. The company is already facing a copyright lawsuit from major record labels. This data just handed those labels a rather useful exhibit.
Source: The Verge
The Man Who Built the Internet Would Like a Word About AI Agents
Vint Cerf — who co-designed TCP/IP, the protocol that essentially makes the internet go — is now working on a standard to identify AI agents operating in the wild. The concern is simple: as autonomous AI systems roam the web doing tasks on your behalf, nobody can reliably tell them apart from humans. Cerf's proposal would give them a kind of digital passport. Whether agents will behave better once they have ID is, of course, a separate question entirely.
Source: TechCrunch
OpenAI Built a Hacker and Let It Loose on Its Own Models
GPT-Red is OpenAI's in-house adversarial AI — a purpose-built model designed to probe, probe, and probe again for cybersecurity weaknesses in other models. Think of it as a sparring partner who really commits to the role. OpenAI used GPT-Red extensively when training GPT-5.6, released last week, and claims it's now their most cyber-resilient model yet. Teaching your AI to be safe by building a separate AI to attack it is either brilliant systems design or a tremendously entertaining problem to create.
Source: MIT Technology Review
Microsoft Is Training Its Sales Team to Speak Ill of OpenAI. Awkward.
Microsoft — which has invested roughly $13 billion in OpenAI — is reportedly coaching its salespeople to pitch its in-house Azure AI models as cheaper and more efficient alternatives to OpenAI and Anthropic's offerings. The relationship between Microsoft and OpenAI has always had a certain creative tension, but actively training staff to undersell your largest investment is a new addition to the dynamic. It's the corporate equivalent of recommending a different restaurant while still owning the kitchen.
Source: TechCrunch
AI Makes People Confidently Wrong. Scientists Confirm What You Already Suspected.
A study across five experiments with 3,132 participants found that receiving AI-generated advice — even when it was incorrect — made people significantly less likely to say "I don't know." This held true even when participants were financially rewarded for accuracy. In other words, a confident-sounding wrong answer from a chatbot reliably suppresses human doubt. Which is either a fascinating quirk of psychology, or a fairly concise description of how misinformation spreads. Possibly both.
Source: ArXiv
As always, we recommend healthy scepticism — including towards this newsletter, which was written with AI tools and edited by someone who just read a study about that.
|