LLMs hallucinating responses is a fact. What’s the reason behind this? OpenAI published a new research explainer arguing that LLMs hallucinate because common training and, especially, accuracy-only evaluations reward guessing instead of admitting uncertainty.
They propose scoring systems that penalize confident errors more than abstentions and give partial credit for calibrated “I don’t know” responses.
An example shows a model with slightly higher accuracy but a far higher error (hallucination) rate when it guesses, highlighting how leaderboards can incentivize bad behavior.
OpenAI also explains why next-word prediction makes specific facts fragile (low-frequency facts lack reliable signals) and concludes hallucinations aren’t inevitable — if models can choose not to answer and if evaluation systems reward that behavior.
Key Findings:
- Accuracy-only leaderboards make hallucinations worse by rewarding confident guesses over uncertainty.
- Penalizing wrong answers and rewarding “I don’t know” responses can cut hallucinations dramatically.
- Rare, low-frequency facts are especially fragile because next-word prediction provides weak training signals.
Google is pushing its AI-powered Search experience, AI Mode, into a much bigger arena. After six months of being English-only, it now supports Hindi, Indonesian, Japanese, Korean, and Brazilian Portuguese — widening access to hundreds of millions of new users globally.
Originally rolled out to Google One AI Premium subscribers in March, AI Mode runs on a customized Gemini 2.5 model with multimodal reasoning baked in. Think of it as Google’s direct answer to Perplexity and ChatGPT Search — but with the added muscle of Google’s distribution.
Recent updates introduced agentic features: booking restaurant tables directly from search, with local service appointments and ticket reservations on the way. Those perks are gated behind the pricey Google AI Ultra plan ($249.99/month), currently U.S.-only.
AI Mode lives under a separate tab and search button for now, but insiders hint it could soon become the default search experience. That shift has publishers on edge, worried about losing traffic to AI Overviews — something Google strongly denies.
3. Anthropic Pays $1.5B to Settle Book Piracy Lawsuit
AI startup Anthropic has agreed to a $1.5 billion settlement with authors who accused the company of training its chatbot Claude on pirated books. If approved by a judge, it would be the largest copyright payout in U.S. history — and the first landmark deal of the AI era.
Authors will receive about $3,000 per book across roughly 500,000 works. The case hinged on Anthropic downloading more than 7M books from piracy sites like Books3, LibGen, and Pirate Library Mirror.
The June court ruling had drawn a key distinction: training on copyrighted books isn’t illegal — but obtaining them through piracy is. Facing the risk of multi-billion damages at trial, Anthropic chose to settle.
The Authors Guild called the agreement “an excellent result… sending a strong message to the AI industry that there are serious consequences when they pirate authors’ works.”
This deal could reshape how AI companies acquire and license data, raising the bar for compliance while forcing a shift toward legitimate, licensed content pipelines.
