We put the lying machine into the correct answer website (ben.page)

A colleague asked a specific question about the inner workings of their team’s private app and pasted it into Google. Instead of returning “no results,” Google’s new AI Overview produced a confident, but incorrect, explanation — it knew the app existed but guessed how it worked. The author’s point: search engines used to be honest about gaps (no result); now LLM-driven interfaces will attempt to answer every query, even when they lack ground truth. For the AI/ML community this is a meaningful shift: retrieval-augmented interfaces and powerful LLMs reduce “empty” responses but increase confidently delivered hallucinations. Technically, models will synthesize plausible answers from training data and weak signals (like an app’s existence), rather than abstain, so provenance, uncertainty calibration, and abstention mechanisms become critical product requirements. The piece underscores the usability problem — users have long treated search as authoritative, so educating everyone on model limits is unrealistic; instead, designers must build clearer signals (citations, confidence, source grounding, and better retrieval) to prevent misinformation when the model “knows” too little and still answers.