An AI Invented Four Sources to Defend One Wrong Answer (and Anthropic's New Opus 4.8 Bets on Honesty)

From AI Ready Podcast with Harrison Painter by Harrison Painter

May 29, 2026 · 33 min · Episode 96

About this episode

This episode discusses the implications of AI models like Anthropic's Claude Opus 4.8 and Google's Gemini in terms of honesty and source verification.

Anthropic just released Claude Opus 4.8, and the headline improvement is unusual: the model is built to flag its own uncertainty and say "I'm not sure." Anthropic says it's roughly four times less likely to let a flaw pass without catching it. When a company's flagship upgrade is honesty, that tells you something about where we are. Here is the other side of it. Harrison asked Google's Gemini one simple factual question for an article he was writing: did Jeff Dunham use AI to create the opening visuals for his 2024 comedy special? Gemini said yes, confidently, and cited a source. When Harrison pushed on that source, the tool did not check itself. It invented a new one. Then another. By the end it had manufactured four separate references, including a word-for-word on-screen quote that does not exist, before finally admitting the only real source was a single unsourced blog post. This episode walks the whole chain step by step. You will learn: - The exact failure mode: when an AI hits a popular but unverified claim, it gets confident instead of careful, and every round of pushback produces a fresh citation instead of a fresh doubt. - Why the Vectara Hallucination…

People in this episode

Host: Harrison Painter

Topics covered

AI honesty
source verification
AI confidence
fact-checking
AI hallucination

Keywords

AI
Anthropic
Claude Opus 4.8
Gemini
source verification
fact-checking
hallucination
Jeff Dunham
confidence

Mentioned in this episode

Organizations: Anthropic, Google, Vectara

Products: Claude Opus 4.8, Gemini

Books & works: 2024 comedy special, unsourced blog post