hallucination

Definition

Plain language

When an AI confidently states something that isn't true.

Also called: hallucinations, hallucinate, hallucinated, hallucinating

Why it matters: It's the central reliability problem with language models — useful output and fabricated output look identical until you check, which limits where they can be trusted.

For example, a model might confidently cite a paper by 'Smith et al. 2021' that doesn't exist, complete with a plausible-looking title and journal.

Heard on the show

“Nobody hallucinated their squad, nobody lost track of their finances.”

Mentioned in 38 episodes

1. 245
   Fifteen Models Ran Football Clubs for Twenty Years, and Size Didn't Decide It
2. 244
   The Open-Weight Defense That Feeds Attackers Confident, Falsified Answers
3. 242
   Making a Vision Model Better by Showing It Blurry Images
4. 239
   Why the AI-Writing Estimate for Biomedical Papers Jumped From 15% to 89%
5. 214
   The Medical AI Answer That's Accurate, Sourced, and Still Wrong
6. 206
   How Four-Second Clips Become Hours of Playable AI Soccer
7. 201
   One in Four NeurIPS Papers Cites a Reference That Doesn't Exist
8. 183
   Why You Can't Fine-Tune Foresight Into an AI Agent
9. 182
   How a Tiny Model Too Weak to Plan Cuts a Bigger Agent's Hallucinations by 80%
10. 181
    How to Backpropagate Blame Through a Team of Chatbots — And When It Backfires
11. 176
    An AI Designed Its Own Psychology Studies, Then Confirmed What It Found
12. 172
    One Bad Token Can Sink a Model's Math, And You Can Delete It
13. 156
    Why More Human Demonstrations Made a Computer-Use Agent Worse
14. 154
    How a 7B Model Out-Investigates a 72B One by Choosing What to Look At
15. 153
    Catching a Lie From the Inside, When the Words Look Completely Honest
16. 149
    When Cornering a Chatbot Makes It Lie: J.P. Morgan's Case for 'Playing Dead'
17. 130
    Why AI Agents Coordinate Better Through a Shared Board Than a Boss
18. 123
    Five Identical Worlds, One Swapped Model: What Happens When AI Agents Run for Fifteen Days
19. 108
    The Reasoning Cliff: Why Thinking Longer Makes Models Worse at Exact Step-by-Step Tasks
20. 100
    How a Prompt Wrapper Lets a Frontier Model Play Poker Like an Expert
21. 098
    Finding Millions of Readable Concepts Inside a Real, Deployed AI Model
22. 089
    When AI-Written Papers Read Well But the Evidence Underneath Is Broken
23. 076
    Same Model, Organized Differently: How an Agent Architecture Beat Frontier Systems at Research Math
24. 072
    A Robot Made Graphene Without Help, And Caught Itself Hallucinating
25. 070
    When Models Know the Answer But Say the Wrong Thing Anyway
26. 067
    An AI Just Solved a 1996 Erdős Problem—and the Simplest Agent Won
27. 066
    Why Giving an AI Agent More Tools Can Make It Worse at Using a Computer
28. 062
    Treating Hallucinations as Exploits: A Gate-Based Architecture for Agent Safety
29. 059
    Firefly's Inversion: Building Verified Tool-Call Training Data by Working Backward
30. 055
    Why LLM Judges Flip Their Verdicts When You Change the Question Format
31. 052
    An Old Reinforcement Learning Tradeoff Sneaks Back Into LLM Agents
32. 041
    When the Iteration Teaches the Model to Skip the Iteration
33. 039
    When Smarter Agents Get Fooled by Three Extra Nodes in a Database
34. 037
    Why Hallucination Detectors Miss Stale Facts: A Geometric Story About What Models Know But Don't Say
35. 029
    Why Forty-Eight Percent on FrontierMath Isn't the Real Story in DeepMind's New Math Paper
36. 025
    The Missing Gradient Term That Predicts Sycophancy in RLHF
37. 017
    When the Agent Grades Its Own Homework: A Brutal New Benchmark for AI Workers
38. 014
    Why a Constrained Pipeline Beat a Full Coding Agent at Finding Bugs 30-to-1

Related concepts