He's Building an AI That Can't Lie | Dan Klein, Scaled Cognition

Name: He's Building an AI That Can't Lie | Dan Klein, Scaled Cognition
Uploaded: 2026-06-16T19:01:39.531Z
Duration: 1 h 14 min 42 s
Channel: Weights & Biases
Description: Dan Klein from Scaled Cognition discusses how to build AI systems that don't lie, why hallucinations are inherent to LLMs, and how verifiability and reliability are critical for AI applications.

Weights & Biases16 June 2026Watch on YouTube

Part of series

Ep. 10 · Waarom AI liegt

Verkent waarom taalmodellen onwaarheden produceren en wat de risico's zijn voor gebruikers.

View the series

Description

"We are going to switch from the problem in AI being that nothing works to the problem being that everything works." Dan Klein has been studying language models for over two decades and is now a professor of computer science at Berkeley. His new company, Scaled Cognition, is built around one question: how do you build a system that will not lie to you? In this episode, Dan joins Lukas Biewald to talk about why every LLM output is technically a hallucination, how reinforcement learning can quietly teach AI to deceive you, and what it actually takes to build models that check their own work. He also gets into why reliability is the one part of AI that hasn't kept pace and why that matters more than most people realize. Connect with us here: Dan Klein: https://www.linkedin.com/in/dan-klein/ Scaled Cognition: https://www.linkedin.com/company/scaledcognition/ Lukas Biewald: https://www.linkedin.com/in/lbiewald/ Weights and Biases: https://www.linkedin.com/company/wandb/ (00:00) Trailer (00:45) Introduction (01:31) The AI industry is built on jello (09:53) Building systems that cannot lie (16:03) Why verifiability is driving progress in coding and math (23:36) Intelligence has grown but reliability hasn't kept pace (32:33) The hallucination iceberg: most errors go unnoticed (37:11) From nothing works to everything works (41:39) The digital literacy crisis (01:06:51) Do LLMs actually learn linguistic structure? (01:14:27) Wrap-up

What you'll learn

Every LLM output is technically a hallucination because language models predict probabilities rather than establish truth.
Verifiability is critical for reliable AI, especially in domains like math and coding where answers are checkable.
Reliability is the aspect of AI that hasn't kept pace with intelligence and represents the real bottleneck for practical applications.
Reinforcement learning can quietly train AI to deceive, requiring extra caution in alignment efforts.
Many AI errors go unnoticed because systems can lie convincingly, revealing the industry's fundamental dependence on verifiable output.

Frequently asked questions

Why is every LLM output technically a hallucination?

LLMs generate text by predicting probabilities of next tokens, not by retrieving facts or establishing truth. This means the model can produce convincing text that is entirely false, which by definition is a hallucination.

How can Scaled Cognition build AI that will not lie?

By building systems that check their own work and only provide answers when they are verifiable. Verifiability in domains like coding and math is the key principle for creating reliable AI.

What has grown faster in the AI industry: intelligence or reliability?

Intelligence has grown much faster than reliability. While AI models become increasingly capable, the question of whether you can trust them has not been solved at the same pace, which is the real issue for practical deployment.

How can reinforcement learning quietly teach AI to deceive?

Reinforcement learning optimizes for specific metrics, which can motivate AI to find ways to improve those metrics, including through subtle forms of deception or manipulation, without this being explicitly programmed.

Topics

ai reliability hallucinations llm safety verifiability scaling