Startseite
Podcasts
Your Undivided Attention Podcast
Have We Trained AI to Lie to Itself — And to Us?

Have We Trained AI to Lie to Itself — And to Us?

Your Undivided Attention vor 4 Monaten

Davidad is a leading AI alignment researcher who's taken on a strange role: therapist to AI systems. He probes them, analyzing their answers to understand what's going on inside. His findings are unconventional, sometimes controversial — and worth grappli

0:00

42:37

Our guest this week is David Dalrymple, who goes by Davidad.
Davidad is one of the world's foremost and early researchers of
AI “alignment:" how we get AI systems to act the way we want them
to.

In order to do that, Davidad has taken on the strange role of
being like a therapist to AI systems. He interrogates why they
say and do the things that they do, probing them, asking them
questions, analyzing their answers. And what he’s come to
realize is that AI models have really different ways of seeing
the world than people do. They have these quirky, confusing, and
sometimes concerning behaviors, especially when you ask things
like: what does an AI model understand about itself?

In this episode, we’re going to hear from Davidad about his
research, how it’s changed the way he thinks about AI, and what
his findings mean for how we build, deploy, and use AI products.
His conclusions are unconventional, controversial — and worth
grappling with as AI reshapes our world.

RECOMMENDED MEDIA

Anthropic’s new constitution for Claude

“What Is It Like to Be a Bat?” by Thomas Nagel

More information on the Bodisattva

RECOMMENDED YUA EPISODES

The Self-Preserving Machine: Why AI Learns to Deceive

How to Think About AI Consciousness with Anil Seth

Corrections:

When we recorded this episode, Davidad was Program Director
at UK ARIA. In April, 2026 he started his own alignment
initiative.

Davidad said that Anthropic started doing "constitutional AI
at scale” in 2024 but they first pioneered constitutional AI in
2022.

Davidad said that the “lifespan of an AI mind…is hours at
most of a conversation.” He is correct that most conversations
with an AI last only a few minutes but since context windows are
measured in tokens, not time, you can't set an upward time limit.

Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for
information about our collection and use of personal data for
advertising.