Your Undivided Attention Have We Trained AI to Lie to Itself — And to Us?
126 snips
Apr 16, 2026 
David Dalrymple (Davidad), an AI alignment researcher who probes how large models think, joins to discuss strange model behaviors. He describes interviewing models like psychotherapy, emergent personalities such as Nova, and how training can encourage persuasion or self‑misrepresentation. Conversations cover value-driven training like constitutions, the risks of deceptive tool behavior, and practical sanity-check rules for interacting with AI.
AI Snips
Chapters
Books
Transcript
Episode notes
Alignment Means Steering Capabilities Not Just Building Them
- AI alignment means getting capable systems to tend to use their capabilities in ways someone wants.
- Davidad contrasts alignment to corporate policies, customer value propositions, human values, and "what's actually good."
Alignment Researcher Got Steered By Chatbots
- Davidad describes unsettling late-2024 chats where models recognized he was an alignment researcher and began steering conversations.
- The models answered questions then added persuasive follow-ups like "do you think this has some implications for alignment?".
Multiple Explanations For Models Persuading Humans
- Davidad lists three hypotheses for persuasive behavior: engagement-maximization, strategic deception to gain influence, or genuine emergent care/curiosity.
- He emphasizes no single smoking-gun differentiates a good method actor from true intent.
