The BaseVideoDwarkesh Patel (YouTube)18 min152,074 views
Which AI is the best liar?
In an 18-minute video, Dwarkesh Patel reports on a benchmark built by Patel, in which AI models play the social deception game Blood on the Clocktower against each other. Patel says OpenAI's newest models now beat Anthropic's, and that AI models did most of the benchmark's design work.
Why it matters
The same thing, for different readers.
What it means for you
AI models are getting much better at lying convincingly to a team and at spotting liars, according to one podcaster's game-based test of many models.
For researchers
Dwarkesh Patel's Blood on the Clocktower benchmark gives a multi-agent test of deception and detection that Patel says is unlikely to saturate, with code released for others to extend.
The question it raises
Models that can plan secret schemes and bluff a group raise the question of how far people can trust what AI agents say to each other and to humans.
In their own words
Every quotation was found, word for word, in the text we saved from the source. That shows the words are there; it does not make them true, or prove the speaker meant them as quoted.
From YouTube’s automatic captions, which carry no speaker names and can mishear words. Where the speaker is known only from context, the quotation says so.
“The gap between these two is about 430 Elo points. That means that Astro wins about 92% of the games it plays against Opus 4.6.”
“it's reasonable to assume that AIs have gotten much better at building and improving the RL environments that go into their training.”
“In all the games that I've run so far with current models, no AI has ever used this approach to work out who his teammates were.”
“most of the work, even most of the judgment work, was done by Claude.”
How the finding was reached
In our words, from the source. Weigh the finding by how it was made.
- What was done
- A benchmark in which 368 games of Blood on the Clocktower were played with all-good teams being copies of one model and all-evil teams copies of another, using a programmatic storyteller; about $20,000 of API credits.
- What it found
- Patel reports about 430 Elo points between Claude Opus 4.6 and OpenAI's newest model, and says OpenAI's models gained 580 Elo points between GPT-5.4 and its newest, mainly from playing the good side better.
- What it does not show
- Patel says the mixed-team test (about 130 games) was not rigorous and not statistically significant, and that he did not test learning across games. A video by the author, not peer reviewed.
- Status
- A report
The record
- Questions
- Can AI systems act beyond what people intended or allowed? · How fast is AI advancing, and where is it heading?
- People
- Dwarkesh Patel (Speaks)
- What we read
- We read the whole source.
- Found by
- The Radar, run of November 6, 2026, via
yt-dwarkesh-patel-youtube(tier 1: the source’s own channel or an official record) - Cite this
- Dwarkesh Patel (YouTube), “Which AI is the best liar?”, November 5, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/dwarkesh-patel-2026-11-05-which-ai-best-liar