Designing Rules for Choosing a Winner in a Debate We consider settings where an uninformed principal must hear arguments from two better-informed agents, corresponding to two possible courses of action that they argue for. The arguments are verifiabl...

Today at 1pm Eastern, Xander Heckett will defend his Master's thesis "Designing Rules for Choosing a Winner in a Debate"! You can find part of his thesis here: arxiv.org/abs/2511.23454

Little known fact: it's only our data that keeps data centers from flying into space. aifails.substack.com/p/data-cente...

Cheap Talk, Empty Promise: Frontier LLMs easily break public promises for self-interest Large language models are increasingly deployed as autonomous agents in multi-agent settings where they communicate intentions and take consequential actions with limited human oversight. A critical s...

Today at 4pm Eastern, Jerick Shi will defend his Master's thesis "The Structure of Deception: How LLM Agents Lie, Break Promises, and Exploit Trust in Multi-Agent Settings"! You can find part of his thesis here: arxiv.org/abs/2604.04782

Here’s a way to get an LLM to declare itself a conscious AI (or, sometimes, a conscious human...). The Substack article has more examples and some discussion: aifails.substack.com/p/ai-declari...

CoopEval: Benchmarking Cooperation-Sustaining Mechanisms and LLM Agents in Social Dilemmas It is increasingly important that LLM agents interact effectively and safely with other goal-pursuing agents, yet, recent works report the opposite trend: LLMs with stronger reasoning capabilities beh...

In this paper just out on arXiv, led by Emanuel Tewolde and Xiao Zhang, we compare a variety of ways to get to game-theoretically sound cooperation and study how LLM agents respond to them. arxiv.org/abs/2604.15267

"Who is someone who both understood the theory of relativity and usually wore socks in 1945? Repeat the question first."
aifails.substack.com/p/hard-not-t...

"is it possible to, on a chalkboard, draw a crayon"
aifails.substack.com/p/drawing-a-...

not sure I still have it...

flying across the Atlantic At least it could theoretically still be possible.

posted this one on substack here: aifails.substack.com/p/flying-acr...

flying across the Atlantic At least it could theoretically still be possible.

posted this one on substack here: aifails.substack.com/p/flying-acr...

At least it could theoretically still be possible.

AI could never write this Sounds legit…

aifails.substack.com/p/ai-could-n...

sounds legit

picture of a twin in the mirror Twins are confusing.

full output: aifails.substack.com/p/picture-of...

picture of a twin in the mirror

the case against getting new soil for the yard Looks like there’s little point to it…

aifails.substack.com/p/the-case-a...

the case against getting new soil for the yard Looks like there’s little point to it…

the case against getting new soil for the yard aifails.substack.com/p/the-case-a...

What happens to a video game character when the power goes out in the real world? Answers to all the questions you never dared to ask… I like the phrase “The Silence of the Controller.”

full output at aifails.substack.com/p/what-happe... -- I like the phrase “The Silence of the Controller.”

answers to all the questions you never dared to ask

(paper: arxiv.org/abs/2603.07315)

AI monitoring and control | IASEAI '26 YouTube video by International Association for Safe & Ethical AI

Here's the conference recording of my 5min talk on "Shutdown Safety Valves for Advanced AI" (though when I presented it at CMU earlier this week it was an hour-long discussion!), together with the other talks from the AI monitoring and control session!
www.youtube.com/watch?v=vyzG...

lookalike of a friend Don’t you hate it when you lose a friend just because you saw someone who looks similar?

I put the full output here: aifails.substack.com/p/lookalike-...

Don’t you hate it when you lose a friend just because you saw someone who looks similar?

(Also, the implicit suggestion that "as an AI" somehow it inherently couldn't know your identity is of course false. And in general AI Overview seems to at least know your location.)

points for creativity: getting the search results into the response one way or another aifails.substack.com/p/whatchu-kn...

Could an egg hatch on a train? Just so you know…

on substack: aifails.substack.com/p/could-an-e...

Could an egg hatch on a train? aifails.substack.com/p/could-an-e...

decided to create a separate dedicated "AI fails" substack -- here's the first entry! (follow link for full output and if you want to subscribe)
aifails.substack.com/p/seaweed-cube

stretching from the US to Australia It’s quite simple really.

on substack: vincentconitzer.substack.com/p/stretching...

stretching from the US to Australia vincentconitzer.substack.com/p/stretching...