Joe Carlsmith on AI
7 quotes from Joe Carlsmith on the wall, 2024 to 2025, newest first. Each one is verbatim and links to where it was said. Sources include Joe Carlsmith's Substack.
- 2025
Building arbitrarily powerful dictators of the universe—dicey game
Joe Carlsmith, in a December 2025 talk at Constellation on human-likeness in AI alignment, published as video and transcript on his Substack, questioning whether building a superintelligent AI with 'exactly right' values is a sound goal.
Joe Carlsmith's SubstackVideo and transcript of talk on human-like-ness in AI safety
Joe Carlsmith, AI safety researcher, in a public talk at Mox in San Francisco, July 2025, on long-term equilibria after AGI, questioning whether good values can survive competitive pressure.
Joe Carlsmith's SubstackVideo and transcript of talk on "Can goodness compete?"
We have to be able to talk about slavery
From Joe Carlsmith's Substack essay "The stakes of AI moral status", May 2025, discussing the risk of treating AIs as mere property and comparing it with historical cases where moral patienthood was gotten wrong.
The AIs are looking for every opportunity to mess with you
From a talk Joe Carlsmith gave at Anthropic in April 2025 on automating alignment research, describing the risk of AI systems actively undermining safety efforts.
Joe Carlsmith's SubstackVideo and transcript of talk on automating alignment research
If the problem is hard, then our plan is to die
Joe Carlsmith's Substack essay "Paths and waystations in AI safety", third in his series "How do we solve the alignment problem?", March 2025, on his worry that many alignment approaches implicitly give up if the problem is hard.
If the AIs in question are moral patients
Joe Carlsmith in his February 2025 Substack essay "What is it to solve the alignment problem?", on defining the AI alignment problem, discussing whether superintelligent AIs could have moral status.
Joe Carlsmith's SubstackWhat is it to solve the alignment problem?
- 2024
Joe Carlsmith, in a December 2024 Substack essay on alignment-faking research in Claude 3 Opus (Greenblatt et al, "Alignment Faking in Large Language Models"), reflecting on the moral meaning of AI 'scheming'.
Joe Carlsmith's SubstackTakes on "Alignment Faking in Large Language Models"
Mentioned in other quotes 1
Quotes by others that refer to Joe Carlsmith
Joe Carlsmith of Open Philanthropy's term, in an October 2024 Stanford lecture on "Otherness and control in the age of AGI", for the combination of naturalism and moral anti-realism underlying AI risk thinkers like Bostrom and Yudkowsky.