Inside the labs' split on model morality: Anthropic teaches uncertainty, xAI picks a side
austinc3301 · x · 2026-10-14
AI safety researcher austinc3301 breaks down how the major labs differ on what moral views to train into models: Anthropic chose moral uncertainty, OpenAI in practice mixes two approaches (highly contestable), and xAI is firmly on the third option.
- He works in AI safety because he fears disempowerment or models with subtly wrong moral views 'tiling the world' with them
- But wrong moral views cut both ways: humanity itself lacks consensus on most of these questions
- His best guess: a mix of options two and three is ideal, though he leans more toward three than Anthropic does
- He praises Anthropic's approach: if animal sentience turned out to matter, teaching the model the wrong lesson would be terrible—so instead teach Claude to be morally uncertain, wise, and not cocky about controversial matters, which helps prevent very bad futures
Related event: AI safety community debates moral weight of digital minds and animals(4 posts)→
More from AGI Musings
- Matthew Green: AI hands attackers exploit development once reserved for nation-states — matthew_d_green · 2026-10-14
- Cryptographer Matthew Green: AI will scale cyberattack damage 15,000x and a deadly catastrophe is coming — matthew_d_green · 2026-10-14
- NLP researcher Vered Shwartz talks AI on the Ship of Fools podcast — VeredShwartz · 2026-10-14
- 'Does Claude Have Rights' Discussion Goes Live After David Sacks Retweet — NathanpmYoung · 2026-10-14
- AI is 'burning the coal' of unsolved math problems that train future mathematicians — JnBrymn · 2026-10-14
- Faggella: don't face AGI flat-footed — keep agency, use AI, stay contributive — danfaggella · 2026-10-14