The AI safety community doesn’t agree on how worried to be. But most discussions flatten this into “doomers vs optimists.” The actual disagreement is more interesting.
I built a semantic search tool that indexes 392 episodes from 80,000 Hours, AXRP, Dwarkesh Patel, The Inside View and more. Searching p(doom) and related terms surfaces some striking contrasts:
Robert Miles puts his p(doom) at 90–99%: “My mainline prediction is doom. It doesn’t look good.”
Eliezer Yudkowsky on those who agree with his arguments but have lower p(doom): they “enact the ritual of the young optimistic scientist who charges forth with no ideas of the difficulties.”
Scott Alexander and Daniel Kokotajlo both land around 20% — the lowest on the AI 2027 team. Alexander notes he’s “not entirely convinced we won’t get alignment by default.”
Will MacAskill sits at 10–20%, calling himself “optimistic today” — but notes this is among the lowest estimates in serious circles.
Sundar Pichai also estimates ~10%.
Zvi Mowshowitz has moved up to ~70%.
What’s interesting isn’t just the numbers — it’s what drives the disagreement. Christiano focuses on conflict scenarios: AI systems acting on their own goals, described as “little green men getting ugly.” Yudkowsky focuses on capability jumps that make intervention impossible. LeCun thinks the whole framing is wrong.
These aren’t just vibes — they’re searchable, timestamped arguments from the primary sources.
You can explore the full disagreement here: AI Safety Search — Leita — search p(doom), existential risk, or any researcher by name
Why the downvotes?
Maybe because P(doom) ranges from 10% to 99% excludes many people that state a lower P(doom) or refuse to state a number.
Maybe because Will MacAskill sits at 10–20%, calling himself “optimistic today” — but notes this is among the lowest estimates in serious circles implies people that state a lower number are not serious.
Those were my reasons to think of downvoting. In the end I didn’t do it because at the time the post was already in the negatives.
There is a huge selection bias coming into play here, where people that appear in AI safety podcasts or use the expression P(doom) have self-selected for higher numbers than people that don’t, and this is not addressed in the post.
Ok—that’s fair. There’s also the issue that doom means very different things to different people.
You’re making good points. The selection bias one especially.
The corpus is AI safety-focused podcasts (AXRP, 80,000 Hours,
The Inside View, etc.), so it skews toward people who already
take alignment seriously. LeCun, Ng, and others who think the
framing is wrong don’t typically appear on those shows, and
when they do they’re pushing back on the premise rather than
giving a number.
The “10 to 99%” framing was honest about what’s in this corpus
but you’re right that it implies a complete range. The “MacAskill
among the lowest in serious circles” line was sloppy. That
phrasing came from his own framing on 80k Hours but it carries
an implication I didn’t intend.
If anyone has pointers to AI safety podcast episodes with people
in the under-5% range or rejecting the framing entirely, I’d
appreciate it. That’s a real gap in coverage I should address.
Thanks for asking, honestly not sure why the early downvotes.
Guess the framing read too promotional first time around. Karma’s
recovered since. Worth mentioning the mobile experience was broken
back then; that’s fixed now if it matters.
I appreciate the initiative! AI Safety is rich with disagreements, and it’s nice to have an opportunity to easily map out the range of existing views. Thanks for sharing!
Thanks! The disagreements are what make this field so interesting to map , the gap between Yudkowsky and MacAskill isn’t just about probability estimates, it’s about fundamentally different models of how AI development unfolds.