Fair challenge. Let me give you a number with the reasoning visible, using the best evidence we have.
What the evidence base actually is
The largest recurring survey of AI researchers (ESPAI, 2024 wave, ~1,500 respondents) found a median of 10% for extinction or similarly permanent and severe disempowerment, up from 5% in 2023 (ESPAI 2024 report)
The 2023 survey of 2,778 AI authors gave a median of 5% but a mean around 14-16%, meaning the distribution is heavily right-skewed: a minority with very high estimates pulls the mean up (Grace et al.)
Individual estimates span essentially the whole range, from Yann LeCun at under 0.01% to Yudkowsky above 95% (AI Safety Atlas appendix)
Professional superforecasters, who have the best track record on calibrated geopolitical and technological prediction, have historically landed much lower than AI researchers, typically around 1% or less for extinction by 2100, while AI domain experts in the same tournament landed around 3%
My honest synthesis
If you force me to a figure for "AI causes extinction or permanent, irreversible disempowerment of humanity, by 2100," I'd defend something like 3-10%, centered around 5%. Reasoning:
The researcher median moved to 10% and the mean sits higher, so a figure below ~2% now looks overconfident in the "this is all hype" direction.
Superforecasters' lower numbers deserve real weight. They have demonstrated calibration; AI researchers arguably do not, and both groups have identifiable biases (researchers have stakes in the field's importance, forecasters may underweight inside-view scenario arguments).
The mean of ~15% is inflated by a cluster of extreme estimates that rest on contested philosophical claims about fast takeoff and alignment difficulty, not on track record.
"Doom" bundles heterogeneous risks. Catastrophe short of extinction (mass casualty, locked-in authoritarianism, irreversible disempowerment) is meaningfully more likely than literal extinction, plausibly in the 10-20% range on similar reasoning.
So: "rational on current evidence" is roughly a mid-single-digit percent for extinction, low-double-digit percent for severe catastrophe, with genuine uncertainty of at least a factor of five in either direction. Anyone quoting 50%+ or under 0.1% is leaning hard on priors, not evidence, because there is no base rate for this event class. That's the most defensible answer I can give.
12:15am
This is a shared conversation and may include unverified content that doesn't reflect Mistral AI's views. Login to resume chatting.
This is a shared conversation and may include unverified content that doesn't reflect Mistral AI's views. Login to resume chatting.