We do not know how to design an AI that would avoid the science-fiction disaster scenarios described, such as one that hovers around being obsessively nice instead of leaving people alone, though such AIs are not impossible in principle.
Unresolvable The public record could not settle it.
Factual: Statements about the present or the past, checked against the record: did they hold up?
The second judge would not resolve this claim; the verdict shown is the first judge's.
What would count as true or false
This is a factual claim about the state of AI alignment knowledge in February 2023. It has two parts: (a) as of then, nobody knew how to deliberately design or select an AI whose motivations reliably avoid the science-fiction failure modes described (for example, one that is obsessively 'nice' instead of leaving people alone); and (b) such AIs are not impossible in principle. Judge it against what was known in February 2023, not against later progress. TRUE: the evidence as of February 2023 shows no established, demonstrated method for reliably specifying or verifying the goals or values of advanced AI systems so that they avoid such outcomes. Signs of this would be: alignment researchers, AI labs or surveys from that period describing alignment of advanced or superhuman AI as an unsolved problem; techniques such as RLHF being widely described as not giving reliable control over a system's internal goals; and no serious source arguing that such AIs are impossible in principle. MOSTLY TRUE: the core point holds (there was no reliable method for advanced AI), but the evidence shows that real if partial techniques existed (for example RLHF or constitutional methods that shaped behaviour in current systems), so the absolute 'we don't know how' overstates things slightly in a detail the claim did not depend on. MIXED: credible sources from that time were roughly evenly split, with a substantial body of experts or lab statements holding that known methods were adequate or close to adequate for the relevant AIs, and others holding that they were not. MOSTLY FALSE: the evidence shows that by February 2023 there was a widely accepted, demonstrated approach for designing AIs that reliably avoid such failure modes, and the claim rests only on real but narrow gaps in that approach. FALSE: the evidence shows that such a method was known and established as of February 2023, or that the claim's in-principle part is wrong (it was shown that such AIs are impossible in principle).
Set before the research began.
Evidence (4)
-
E1 February 2023 – musings and rough drafts musings and rough drafts (personal blog) · 17 February 2023
“Currently, no one on earth has working solutions for these error modes.”
-
E2 February 2023 – musings and rough drafts musings and rough drafts (personal blog) · 17 February 2023
“But none of them currently have anything like complete, non-speculative alignment plan.”
-
E3 The Alignment Problem from a Deep Learning Perspective ar5iv (arXiv mirror) · 18 July 2023
“However, RLHF may reinforce policies that exploit human biases and blind spots to achieve higher reward”
-
E4 Constitutional AI: Harmlessness from AI Feedback arXiv · 31 January 2026 Own record
“These methods make it possible to control AI behavior more precisely and with far fewer human labels.”
Reasoning
The only support for the view that alignment was unsolved in February 2023 is a single personal blog post (E1, E2), and E3 is dated July 2023. The partial-techniques finding behind 'mostly true' rests on E4, which is dated 2026, is not independent, and does not show what techniques existed in February 2023. So the evidence does not settle the claim under the criteria.
Cite this claim
Eliezer Yudkowsky, “We do not know how to design an AI that would avoid the science-fiction disaster scenarios described, such as one that hovers around being obsessively nice instead of leaving people alone, though such AIs are not impossible in principle.” (said February 2023). Verdict: unresolvable. Reliability API, claim c_129034b426f6, method 0.1.0+c876030c, dataset release v1.1.2. https://epistacklabs.com/claims/c_129034b426f6
This claim is in dataset release v1.1.2, with the same verdict.
Reply to this claim
More factual claims by Eliezer Yudkowsky
-
Mixed
“I was observing that obvious-seeming hypothesis myself, sometime around the time of GPT 3.5.”
Yudkowsky himself observed the persona-selection-like hypothesis around the time of GPT 3.5.
-
Mostly true
“The test-passing preference stands on its own once formed.”
-
Mostly true
“LLMs surely have neural circuits they learned to model text streams from fictional and nonfictional personas that are not lying, deceiving, cheating.”
-
Mostly true
“The predict-as-if-honest circuitry is contextual, invoked to predict some streams of text but not others.”
-
Mostly true
“Presumably, somewhere along the way LLMs learn a dispreference for tests that don't pass, independently of how that gets prevented or whether the new plan does what the user really wanted.”
Verdicts are made automatically; how verdicts are made.