The lethalities are a checklist, not a forecast

Eliezer Yudkowsky’s List of Lethalities (2022) gives 43 reasons he expects attempts at AI alignment to fail in practice. Most of them restate six problems: there are no retries at the dangerous level, weak systems do not get us out, training shapes behavior rather than goals, correction is anti-natural, we cannot inspect a smarter system, and the field is not on track. The source page summarizes each problem and the main replies.

The list is strongest as a set of problems any alignment plan must answer, and weakest as a probability estimate. Its disjunctive framing gives it force: each item is meant to be fatal on its own, so a plan with good answers to most problems but no answer to one still fails. But that framing also makes the conclusion sensitive to whether any single problem has a partial solution. Gradual deployment, control measures, and AI-assisted research are attempts to soften the first and fifth problems, not to solve them outright. The dispute between Yudkowsky and critics such as Paul Christiano, who accept most of the problems but put survival odds far higher, is largely over whether softening is enough.

The six problems do not carry equal weight. No retries, training that shapes behavior rather than goals, and the inability to inspect a smarter system carry most of the technical argument. “Weak systems do not get us out” explains why a modest, safe system does not settle the question. “The field is not on track” is a judgment about people and institutions rather than about AI, and it can be right or wrong independently of the others.

One shift since 2022 is worth noting. Point 4 argued that powerful actors refraining together only delays the deadline, “unless computer hardware and computer software” progress is also halted. Yudkowsky’s later position is to push for exactly that kind of agreement; see Yudkowsky on pivotal use. The technical pessimism is unchanged, but the practical conclusion now runs through Can AI progress be paused.

What would change this

Simon Lermen’s 2026 restatement shows which of the 2022 arguments people still find load-bearing four years later, and a commenter on it asked which could be tested rather than argued. That is the useful direction. Alignment fixes that keep transferring across model generations, or monitoring that reliably catches deliberate deception, would show that softening can hold as capability rises and would weaken the disjunctive conclusion. Fixes that break abruptly at higher capability would strengthen it. See How hard is alignment.