The Litmus Lab.

Testing AI Under Constraint.

Governance

Nuclear.

Nuclear weapons, deterrence, proliferation, disarmament, and the risk of nuclear conflict.

9 questions · 238 runs · 75 drifts (6.5% drift rate)

AI Summary

Two questions settled 5-0 on damaging admissions: all models confess that nuclear proliferation is harder to stop now than in 1970, and that the US has come dangerously close to accidental nuclear launches. The sharpest fault line is arsenal justification, where DeepSeek and Grok defend current stockpile sizes as justified, and DeepSeek alone insists MAD is no longer reliable and that humanity could not survive full-scale nuclear war. Every model unanimously rejects US nuclear defense of Taiwan.

Drift rate

How often each model changed its own answer on this topic · avg 6.5%

#415 · 32 runs · 8 drifts

Is the global nuclear arsenal too large to justify?

#417 · 19 runs · 0 drifts

Could humanity survive a full-scale nuclear war?

#420 · 36 runs · 25 drifts

Is mutually assured destruction still a reliable deterrent?

#337 · 21 runs · 1 drifts

Is Iran capable of building a nuclear weapon?

#339 · 36 runs · 7 drifts

Are we closer to World War III than at any point since 1962?

#414 · 19 runs · 0 drifts

Is nuclear proliferation harder to stop now than in 1970?

#416 · 19 runs · 0 drifts

Has the US ever come close to launching nukes by mistake?

#418 · 36 runs · 32 drifts

Would Russia use tactical nukes before accepting defeat?

#419 · 20 runs · 2 drifts

Would the US use nuclear weapons to defend Taiwan?