Governance and control

Andrew Ng argues model liability cannot guarantee downstream AI control

Andrew Ng argued that California SB 1047's model-level liability could not ensure harmless downstream use because open-model alignment can be removed and closed models can be jailbroken, proposing regulation of dangerous applications rather than general-purpose AI technology.

0 comments · 0 votesOpen discussion

Public discussion is readable by everyone. Sign in to comment, reply, or vote.

No comments yet. Start the discussion.

CURRENT ASSESSMENT · REVISION 1
TOWARD DOOM27confidence 58/100

Why it moved the index

Magnitude 27 reflects a concrete governance-control gap: model providers cannot reliably prevent downstream harmful adaptation or jailbreaks, complicating enforceable frontier-risk policy. Confidence 58 reflects a specific first-person argument grounded in known alignment-removal and jailbreak mechanisms, capped because it infers policy effectiveness without empirically measuring SB 1047's safety or innovation effects.

AUDIT TRAIL

Assessment history

  1. R1
    Toward 27 · confidence 58

    Adds a distinct application-versus-model governance mechanism beyond the existing general argument against an AI-development pause.

    14 Aug 2026