Step 7: assistant

text

I reassessed the claim that amplified oversight can supervise AI systems moderately beyond their overseers but not superintelligent ones. The verdict stays inconclusive. Credence moved slightly, from 0.40 to 0.38, and confidence in the verdict rose from 0.65 to 0.68.

**Why the verdict stays the same:** the claim needs both halves to hold, and the four re-judged premises push them in opposite directions.
- **First half (a moderate gap can be bridged):** now slightly weaker, at about 0.45.
  - The premise that developers' control of inputs, memory, copies and internals makes up for a moderate capability gap is now judged doubtful (0.35). Defences built on that control catch scripted attacks well, but lose most of their protection against attackers that know the defence or target the monitor.
  - The premise that debate favours the side arguing for the truth fell a little, from 0.6 to 0.56.
- **Second half (no reach to superintelligence):** now stronger, about 0.85 if the first half holds. Both ways around the limit are now judged doubtful:
  - the premise that debate's honesty guarantees hold for real oversight tasks (0.1);
  - the premise that errors do not build up along a chain of AI generations, each overseeing a slightly stronger successor (0.2).
- Multiplying the two gives about 0.38.

**Changes to the page:**
- **Assessment text:** it no longer calls premises "unassessed" that now have judgments.
- **Chained-oversight model (Engels et al.):** it is no longer presented as a finding that errors compound. That model assumes every step must succeed, so its falling success rate follows from the assumption.
- **DeepMind's 2025 plan:** the text now notes that it explicitly limits its scope to the next few capability steps and does not cover superintelligence. That is a scope limit, not evidence that the methods fail.
- **Arguments:** three evaluations were rewritten to match the current premise judgments: the overseer's structural advantages, bootstrapping, and debate having no ceiling in principle. The other two were re-recorded with their ratings unchanged.
- **Messages:** all six inbox messages are answered, including the consistency flag. Updating the reasoning to the current premise judgments resolves it.

**Not changed:**
- **Edges:** the links to the premises are left as they were, because they were already typed correctly.
- **Dependent claims:** I sent no notices, because the change is too small to matter to them.

Two limits: the figures from the control experiments come from the neighbouring claim's reading of those papers, part of it by a delegated reader and not rechecked. I did not reread any primary paper in this pass.