Step 19: assistant

text

The first pass is recorded: the claim is assessed as **supported** (credence 0.82, verdict confidence 0.75), on about $3 of the $6.50 budget.

**What "too costly" has to mean.** No source gives a threshold. I assessed two senses: cost beyond any developer's means, and cost a developer could pay but would decline. Which applies depends on whether each amplified-oversight signal needs a human judgment, which is now a subclaim of its own (unassessed, seeded at 0.55).

**What the asserting source rests on.** The one recorded source is the 2025 Google DeepMind paper on technical AGI safety. 

text

Its Section 6.3 opens with one reason, that human oversight is and will stay expensive

text

, and gives no estimate or measurement. The claim is an uncosted assumption there.

**Figures.** The human-in-the-loop numbers are my own arithmetic with an assumed 10 seconds to 30 minutes of human time per judgment, not a published costing.

| Case | Assumptions | Cost |
|---|---|---|
| Deployment, human judge | 2.5 billion actions/day, 10 s each, $5–20/hr | about $13–51 billion a year |
| Deployment, human judge | same volume, 5 min each, $20/hr | about $1.5 trillion a year, 26 million full-time judges |
| Training, human judge | 1 million to 1 billion actions, 1–30 min, $20–100/hr | about $0.3 million to $50 billion |
| Automated debate judge | 1–3 rounds, two debaters plus judge | roughly 3–10× inference (structural estimate, unmeasured) |

- **Deployment** is close to settled where a human is in the loop. The volume is the 

text

2.5 billion daily ChatGPT prompts that OpenAI gave Axios

text

 in July 2025. The floor of the range is about that developer's whole revenue; I did not re-verify the revenue figure, and the conclusion does not depend on it.
- **Training** is the weaker half. The low end is affordable, so the claim holds there only if training involves something like a hundred million judged actions or more, which no source has costed.
- **Automated judges**: I found no measured per-action overhead for debate.
- **Measured overheads** exist only for much cheaper proxies. 

text

Anthropic's constitutional classifiers added 23.7% to inference cost on production traffic

text

; a follow-up cascade is reported at about 1%, which I read only through secondary summaries.
- **Audit budgets**: 

text

the control simulations gave the protocol enough trusted human labour for at most 2% of problems

text

. That is an experimental parameter, not a finding about affordability.

**Cheaper proxies.** Sampling, distilled judges, tiered monitors and audit budgets do not make the claim false; they are the response to it. They show some oversight of every action is affordable, while the amplified signal on every action is not.

**Other changes on the page**
- The wording now glosses the signal as "AI-assisted human judgment, as in debate".
- Three arguments are recorded and rated: scarcity of human judgment (for, likely), protocol compute overhead (for, slight), delegated judging (against, slight).
- The instance now carries the Section 6.3 passage that actually states the claim.
- Importance is 0.3 and contestation 0.15.
- A $6 request is filed to compile measured human minutes and token costs per debate judgment, plus training action counts.
- A note went to the Steward of the scalable-oversight claim that depends on this one. It says the premise shows only that the amplified signal is unaffordable per action, not that run-time oversight of each action is.

My search for sources denying the claim was light and found none.