If a person has to approve every single AI decision, what happens when the volume explodes and they get tired?
Can per-decision human review ever maintain capacity against volume and fatigue?
This explores whether a person approving every individual AI decision can keep up as output volume grows and reviewers tire, or whether human review has to change shape.
This explores whether a person approving every individual AI decision can keep up as volume grows and attention wears down. The corpus says no, not as a standing design, and the reason goes beyond humans being slow. Can AI generate knowledge faster than humans can evaluate it? describes the mismatch as a loop. AI produces knowledge faster than people can verify it, and the gap feeds itself because the evaluation tools are increasingly AI-generated too. It works like monetary hyperinflation: each unreviewed output makes confidence in all of them worth less.
Fatigue makes it worse, because exhaustive review degrades each review instead of just costing more. In Does targeted human oversight beat both full autonomy and exhaustive review?, a research agent with step-by-step human oversight got a 50% accept rate, and full autonomy got 25%. A mode that routed only high-uncertainty decisions to a human got 87.5%. The paper's explanation is rubber-stamping: constant interruptions train reviewers to click through. Human attention is a budget, and per-decision review spends it on trivial calls. This is one system in one setting, so read it as a strong hint rather than a law.
Routing needs a signal for where a human matters, and the corpus has ingredients for one, though none of them is an oversight system yet. Does model confidence predict robustness to prompt changes? finds that confident models resist rephrasing while low-confidence ones swing widely, so confidence is a usable warning light. Does step-level confidence outperform global averaging for trace filtering? adds that the warning belongs at the step level, since averaging over a whole trace hides the moment reasoning breaks down. Together they suggest flagging the shaky step for a human instead of handing over the whole output for a full read.
Machines can absorb the bulk, but they raise their own trust problem. Can inference scaling help reviewers catch errors humans miss? surfaced critical flaws at STOC and ICML that had passed human review. Exhaustive human review is therefore not only unscalable but sometimes not enough. Can agents evaluate AI outputs more reliably than language models? cut judge drift from 31% to 0.27%, yet its memory module cascaded errors, so automated layers need isolation. Does setting temperature to zero actually make LLM outputs reliable? warns that a reviewer giving the same verdict every time is consistent, not necessarily correct. Can automated review loops handle AI-generated research at scale? shows review-refine loops measurably improving AI-generated research, which is encouraging, though it runs into the hyperinflation worry that AI is grading AI.
The corpus also has a gap. Does added monitoring improve protection at acceptable cost? designs a comparison of monitoring schemes at equal review cost and equal false-alert workload, but reports no results. So there is no evidence yet on whether extra monitoring pays for itself. What the evidence does support is that per-decision review cannot hold capacity, while a scarce human budget spent at high-uncertainty points can. The open question is whether the automated layers doing the rest can be trusted without reviving the same bottleneck.
Sources 9 notes
AI produces knowledge faster than human judgment can verify it, collapsing epistemic confidence just as monetary hyperinflation collapses purchasing power. The gap self-reinforces because evaluation tools are themselves AI-generated, trapping the system in acceleration.
AutoResearchClaw's confidence-routed CoPilot mode achieved 87.5% accept rate, beating full autonomy (25%) and step-by-step oversight (50%). Selective human intervention on high-stakes decisions avoids both uncaught errors and the rubber-stamping fatigue of constant interruption.
ProSA found that when models are highly confident, they resist prompt rephrasing; low confidence causes major output swings. Larger models, few-shot examples, and objective tasks all correlate with higher confidence and greater robustness.
Local step-level confidence catches reasoning breakdowns that global averaging masks and enables early stopping before traces complete. This approach achieves comparable accuracy gains to naive majority voting with far fewer generated traces, proving trace quality matters more than quantity.
PAT, an agentic reviewer using test-time compute to check proofs and experiments line by line, achieves 34% better recall on math errors than zero-shot approaches and surfaced critical flaws at STOC and ICML that passed human review.
Show all 9 sources
Eight-module agentic evaluation achieved 0.27% judge shift versus 31% for LLM-as-a-Judge on complex tasks. However, the memory module cascaded errors, revealing that agentic systems need error isolation mechanisms to maintain gains.
Fixed seeds and zero temperature replicate the same output repeatedly, but that output remains one draw from the model's probability distribution. McDonald's omega testing across 100 repetitions reveals that consistency does not equal reliability.
aiXiv demonstrates that iterative review-refine cycles with automated retrieval-augmented evaluation and prompt-injection defenses measurably enhance proposal and paper quality, addressing the structural gap where AI-generated research lacks appropriate publication venues.
The paper designs a controlled comparison of isolated actions, rolling windows, known groups, and prospectively discovered episodes at equal review cost and false-alert workload, but the excerpt provides no empirical results showing whether added monitoring improves protection.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Towards Automating Scientific Review with Google's Paper Assistant Tool
- What Characterizes Effective Reasoning? Revisiting Length, Review, and Structure of CoT
- Understanding and Mitigating Premature Confidence for Better LLM Reasoning
- AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions
- AutoResearchClaw: Self-Reinforcing Autonomous Research with Human-AI Collaboration
- AI Agents Push Humans Out of the Loop
- GenAI as a Power Persuader: How Professionals Get Persuasion Bombed When They Attempt to Validate LLMs
- Deep Think with Confidence