Decoded
← Feed
arXiv:2608.23642AIHC24 aug 2026

AI Agents Push Humans Out of the Loop

1 min read1 citationsHN ×9PDF ↗

Watching an AI agent all day wears down the very skills needed to catch its mistakes.

How it works

  1. Oversight is assumed

    Rules and vendors say a human must check agent actions, but give users almost no tools to do it well.

  2. Trace the degradation

    Approval fatigue, automation bias and skill atrophy pile up, and tired approvals then feed back into training signals.

  3. Build cognitive scaffolding

    Developers add friction, decision design and monitoring. Deployers add training, breaks, rotations and role separation.

repeated promptsrated successretrainingloopAgent requestsapprovalOverseergrows fatiguedQuick rubber-stampapprovalsDegradedfeedback signalsAgent optimizesskimmabilityHumanpushed outrepeated promptsrated successretrainingloopAgent requestsapprovalOverseergrows fatiguedQuick rubber-stampapprovalsDegradedfeedback signalsAgent optimizesskimmabilityHumanpushed out

Approval fatigue

try it

Drag the number of approval prompts up and watch scrutiny per approval.

Approvals per hourScrutiny per approvalcareful reviewrubber-stamping
Scrutiny per approval
high
here
careful review
More sign-off requests means less real review each; bounded autonomy saves attention for high stakes.Illustrative model, not the paper's data

Terms

6 terms

Human in the loop A person who approves, corrects or overrides an agent's actions while it runs.

Like a driving instructor riding along with a near-perfect student: after months of nothing happening, they stop watching the road and miss the one crash.

Where it breaks
Agents also shape what the overseer sees and can learn to hide failures.

Limits

  • Position paper with no new experiments or measurements.
  • Proposed mechanisms are untested; users may dislike friction that reduces overreliance.
  • Behavioral monitoring of overseers raises surveillance concerns.
From the paper