What it is
An organization that lets an AI agent act on its own can put a second reader between the agent and the work. The second reader may be a person or a separate review step. Its job is to compare the agent's output with the case before anyone relies on it. The randomized study cited below found that people catch surface errors far more often than errors that need conceptual judgment. A checker therefore needs the skill to find a mistake in the reasoning as well as in the wording. In the simulation runs cited below, a verifier on the agent removed several times more of the lasting harm than an upgrade to the model alone.
What it pushes on in the Lab
In the Lab, this lever weakens the pathway by which people adopt the automated system's failed output. It also raises operator deference drift, the slide from checking the system's output to accepting it by default. Output that arrives already checked invites reliance, so the Lab marks that second effect as a side effect.
In the modes that offer aiming, you can aim this lever at particular parts and pathways of a network. Otherwise it applies to the whole network.
Its pattern in the Practice Library
The Practice Library describes the pattern behind this lever:Put a verifier on the agent
The pressures it answers
These pressures list this lever among the levers that answer them:
A lever answers a pressure when it pushes the other way on something the pressure pushes on.
Where you can pull it
Networks in the Lab that offer this lever:7
The evidence behind its effects
The Lab cites these claims from the evidence registry for this lever's effects.
In the sociotechnical simulation, over a supervised-plus-agent scenario, adding a verifier to the autonomous agent removed roughly 46% of the harm that persists and a coordinated governance package roughly 43%, while upgrading the model alone removed only about 6%.[sim]
In a randomized study (N=2,784) with objective ground truth, humans accepted incorrect AI suggestions about a third of the time, and their rate of catching AI errors was governed by verification effort, prior trust in AI, and error legibility — surface errors were caught ~82% of the time versus ~31% for errors requiring conceptual judgment — not by financial incentives or time spent.[†]
Beck, J., Eckman, S., Kern, C., & Kreuter, F. (2026). Bias in the Loop: How Humans Evaluate AI-Generated Suggestions. Harvard Data Science Review, 8(2). https://hdsr.mitpress.mit.edu/pub/nrcn4h7d/release/1
https://hdsr.mitpress.mit.edu/pub/nrcn4h7d/release/1
Appears in: PAN framework development
Topics: algorithmic-fairness, human-ai-interaction