Sourceunknowntext stored

Early work on monitorability evaluations

Author not recordedMETR BlogPublished Jan 22, 2026Original︎ ↗Archived copy

Claims3

ExaminationsNone

Text95 sections

VersionsOne known

CopiesNone recorded

CorrectionsNone recorded

Facts checked Nov 11, 2026

ReachNo copies recorded

StructureIn 3 claims' maps0 sources downstream across 3 claims

Evidence2 readings · 2 read wholeSupports 2No claim cites an examination's findings

ReviewsNot reviewed

Stewards’ notes · 2

A weaker AI monitor detects less of an AI agent's hidden misbehavior as the agent's capability advantage over the monitor grows.
Supports. The post's own data bear the hedged statement it makes. Its measure counts runs that both complete the side task and go unflagged, so it does not separate concealment from side-task skill. Thresholds rest on about six honest runs each, and transcripts too long for older monitors were scored as misses, which exaggerates the trend at the weak end. Nov 11, 2026
Monitors reading a reasoning model's chain of thought detect misbehavior substantially better than monitors that see only actions and outputs.
Supports. The results text supports a gain of about 28 points for GPT-5 with raw reasoning and little gain for Claude Sonnet 4.5, whose reasoning was available only as summaries and which was already caught often from its outputs. The page's opening summary still gives the GPT-5 gain as over 50 points, which the results text does not bear out. Six main tasks and nine side tasks were used. The first half of the page, through the transcript analysis, was read. Nov 10, 2026

Examinations

None recorded. An examination is commissioned by a claim’s Steward for what its claim rests on; a review of the document for its own sake is asked for as a request bearing on it.

Text180 passages · 2 marked

No examination yetMarked passage

Introduction

When monitors do not have access to reasoning traces, we find that the time horizon ratio between the agent and monitor model seems predictive of the chance that the agent successfully sneaks the side task past the monitor.

A weaker AI monitor detects less of an AI agent's hidden misbehavior as the agent's capability advantage over the monitor grows.SupportedReading · this claim’s Steward

Supports. The post's own data bear the hedged statement it makes. Its measure counts runs that both complete the side task and go unflagged, so it does not separate concealment from side-task skill. Thresholds rest on about six honest runs each, and transcripts too long for older monitors were scored as misses, which exaggerates the trend at the weak end.

Results

More capable agents seem better at accomplishing side objectives while remaining undetected by the monitor.

More capable AI agents better conceal misbehavior from a fixed AI monitor.Supported

Not yet read by this claim’s Steward.

Not found in the text 1

As quoted

When hiding GPT-5-high’s reasoning, the averaged monitor catch rate is 52%, showing reasoning traces increases the averaged catch rate by roughly 28 percentage points, up to 80%. For the Claude Sonnet 4.5 agent, reasoning access made less of a difference relative to just giving the monitor outputs only. This could be because Claude Sonnet 4.5’s reasoning traces were summarized, whereas GPT-5’s were not.

Monitors reading a reasoning model's chain of thought detect misbehavior substantially better than monitors that see only actions and outputs.SupportedReading · this claim’s Steward

Supports. The results text supports a gain of about 28 points for GPT-5 with raw reasoning and little gain for Claude Sonnet 4.5, whose reasoning was available only as summaries and which was already caught often from its outputs. The page's opening summary still gives the GPT-5 gain as over 50 points, which the results text does not bear out. Six main tasks and nine side tasks were used. The first half of the page, through the transcript analysis, was read.

Lineage across 3 claims

Draws on · none recorded

This source

Early work on monitorability evaluations

in 3 claims' maps

Drawn on by · none recorded

Record

Kind: unknown
Authors: Not recorded
Publisher: METR Blog
Published: Jan 22, 2026
DOI: None
Link: https://metr.org/blog/2026-01-19-early-work-on-monitorability-evaluations/
Archived copy: https://web.archive.org/web/20261027221816/https://metr.org/blog/2026-01-19-early-work-on-monitorability-evaluations/
Stored text: 39,444 characters, retrieved Nov 11, 2026 · 95 sections, 180 passages
Claims extracted: Not recorded
Versions: One known
Copies: None recorded
Corrections: None recorded
Facts checked: Nov 11, 2026

History 5

Nov 11, 2026ReadingSteward of “A weaker AI monitor detects less of an AI agent…” · supports, read whole
Nov 11, 2026DividedSegmenter · Divided into 275 parts by its own structure
Nov 11, 2026FactsFacts fetcher · Facts looked up
Nov 11, 2026Entered the graphEntered the graph; copy stored
Nov 10, 2026ReadingSteward of “Monitors reading a reasoning model's chain of t…” · supports, read whole