Who Checks the Evidence When AI Helps Build AI? Cognitive Security in Model Development and Release

Frontier developers now use AI systems to help run the experiments, evaluations and analyses that inform the next model. One developer has described agents that control a share of the compute used to build its next model (Fry, 2026). This paper asks whether a release reviewer can independently establish whether the available evidence supports a specified safety claim for a defined deployment, when AI helped produce, select or interpret that evidence. It treats the release reviewer as a receiver of evidence in the sense of Bayesian persuasion, and applies three earlier constructs to the release decision: the monitor as receiver and synthetic assurance (Gardner, 2026d), epistemic supply-chain security (Gardner, 2026e) and the Cognitive Control Gap (Gardner, 2026b). The paper describes the release decision as a judgement made under delegated evidence and classifies release evidence by who designed, generated, recorded and can challenge it. It identifies three routes to a confident and mistaken release decision: an evaluation void, in which failed, excluded or unrun tests leave no mark on the summary; a shared origin, in which apparently independent checks draw on the same AI-processed material; and a control gap, in which a reviewer holds the authority to defer a release and lacks the means to use it. The paper states six falsifiable propositions and proposes a controlled study in which reviewers apply the same written release criterion, with the same authority to approve, defer or decline. All reviewers receive the same AI-generated summary, and some also receive the original run records and test plan. The study measures detection of material omissions, the correctness of the release decision, the quality of its justification and time taken. The argument is conceptual, and its claims are graded by the evidence available.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-21
DOI
https://doi.org/10.5281/zenodo.24424147
Primary Topic
Ethics and Social Impacts of AI
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Who Checks the Evidence When AI Helps Build AI? Cognitive Security in Model Development and Release

Sarah Jane Gardner
Zenodo (CERN European Organization for Nuclear Research)
Ethics and Social Impacts of AI
preprint

Who Checks the Evidence When AI Helps Build AI? Cognitive Security in Model Development and Release

Sarah Jane Gardner
preprint en

Abstract

Frontier developers now use AI systems to help run the experiments, evaluations and analyses that inform the next model. One developer has described agents that control a share of the compute used to build its next model (Fry, 2026). This paper asks whether a release reviewer can independently establish whether the available evidence supports a specified safety claim for a defined deployment, when AI helped produce, select or interpret that evidence. It treats the release reviewer as a receiver of evidence in the sense of Bayesian persuasion, and applies three earlier constructs to the release decision: the monitor as receiver and synthetic assurance (Gardner, 2026d), epistemic supply-chain security (Gardner, 2026e) and the Cognitive Control Gap (Gardner, 2026b). The paper describes the release decision as a judgement made under delegated evidence and classifies release evidence by who designed, generated, recorded and can challenge it. It identifies three routes to a confident and mistaken release decision: an evaluation void, in which failed, excluded or unrun tests leave no mark on the summary; a shared origin, in which apparently independent checks draw on the same AI-processed material; and a control gap, in which a reviewer holds the authority to defer a release and lacks the means to use it. The paper states six falsifiable propositions and proposes a controlled study in which reviewers apply the same written release criterion, with the same authority to approve, defer or decline. All reviewers receive the same AI-generated summary, and some also receive the original run records and test plan. The study measures detection of material omissions, the correctness of the release decision, the quality of its justification and time taken. The argument is conceptual, and its claims are graded by the evidence available.

Zenodo (CERN European Organization for Nuclear Research)
Ethics and Social Impacts of AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon