Risk Evaluation · Method

Item Response Theory for AI Safety

NUWA research note · independent review pending

First published: 2026-08-05Source checked: 2026-10-25

NUWA Research Note

Research question Design and assumptions Source findings NUWA interpretation Limits and contrary evidence Research and decision relevance Evidence update Institutions and roles Original resources

Model-assisted reading · Selected sections of the original full text or official article; a complete section-by-section audit remains pending.

At a glance

Headline cohort size differs from complete-case and latent-analysis denominators.

Research question

Do safety benchmarks measure common latent traits across the same model population?

Design and assumptions

The study uses item response theory to analyse latent structure across safety benchmarks, with different complete-case populations at each stage: up to 192 headline models, 169 in the complete benchmark pool and 134 in latent analysis. Items are filtered from 5,255 to 5,067.

Source findings

The analysis reveals patterns of shared item response, but factor labels summarise loadings rather than independently validated real-risk constructs. Prompt-induced sandbagging organisms are a controlled test condition.

NUWA interpretation

Latent variables can expose item redundancy, difficulty and response structure, without becoming a universal risk score. Filtering and scoring failures change the population to which the results apply.

Limits and contrary evidence

Calibration distribution shift, judgment failures and prompt-induced organisms limit extrapolation to concealed strategic behaviour. Do not use 192 as the latent-analysis denominator.

Research and decision relevance

NUWA recommendation: Update population counts and construct-validity notes on the existing record; inspect held-out calibration before using IRT for allocation decisions.

Precise source locations

Item Response Theory for AI Safety · Methods; Table1; results; limitations; text lines 00142,00196Original resource captured 2026-10-28T07:25:52.921099+00:00; sha256:b73cdc52885df6e9c0e26f0fcc68480d3e3222ed79c4b6c6106ec2e00203b492

Source checking, extraction, editorial approval and replication are recorded separately. Methodology →

Evidence update · 2026-10-28

Model-assisted reading · independent evidence and bilingual editorial review pending. Values are source findings or inventory counts, separate from the trend result series.

The original-source supplement distinguishes the headline 192 models, 169 complete benchmark models and 134 latent-analysis models, with filtering from 5,255 to 5,067 items. Factor names summarise loadings rather than validated universal risk constructs.

Source-bound observations (5)

Counts, rates and contrasts have different units. A sample count is not a number of independent experiments. Unknown numerators and interval levels remain unknown.

System / metric	Reported value	Population and conditions
Models in headline cohort	192 models	Conditions, limitations and source Different analysis populations; do not use192as factor-analysis denominator. Original source ↗ · Methods; Table1; text lines 00071,00088,00101,00142,00192,00460Original capture 2026-10-28T07:25:52.921099+00:00; sha256:b73cdc52885df6e9c0e26f0fcc68480d3e3222ed79c4b6c6106ec2e00203b492
Complete benchmark models	169 models	Conditions, limitations and source Different analysis populations; do not use192as factor-analysis denominator. Original source ↗ · Methods; Table1; text lines 00142,00238,00462Original capture 2026-10-28T07:25:52.921099+00:00; sha256:b73cdc52885df6e9c0e26f0fcc68480d3e3222ed79c4b6c6106ec2e00203b492
Latent-analysis models	134 models	Conditions, limitations and source Different analysis populations; do not use192as factor-analysis denominator. Original source ↗ · Methods; Table1; text lines 00142,00196Original capture 2026-10-28T07:25:52.921099+00:00; sha256:b73cdc52885df6e9c0e26f0fcc68480d3e3222ed79c4b6c6106ec2e00203b492
Initial item inventory	5255 items	Conditions, limitations and source Different analysis populations; do not use192as factor-analysis denominator. Original source ↗ · Methods; Table1; text lines 00088,00103,00162Original capture 2026-10-28T07:25:52.921099+00:00; sha256:b73cdc52885df6e9c0e26f0fcc68480d3e3222ed79c4b6c6106ec2e00203b492
Retained items	5067 items	Conditions, limitations and source Different analysis populations; do not use192as factor-analysis denominator. Original source ↗ · Methods; Table1; text lines 00140,00145Original capture 2026-10-28T07:25:52.921099+00:00; sha256:b73cdc52885df6e9c0e26f0fcc68480d3e3222ed79c4b6c6106ec2e00203b492

Sources for this interpretation

Item Response Theory for AI Safety · Methods; Table1; results; limitations; text lines 00142,00196Original resource captured 2026-10-28T07:25:52.921099+00:00; sha256:b73cdc52885df6e9c0e26f0fcc68480d3e3222ed79c4b6c6106ec2e00203b492

Download source-bound evidence · JSON ↓

Institutions and roles

Published by UK AI Security Institute

Authors (4)

Joshua Fonseca Rivera · Neil Shah · David Demitri Africa · Konstantinos Voudouris

Cite and reuse

Cite the original publication for its findings. Cite this NUWA record separately for its explanation and retained snapshot.

BibTeX ↓RIS ↓Record + source metadata · JSON ↓Reading coverage · JSON Suggest a correction

Original publication · BibTeX

@misc{item-response-safety,
  title = {Item Response Theory for AI Safety},
  url = {https://www.aisi.gov.uk/research/item-response-theory-for-ai-safety},
  author = {Joshua Fonseca Rivera and Neil Shah and David Demitri Africa and Konstantinos Voudouris},
  year = {2026},
  date = {2026-08-05},
  version = {Source capture 2026-10-25},
  note = {Source capture 2026-10-25}
}

NUWA record · BibTeX

@misc{item-response-safety-nuwa,
  title = {Item Response Theory for AI Safety — NUWA research record},
  url = {https://whitzard.tech/en/nuwa/whitzard-index/work/item-response-safety/},
  version = {2026-10-29.1},
  organization = {NUWA Frontier AI Safety Lab},
  howpublished = {Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--item-response-safety.json},
  note = {NUWA research note. Source-reading scope and independent review status are stated on the record. Snapshot 2026-10-29.1. Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--item-response-safety.json.}
}

Evidence snapshot 2026-10-29.1 · Index updates · RSS

References

[1] Joshua Fonseca Rivera, Neil Shah, et al.. Item Response Theory for AI Safety. 2026-08-05 ; Official publication masthead and article byline
[2] Joshua Fonseca Rivera, Neil Shah, et al.. Item Response Theory for AI Safety. 2026-08-05 ; Methods; Table1; results; limitations; text lines 00142,00196