Training improved discrimination sensitivity (d′) with only a small, non-significant criterion shift toward labeling images as AI-generated¶
Claim · All claims
Evidence · 1 study · causalr2·q3
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person
Subclaims¶
Reading the evidence codes
q is the design tier, i the impact, n the sample, and the kind with its r
the rigour judged within that kind: what each code means.
This claim: no standardized effect is recorded in observations/ for it yet, so no benchmark reading is possible.
q3 i? Signal detection analysis showed training produced a substantial improvement in discrimination sensitivity (∆d′ = +0.36, p = 0.003). → Negar Kamali 2026
q3 i? The criterion shift toward classifying images as AI-generated was small and non-significant (∆c = −0.12, p = 0.230). → Negar Kamali 2026
Evidence¶
Negar Kamali 2026¶
Negar Kamali, Candice Rockell Gerstner, Jessica Hullman, Matthew Groh. (2026). Generative AI Literacy Training Improves Intelligence Analysts' Discrimination of Real and AI-Generated Images. https://arxiv.org/abs/2606.28510 doi:10.48550/arxiv.2606.28510
q3 · i? · causal · r2
Signal detection analysis aggregating each participant's hit and false-alarm rates from raw pre- and post-training trials, averaged across the AB and BA counterbalance orderings so set-difficulty differences cancel. The article reports "a substantial improvement in discrimination sensitivity" while the criterion barely moved.
"Training produced a substantial improvement in discrimi- nation sensitivity (∆d′ = +0.36,p= 0.003) alongside a small, non-significant criterion shift toward classifying im- ages as AI-generated (∆c=−0.12,p= 0.230)."
Discussion¶
Related Claims¶
- A 30-minute generative AI literacy training increased intelligence analysts' overall accuracy at distinguishing real from AI-generated images by 9 percentage points — related
- The training gain was driven by a 14.2-percentage-point improvement in accuracy on real images, while the gain on AI-generated images was not statistically significant — related
- After training, analysts' free-text detection reasoning shifted away from physics-violation and anatomical cues toward functional and stylistic artifacts and became more specific — related
- Training effects concentrated in portrait images (+39.8 percentage points, significant), with directional non-significant negative estimates for full-body and posed-group images — related
- Training gains were distributed broadly across the 97 stimulus images rather than concentrated in a small subset, though stimulus-level effects were heterogeneous — related
Page metadata
| Field | Value |
|---|---|
type |
claim |
id |
training-improves-discrimination-sensitivity-not-response-bias |
title |
Training improved discrimination sensitivity (d′) with only a small, non-significant criterion shift toward labeling images as AI-generated |
description |
Training improved discrimination sensitivity (d′) with only a small, non-significant criterion shift toward labeling images as AI-generated |
status |
review |
evidence_strength |
moderate |
newest_evidence |
2026 |
generated |
by process:wiki-ingest on 2026-10-10 |
trust |
unverified — no one has confirmed this page's content |
sources |
1 — negar-kamali-2026 |
Frontmatter is the page's machine-readable record — see the Schema & Guide.