After training, analysts' free-text detection reasoning shifted away from physics-violation and anatomical cues toward functional and stylistic artifacts and became more specific¶

Claim · All claims
Evidence · 1 study (2 entries) · causal r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? References to physics violations in analysts' comments decreased from 20.0% pre-training to 7.0% post-training, the only between-phase difference with non-overlapping 95% Wilson confidence intervals. → Negar Kamali 2026 q2 i? High-specificity comments naming a concrete cue rose from 33.2% pre-training to 41.7% post-training (+8.5 percentage points). → Negar Kamali 2026 (2)

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

q2 · i? · causal · r2

Exploratory LLM-assisted coding of 365 free-text comments from 24 analysts along the five-category artifact taxonomy; 250 comments pre-training and 115 post-training. The article reports "the only between-phase difference for which the 95% Wilson confidence intervals do not overlap".

"References to physics violationsdecreasedfrom 20.0% (95% CI: [15.5, 25.4]) pre- training to 7.0% (95% CI: [3.6, 13.1]) post-training, the only between-phase difference for which the 95% Wilson confi- dence intervals do not overlap."

Negar Kamali 2026 (2)¶

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

q2 · i? · causal · r2

Same exploratory coding of free-text detection reasoning. The article reports high-specificity comments "rose from 33.2% pre-training to 41.7% post- training (+8.5 percentage points)", with post-training comments citing concepts taught in the session.

"High-specificity comments, those naming a concrete cue instead of a gen- eral impression, rose from 33.2% pre-training to 41.7% post- training (+8.5 percentage points)."

Discussion¶

Related Claims¶

• Training improved discrimination sensitivity (d′) with only a small, non-significant criterion shift toward labeling images as AI-generated — related
• Word-frequency changes in reflections suggest a shift away from given-information-based strategies — related
• A 30-minute generative AI literacy training increased intelligence analysts' overall accuracy at distinguishing real from AI-generated images by 9 percentage points — a broader claim this one bears on
• Training shows small trends toward better conflict detection on bat-and-ball and base-rate tasks but the opposite trend on the conjunction fallacy task — related