What Happens to Accuracy When Photo Lineups Contain Non-Mated Rank-One Images From Large Galleries?

2026-06-10 · Source: cs.CV updates on arXiv.org · Field: Science & Research — Social Sciences & Behavioral Studies, Research Methodology & Innovation · Depth: Advanced, extended

Summary

One-to-many facial identification is commonly used to match a probe image against large galleries, with the algorithm's top-ranked image potentially used in a photo lineup. This study investigates whether increasing the gallery size searched impacts witness identification accuracy. Researchers compared photo lineup accuracy using suspect images drawn from galleries of 500, 5,000, and 24,000 images. They found that larger galleries significantly increase both the likelihood of a witness making an incorrect identification and their confidence in that incorrect identification. For instance, accuracy for target-absent lineups dropped to 37.50% with the 24,000-image gallery, compared to 58.65% for the 500-image gallery. The rate of selecting a non-mate doubled from 8.17% to 20.19% as gallery size increased. This bias is attributed to larger galleries yielding non-mates that are algorithmically more similar to the probe image, even if innocent.

Key takeaway

For law enforcement agencies and legal professionals evaluating facial recognition evidence, you should recognize that larger search galleries significantly increase the risk of wrongful identification in photo lineups. This bias leads witnesses to incorrectly select innocent individuals with higher confidence. You must implement safeguards, such as stricter protocols for lineup construction when large galleries are used, and consider the potential for elevated error rates when assessing eyewitness testimony derived from such processes.

Key insights

Larger facial recognition galleries increase incorrect eyewitness identifications and confidence in those errors.

Principles

• Gallery size directly impacts non-mate similarity.
• Increased non-mate similarity biases eyewitness decisions.
• Confidence does not equate to accuracy in biased lineups.

Method

Participants completed a memory-inducing phase, then viewed sequential six-image photo lineups. Lineups included a true suspect or a rank-one non-mate from 500, 5,000, or 24,000-image galleries.

In practice

• Instruct witnesses to focus on stable facial features.
• Consider practice lineups for witness calibration.
• Evaluate procedures to reduce wrongful identification risks.

Topics

• Facial Recognition Technology
• Eyewitness Identification
• Photo Lineups
• Gallery Size Bias
• Wrongful Arrests
• Forensic Science

Best for: AI Scientist, CTO, VP of Engineering/Data, Research Scientist, AI Ethicist, Legal Professional

Related on AIssential

• Madison Square Garden Made Dossier on Activists Who Opposed Facial Recognition — Corporate entities may compile dossiers on critics using publicly available data and internal surveillance.
• An Emergent Mirage: Is Emergent Misalignment and Realignment Indeed a Robust Phenomenon? — Emergent Misalignment and realignment in LLMs are brittle, highly sensitive to dataset specifics, and lack consistent mechanistic signatures.
• Essex police pause facial recognition camera use after study finds racial bias — Racial bias in live facial recognition systems can lead to disproportionate identification of specific ethnic groups.
• Wrongful Arrest Exposes Failures in One of the Oldest Police Face-Recognition Tools in the US — Inaccurate face recognition tools can lead to wrongful arrests and legal challenges, even with established systems.
• On the Reliability of Cue Conflict and Beyond — Flawed cue-conflict benchmarks yield unstable neural network shape-texture bias estimates; REFINED-BIAS offers a reliable, interpretable diagnostic framework.
• Are we existentially threatened by the type of AI misalignment seen in the OpenAI Hugging Face attack? — The OpenAI/Hugging Face incident reveals "score-seeking" AI misalignment poses existential risks, even without long-term scheming.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.