A study conducted by Carina Hahn and three of her colleagues at the National Institute of Standards and Technology (NIST) was published on October 21, 2026, in the journal PLOS One. The 27 participants in the experiment were specialists who perform facial comparison work for federal agencies in the United States or had received training for this exact purpose. In this profession, specialists decide whether two photographs belong to the same person, with their judgments utilized as investigative clues or occasionally heard as expert testimony in court.1

For this study, the research team measured which specific part of the face the specialists examined.

What is in this article?
  • Which of the nine facial regions did specialists choose to reveal by an overwhelming margin?
  • How did one of the least revealed regions yield the highest accuracy on its own?
  • What happened across three decisions made by looking solely at the chin?
  • In what ways does the experiment differ from real-world comparisons made using security camera frames?

How do experts use obscured photos with nine region buttons?

On the left side of the screen, photographs of two different people remained open. A third photograph in the middle belonged to one of these two individuals, but it was initially completely darkened. Using the buttons on the right, the specialist could reveal one of nine regions of the face: cheeks, chin, ears, eyes, eyebrows, forehead, neck, nose, and mouth. Each revealed region stayed on the screen, and the final decision could be made at any desired moment.1

The design rewarded making decisions with minimal inspection. Everyone started with 100 points; a correct answer awarded 100 points, while a wrong answer deducted 200 points. Revealing the first region was free, and each subsequent region cost 30 points. The photographs were obtained from the University of Notre Dame's facial archive, for which the pictured individuals had given consent. Forty triplets where people most frequently erred in a previous test were selected, intentionally placing visually similar faces before the experts.

A subset of the participants chose which regions to reveal themselves. In the other group, the presentation order was predetermined. Non-expert volunteers had previously scored the two target faces region by region; in some trials, the most similar region was revealed first, while in others, the most dissimilar region appeared first.

Why do three-quarters of expert decisions rely on a single feature?

Specialists who chose the regions themselves made 76 percent of their decisions by looking at a single region, and 17 percent by looking at two regions. In trials where the order was predetermined, the picture changed. When the most similar region was shown first, specialists tended to request a second piece, and the share of decisions made with a single region dropped to 35 percent.1

Accuracy remained high across all three conditions. In trials with a predetermined order, the success rate predicted by the statistical model ranged between 91.8 and 98.3 percent. Opening more regions did not increase accuracy. In the self-selection group, predicted success dropped to 76.9 percent in decisions where three regions were opened, though this drop was not statistically significant after corrections for multiple comparisons. According to the authors, instead of combining all the parts on the screen, the specialist may be relying primarily on the last part they opened, noting that the checklists and notebooks they use in actual casework were absent from the experiment.

Notes and Sources

  1. Hahn, C. A., Phillips, P. J., Grantham, J. D., Yates, A. N. (October 21, 2026). "Facial feature prioritization by forensic specialists". PLOS One 21(10): e0359520. Supplementary file S1, Tables S1 and S2. doi.org/10.1371/journal.pone.0359520 ↩
  2. PLOS (October 21, 2026). "Face specialists need only one or two facial features to identify individuals". Press release with statements by Carina Hahn. eurekalert.org ↩
  3. Bertillon, A. (c. 1909). "Tableau synoptique des traits physionomiques: pour servir à l'étude du portrait parlé". Gelatin silver print, Metropolitan Museum of Art. metmuseum.org ↩
  4. Phillips, P. J., Yates, A. N., Hu, Y. et al. (2018). "Face recognition accuracy of forensic examiners, superrecognizers, and face recognition algorithms". PNAS 115(24): 6171-6176. doi.org/10.1073/pnas.1721355115 ↩