{"resourceId":"nist-fbi-plos-one-forensic-face-specialists-feature-prioritization-ears-20261021","versions":[{"version":"external-414c2febd6a6f9cb4cc3f2353cb1327d038fb3919f812e862056f2af934452ab","resource":{"id":"nist-fbi-plos-one-forensic-face-specialists-feature-prioritization-ears-20261021","title":"NIST/FBI peer-reviewed study: forensic face specialists often identify correctly from one or two features, favour ears, and rarely choose the forehead despite similar accuracy","organization":"National Institute of Standards and Technology (NIST), with FBI CJIS Division participation","sector":"Forensic facial comparison / human review of face identification","geography":"United States (federal agency participants)","publishedAt":"PLOS One, Oct. 21, 2026 (NIST blog post Oct. 28, 2026)","publicationDate":"2026-10-21","eventDate":null,"sourceName":"PLOS One","sourceLabel":"Facial feature prioritization by forensic specialists (Hahn, Phillips, Grantham, Yates)","sourceUrl":"https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0359520","evidenceClass":"academic-research","outcomeClass":"emerging","topics":["knowledge-work","accessibility-workforce","governance-procurement"],"finding":"Hahn, Phillips, Grantham and Yates (PLOS One, Oct. 21, 2026) tested 27 face specialists from federal agencies, recruited October 2023 to July 2024, on 24 trials each (648 trials). Participants matched an occluded probe to one of two identities by revealing isolated features, either self-selected or in a predetermined order based on similarity ratings from 34 benchmark raters at NIST and FBI CJIS. Specialists often decided correctly after one or two features; 396 of 648 trials used a single feature. Ears were strongly preferred, with probability of success 0.93 (95% CI lower bound 0.88); foreheads reached 0.94 but with a lower bound of 0.71 and were rarely chosen. Specialists adapted choices to feature similarity. The authors state results do not reflect real-world accuracy, use 40 difficult triads mainly of Caucasian and East Asian faces, and that comparisons such as security-camera stills against mugshots still need testing. The work was supported by a NIST-FBI interagency agreement, and the FBI gave feedback on study design.","sledRelevance":"Human examiner review is the main safeguard U.S. agencies cite after a facial recognition search; this study gives evidence on how trained reviewers actually decide, and on its limits.","evidence":"Opened and read the PLOS One article (abstract, methods, results, discussion and limitations); NIST’s Oct. 28 blog post and an Oct. 30 Biometric Update summary were also read. Peer reviewed per the journal record (received Feb. 4, accepted Sept. 15, 2026).","architectureImplications":"Findings concern human comparison after algorithmic search; the authors suggest specialist-chosen features could inform algorithm training, a proposal not tested here.","governanceImplications":"Supports examiner training and procedure standards; the authors note FISWG guidance calls for full-face morphological analysis, which single-feature shortcuts do not replace.","securityPrivacyImplications":"Not addressed; relevant indirectly because over-confident human confirmation can turn an algorithmic candidate into an arrest.","caveats":"Laboratory task on high-quality, demographically narrow images; does not measure real-world accuracy or test security-camera stills against mugshots; participants’ professional roles and experience were not recorded; supported by a NIST-FBI interagency agreement; findings concern human specialists, not algorithms. Published Oct. 21 but publicised by NIST Oct. 28 and in trade press Oct. 30, which is why it appears now.","streamIds":["public-safety"],"roles":{"sales":"Sales framing: agencies defend facial recognition by pointing to human review, and this NIST study shows trained specialists can be accurate yet overlook useful features. For police and prosecutors, the buyer problem is whether their reviewers are trained and tested. A bounded engagement could assess an agency’s examiner training and proficiency testing against FISWG guidance. The evidence cannot support claims about real-world accuracy; it is a laboratory task with high-quality images and a narrow demographic range.","engineering":"Engineering framing: the study concerns the human step after an algorithm returns candidates. Pre-sales should ask whether review tools support structured, feature-by-feature comparison, record which features drove a decision and separate trained examiners from investigators. A proof of value would add structured comparison notes to the review workflow and test them on a proficiency set. Constraint: the study did not test security-camera stills against mugshots, a common real-world case, so local validation is needed.","delivery":"Delivery framing: the work is to formalise human review: define who may confirm a facial recognition lead, require documented feature-level comparison and run periodic proficiency tests. The operational owner is the agency’s forensic or identification unit. Acceptance criteria are proposals: every confirmation done by a trained examiner, comparison notes retained with the case, and annual proficiency results reviewed. Main risk: treating any officer’s glance as the human safeguard the policy promises."},"roleDetails":{"sales":{"customerProblem":"Agencies rely on human review to check facial recognition leads without evidence that reviewers are trained or tested.","stakeholders":["Forensic and identification units","Police leadership","Prosecutors","Defense bar"],"discoveryQuestions":["Who confirms facial recognition candidates, and with what training?","Are feature-level comparison notes recorded?","Are reviewers proficiency-tested?"],"valueHypothesis":"Structured, tested human review could make the stated safeguard auditable.","potentialEngagement":"Assessment of examiner training and proficiency testing against FISWG guidance.","unsupportedClaims":["Real-world accuracy figures","That single-feature checks are sufficient","Algorithm performance claims"]},"engineering":{"fit":"Relevant to any agency using facial recognition with human confirmation.","architectureIntegration":"Review tools after algorithmic search; documentation of feature comparisons in case records.","prerequisites":["Trained examiner roster","Structured comparison template","Proficiency test set"],"constraints":["Laboratory images only","Narrow demographics","Security-camera stills not tested"],"security":"Retain review notes as discoverable records with access controls.","proofOfValue":["Pilot structured notes and test reviewers on a local proficiency set including CCTV stills."]},"delivery":{"work":"Define confirmation roles, require structured notes, run proficiency tests.","dependencies":["Training budget","Case records integration","Prosecutor consultation"],"ownership":"Agency forensic or identification unit","skillsAdoption":"Train examiners on diagnostic features, including under-used ones.","governanceCheckpoints":["Policy approval","Prosecutor and defense disclosure review","Annual proficiency review"],"acceptanceCriteria":["100% of confirmations by trained examiners","Comparison notes stored with each case","Annual proficiency results reviewed by command"],"risks":["Over-reliance on single features","Untrained officer confirmations","Lab-to-field gap"]}},"retrievedAt":"2026-10-31T02:56:14Z","enrichedAt":"2026-10-31T03:04:14Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Points to training on under-used diagnostic features such as foreheads; professional roles and experience of participants were not recorded.","procurementImplications":"None directly; agencies buying facial recognition should budget for examiner training and proficiency testing.","operatingModelImplications":"Human review quality depends on trained specialists using structured methods, not ad hoc officer confirmation.","peerReview":"peer-reviewed","sourceVerification":{"openedUrl":"https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0359520","referenceExcerpt":"Results indicate that specialists are highly efficient, frequently achieving accurate identifications while viewing only one or two facial features.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}