Blog · arXiv Analysis · Last reviewed July 10, 2026

The Linkage Score Becomes the Analyst Record

A July 2026 arXiv paper studies an AI-enabled crime-linkage tool inside a UK law-enforcement setting. Its useful lesson is not that the score should replace judgment. It is that high-stakes AI must be designed around the evidence analysts use to challenge the score.

Published: July 10, 2026 Modified: July 10, 2026 Last reviewed: July 10, 2026 Tags: arXiv, policing, surveillance, decision support, governance

The Paper

The paper is Jessica Woodhams, Amy Burrell, Wanyin Li, Fahim Ahmed, Matthew Tonkin, Jan Lemeire, Arkady Konovalov, Steven Frisson, Mark Webb, Sarah Galambos, Vesna Nowack, and Dalal Alrajeh's How Analysts Use AI in High-Stakes Crime Linkage: An Industrial Study, arXiv:2607.08274 [cs.HC, cs.SE]. The arXiv API lists version 1 as submitted on July 9, 2026, with the comment that it is a 12-page paper with 6 figures for FSE Industry. The PDF metadata confirms 12 pages, and the arXiv HTML lists a CC BY 4.0 license.

The paper belongs beside this site's work on real-time crime-center dashboards, police-report model memory, human oversight, AI audit trails, and explanation rights. Its fresh angle is how expert analysts use an AI score in crime linkage, not general prediction or patrol allocation.

What Was Studied

Crime linkage is the practice of identifying whether multiple offences may have been committed by the same person. Specialist analysts compare behavioral and situational information across large crime databases, a task the authors describe as time-consuming, cognitively demanding, and sometimes involving repeated exposure to disturbing material. The tool under evaluation is LATIS, formerly DST, an AI-enabled crime-linkage decision-support tool developed for the National Crime Agency's Serious Crime Analysis Section, or SCAS.

LATIS does not make the final linkage decision. It returns ranked candidate offences for a user-specified index offence, using machine-learning models trained on real crime data, and presents predictions alongside model features and non-AI behavioral evidence. The study used operationally realistic tasks on a secure site with sensitive real crime data. Six participants took part: three analysts and three senior analysts. The evaluation produced 16 experimental sessions because four participants completed all three cases while two completed the first two cases only.

The methods matter. The authors combined direct observation, feedback, eye-tracking, mouse-tracking, and post-session surveys. Sessions typically lasted 75 to 120 minutes. Analysts and researchers were blind to the linked offence in the top-20 ranked list. This is still a one-team study, but it is not a toy benchmark detached from work practice.

How Analysts Used It

The headline finding is disciplined partial trust. Analysts did not accept the AI-ranked list as a verdict. They used predictions selectively, attended to feature-based explanations, and checked candidate links against traditional behavioral evidence. Eye-tracking showed attention to the case identifier, modus-operandi similarity, and geographical proximity. The interface became a navigation aid for existing analytic work, not a replacement for it.

The paper also reports positive reactions to seeing model features beside probability scores, while analysts wanted better filtering, progress marking, and feature removal during review. The usability lesson is that explanation is not just a paragraph saying why the model thinks something. In this setting, explanation is work surface: it has to support scanning, retrieval, comparison, doubt, and memory across a session.

Why the Matrix Matters

The strongest governance signal is the behavioral matrix, the non-AI comparison surface analysts used to inspect candidate cases against the index offence. The paper reports that the matrix was opened at least ten times in 12 of 16 sessions. In 260 of 279 matrix openings, participants selected only one case, suggesting repeated targeted checks rather than broad batch acceptance. Average matrix-open time per participant was 2.52 minutes in session 1, 1.67 minutes in session 2, and 1.59 minutes in session 3.

That matrix is not decorative transparency. It is where the score becomes accountable to case details. Without it, the analyst would see rank and probability but lose the friction that turns a candidate into something contestable. In high-stakes decision support, the auxiliary evidence surface may be the safety feature.

Governance Reading

The authors explicitly frame the system as decision support rather than automated decision-making. They state that final decisions remain with human analysts, and they discuss bias, transparency, and accountability as ethical concerns. They also report that prior work assessed training data for demographic representativeness and did not identify evidence of bias. That sentence should not be inflated into a general claim that the system is unbiased. It is evidence about an assessment in this tool's development path, not a guarantee for every deployment, dataset, or jurisdiction.

The paper is candid about limits. A radar-plot component had incorrect computed data, so findings about it were excluded. The study involved one law-enforcement team and a tool designed for that team. Questionnaire scores may also reflect learning bias, because sessions were completed in the same order and the first task was more complex.

For Spiralist purposes, the lesson is that high-stakes AI governance cannot stop at model accuracy. It needs a record of how the score was used, what non-AI evidence was consulted, which visualization failed, what the analyst accepted or rejected, and who owned the final judgment.

The Receipt

A crime-linkage AI receipt should name the index offence, candidate list, model version, training-data boundary, probability score, model features shown, non-AI behavioral matrix checks, analyst notes, rejected candidates, accepted candidates, visualization defects, bias assessment, explanation surface, database searches outside the tool, reviewer, final decision owner, and downstream investigative use limit.

The practical rule: a linkage score is not operational evidence by itself. It becomes part of a responsible record only when the analyst's validation path is visible enough to inspect, challenge, and repair.

Sources

Return to Blog