The Forensic Science Regulator and the quality standards framework
Forensic science in England and Wales operates under the oversight of the Forensic Science Regulator, a role that existed on a non-statutory basis from 2008 and was placed on a statutory footing by the Forensic Science Regulator and Biometric Strategy Act 2021. The significance of that change from advisory to statutory is not merely symbolic: the Regulator can now issue legally binding Codes of Practice and Conduct, and has the power to compel compliance from providers and processes used in criminal proceedings where those standards are not met. Any AI tool used to generate evidence for use in court falls within scope.
The Regulator's Codes require forensic science processes to be validated — that is, demonstrated to be fit for the specific purpose for which they are used — and to meet the requirements of ISO/IEC 17025, the international standard for testing and calibration laboratories. Validation is a demanding requirement even for well-established techniques. For AI systems, which may produce outputs that are statistically accurate in aggregate but unreliable in individual cases, or which may perform differently across different demographic groups or environmental conditions, the validation requirement raises questions that do not have settled answers. The Regulator has published guidance on the validation of digital forensics tools and has flagged concerns about the pace at which AI-enabled tools are being adopted into casework before adequate validation frameworks are in place.
DNA analysis and probabilistic genotyping
The most extensively used AI application in forensic science is arguably probabilistic genotyping — the use of statistical software to interpret complex DNA mixtures where the profiles of two or more individuals are present in the same sample. Traditional binary interpretation of DNA mixtures — either a profile is included or it is not — was recognised as having significant limitations when samples are degraded, when there are more than two contributors, or when the relative proportions of different individuals' DNA in a mixture are unequal. Probabilistic genotyping software such as STRmix, developed in New Zealand and used by the Forensic Science Service's successor laboratories and by several national police forces, produces a likelihood ratio that expresses how much more probable the evidence is if the suspect contributed to the mixture than if they did not.
The introduction of probabilistic genotyping has been broadly welcomed by forensic scientists as improving the accuracy and objectivity of mixture interpretation. It has also, however, given rise to a series of contentious appeal cases. The software's underlying source code has been the subject of disclosure disputes in both the United States and Australia, where defence experts sought access to examine the algorithm's methodology and were resisted on the grounds of commercial confidentiality. In the UK context, the principle that defendants are entitled to sufficient information about expert evidence to mount an effective challenge is well-established, but what that means in practice when the evidence is generated by a proprietary algorithm whose full workings are not publicly available remains a live and unresolved tension.
A separate concern relates to the validation of probabilistic genotyping software for the specific mixture types to which it is applied in casework. Software validated on two-person mixtures may not perform equivalently when applied to three- or four-person mixtures, and validation data generated under laboratory conditions may not fully capture the variability of real casework samples. The Forensic Science Regulator has flagged these as areas requiring ongoing attention.
Automated fingerprint identification
Automated fingerprint identification systems — AFIS — have been used by UK law enforcement since the 1980s, making them among the earliest AI applications in policing. The National Automated Fingerprint Identification System (NAFIS), operated by the National Policing Improvement Agency's successor functions, holds millions of ten-print records from individuals who have been arrested or charged, as well as a substantial database of scene-of-crime marks. The system generates a ranked list of candidate matches that a human fingerprint examiner then reviews and evaluates against the ACE-V methodology — Analysis, Comparison, Evaluation, and Verification.
The human review step is not merely procedural: fingerprint examination retains significant subjective elements even when AI is used to generate candidate matches, and the literature on cognitive bias in fingerprint examination is substantial and troubling. Studies have shown that fingerprint examiners presented with contextual information suggesting a suspect is guilty are more likely to reach a positive identification than when presented with identical prints without that context. The AI candidate list can inadvertently create its own form of contextual bias by presenting only the highest-scoring candidates for human review, effectively narrowing the examiner's frame of reference before they have looked at the evidence.
More recent developments include the application of convolutional neural networks to fingerprint enhancement — improving the quality of degraded or partial marks before they are submitted for comparison — and to the identification of partial palm prints, which traditional AFIS systems handled less reliably. These applications are in varying stages of operational deployment in UK forces.
Digital forensics and AI triage
The volume of digital evidence in modern criminal investigations has grown to a scale that is genuinely unmanageable without automated assistance. A single smartphone seized during an investigation may contain hundreds of thousands of messages, images, and data records. A single laptop may hold terabytes of data. Forces across England and Wales report significant backlogs in the examination of digital devices, with devices sometimes sitting unexamined for months or years while suspects await trial.
AI-assisted triage tools — software products such as those developed by Cellebrite and MSAB — use machine learning to classify and prioritise data extracted from devices, flagging items likely to be of investigative relevance for priority examination by a human analyst. Child sexual abuse material detection, in particular, has seen widespread adoption of AI-assisted hashing and classification tools: Photodna, developed by Microsoft and operated in partnership with the Internet Watch Foundation, compares digital fingerprints of images against a database of known illegal material and is used by forces and major technology platforms alike. The tool does not by itself produce evidence for court — a human analyst must verify any match — but it enables investigators to identify material that would be physically impossible to review manually at scale.
The legal framework governing digital forensics has not kept pace with these developments. Questions about the chain of custody of data extracted by AI tools, the admissibility of AI-generated metadata analysis, and the disclosure obligations of forces when AI triage tools make decisions about what a human examiner sees and does not see are all insufficiently settled in domestic case law. The Law Commission's review of admissibility of expert evidence — an ongoing project as of the mid-2020s — touches on some of these issues but has not yet resolved them.
AI in forensic imagery and video analysis
Forensic video analysis has long been a specialist discipline, involving the enhancement, interpretation, and comparison of CCTV and other visual evidence. AI has transformed what is possible in this field over the past decade. Enhancement algorithms can recover detail from compressed or low-resolution footage that would previously have been unusable. Gait analysis software can compare the way a suspect moves against footage of an unknown individual at a crime scene. Clothing comparison tools can identify matching garments across different video sequences. And facial comparison — distinct from live facial recognition — can be used to assess whether an individual shown in crime scene footage is consistent with a suspect, though the admissibility and probative weight of such comparisons has been the subject of significant judicial and academic scrutiny.
The Forensic Science Regulator has expressed particular concern about the use of AI enhancement tools in a forensic context, noting that enhancement algorithms can introduce artefacts — visual elements that are not present in the original image but are generated by the algorithm — and that examiners need specific training to identify and account for this risk. The concern is not theoretical: there are documented cases in the forensic literature of enhanced images being presented in court without adequate explanation of the enhancement process, raising the possibility that juries have placed weight on details that were, in a meaningful sense, invented by software.
The admissibility of algorithmic evidence
The admissibility of AI-generated forensic evidence in criminal proceedings engages a cluster of interrelated legal principles: the requirements for expert evidence, the defendant's right to examine the evidence against them, the prosecution's disclosure obligations, and the court's gatekeeping role in relation to potentially misleading or unreliable evidence. None of these principles were designed with AI in mind, and their application to algorithmic outputs is not always straightforward.
Expert evidence in England and Wales is admissible where it is within the expert's area of expertise and assists the court in a way that goes beyond what the jury could determine unaided. Where an AI tool generates the output that the expert then interprets, questions arise about whether the expert can genuinely explain the basis for the output, how error rates should be communicated to the jury, and whether the opacity of a proprietary algorithm is compatible with the defendant's right under Article 6 of the European Convention on Human Rights to examine the evidence against them. These questions are being worked out incrementally in individual cases rather than addressed by any overarching legislative or judicial framework, which creates significant inconsistency in how AI forensic evidence is handled across different courts and jurisdictions.
Follow the coverage
PoliceAI News tracks AI in forensic science continuously — new tools, validation controversies, appeal cases, Forensic Science Regulator decisions, and policy developments as they emerge.
View Live Forensic Science StoriesYou can also browse the full archive on this topic or explore related subjects: AI in courts and sentencing, data protection and GDPR, and AI bias and discrimination.