Police need clearer AI rules to accelerate investigations without putting evidence at risk

Uncertainty over the use of AI in police investigations is slowing adoption of technology that could help forces tackle growing digital evidence backlogs, according to John Lucey, VP EMEA at digital forensics company Cellebrite.

AI is becoming increasingly important as investigators contend with growing volumes of data from mobile phones, cloud services, social media and connected devices. However, Lucey warned that deploying the tech without appropriate safeguards could jeopardise investigations and potentially cause cases to collapse.

Speaking to Think Digital Partners, Lucey said clearer national guidance would help forces adopt AI with greater confidence while maintaining the standards required for evidence to withstand scrutiny in court.

“Policing decisions carry consequences that most commercial AI tools were never built to withstand, so when error happens, the stakes are much higher,” he said. “Rushing deployment can risk creating weaknesses in evidence that cause cases to collapse.”

Tackling the digital evidence backlog

The growing volume of digital evidence is placing considerable pressure on police resources, particularly smaller teams with limited capacity to review large quantities of footage, messages and other material.

“This hits smaller teams hardest,” said Lucey. “A small police zone with limited resources and a modest budget may not have the capacity to review hours of footage frame by frame, and that is exactly where AI can support.”

He described one case involving a drone used to deliver drugs into a prison. Although the device had been destroyed, investigators recovered an SD card containing hours of footage.

“Reviewing that by hand would have taken days, and the key detail could easily have been missed,” he said.

AI analysis identified footage of a test flight in a living room, where a document visible beneath a television turned out to be a hospital appointment notice. The technology extracted details including the appointment date, hospital and doctor’s name. According to Lucey, investigators had the suspect in their office within an hour of verifying the information.

He also pointed to the use of AI to search recorded prison phone conversations for references to drugs or firearms, reducing work that could otherwise require dozens of hours of manual review to minutes.

Such capabilities could be particularly valuable for forces struggling with digital forensic backlogs, he said.

“Doing it without AI isn’t impossible, but it isn’t a smarter way of working, and it doesn’t benefit the police force or the community it serves.”

However, he cautioned that faster processing does not necessarily mean better investigations.

“Speed without proper safeguards can overwhelm officers with false positives, so AI is extremely beneficial to manage the data challenge but only when paired with human review.”

Ensuring evidence can withstand scrutiny

The increasing use of AI in investigative workflows also raises questions about how forces establish the provenance and reliability of evidence, particularly where AI-generated material, summaries or transcripts form part of a case file.

Lucey said investigators must be able to establish exactly where AI has been used, which data it has processed and how its outputs have been produced.

“When AI flags a detail in a piece of footage, investigators must corroborate it independently, as they would with any other intelligence before acting on it,” he said. “With this, police forces need a consistent and auditable record of which tool was used, on what data source and what the output generated.”

Without that record, he warned, it becomes harder to challenge or defend evidence in court, or establish whether AI-generated or manipulated content has entered an investigation.

He also called for continuous testing of investigative AI tools to ensure consistent standards across forces, rather than individual organisations relying solely on suppliers’ assurances.

Purpose-built AI versus generic tools

Elsewhere, Lucey argued that policing organisations should distinguish between general-purpose AI applications and tools designed specifically for investigative work.

While generic AI systems may be useful for admin tasks, he said their limitations around accuracy, traceability and data handling can create risks when applied directly to evidential material. Investigative tools, by contrast, should allow outputs to be traced to their original sources and maintain audit logs showing what actions were performed, when and under which instructions.

Data security is another consideration, particularly where investigations involve sensitive personal information. Lucey said forces should expect strong access controls and the ability to process information within secure environments, rather than automatically transferring sensitive material to external services.

Clearer guidance needed

Despite the potential benefits, Lucey said uncertainty around liability, data protection and the admissibility of AI-assisted evidence is contributing to hesitation among policing organisations.

He claimed that forces across England and Wales are either using AI or moving towards adoption, although the maturity of those deployments varies.

Procurement and legal teams are understandably cautious, he said, particularly given the potential consequences of errors in criminal investigations. However, he argued that prolonged uncertainty could leave forces dependent on manual processes while digital evidence volumes continue to grow.

“The answer isn’t less caution or less regulation, but rather clearer regulation,” he said. “There should always be transparency, continuous testing and human oversight regardless of the level of regulation, but clearer national guidance would let forces move with confidence.”

Beyond the technical and legal challenges, Lucey said forces must demonstrate that their use of AI is proportionate and respects privacy and civil liberties. This includes explaining to the public how AI is being used, what happens to personal data and where human decision-making remains part of the process.

He also argued that responsibility for the safe use of investigative AI should be shared between police forces and technology suppliers.

Officers must be trained to understand the limitations of the tools they use and know how to respond when something goes wrong. Suppliers, meanwhile, must be transparent about how their systems operate and where they may be unreliable.

Ultimately, Lucey said AI should support investigators rather than make decisions on their behalf.

“AI is an assistant to accelerate casework and connect siloed data, not an investigator – it helps us get to the truth faster, but it doesn’t decide the outcome of a case.”