Investigative interviewing – the process of obtaining accurate and complete accounts from victims, witnesses and suspects – is the lifeblood of the criminal justice system.
When a crime occurs, someone usually knows something. But the way a police interview is conducted doesn’t simply determine whether information is obtained. It shapes the reliability and completeness of that information – and the credibility of everything that follows in the criminal justice process.
For much of the 20th century (and in many places still today), police largely used accusatory, non-evidence-based interrogation methods that heighten the risk of false confessions.
Research examining police interviews with suspects – notably John Baldwin’s 1992 study of over 600 interviews in England and Wales, commissioned by the UK Home Office – showed that officers routinely relied on assumption and confirmation-seeking, rather than genuinely open-minded information gathering.
False confessions were occasionally extracted from innocent people, witness accounts were shaped by the questions put to them, and the reliability of evidence was often compromised from the moment of its collection.
AI has long been discussed as a threat to jobs and livelihoods. But what’s the reality? In this series, we explore the impact it is already having on different occupations – and how people really feel about their AI assistants.
At the core of investigative interviewing is the often underappreciated, but scientifically accepted, view that human memory is not a recording.
Rather, it is a reconstructive process, fallible to distortion from details that a person encounters after witnessing an event – including leading questions, media coverage and input from other witnesses. This has implications for how police interviewers conduct interviews.
The “misinformation effect”, whereby exposure to inaccurate post-event information alters a witness’s recall, is among the most robust findings in cognitive psychology. Co-witness contamination is a related challenge, because witnesses who discuss an event can inadvertently synchronise their accounts – sometimes importing details they had never actually witnessed.
In 2001, a national evaluation conducted by one of us (Rebecca Milne) with Colin Clarke found significant deficits in police interviewing practices in England and Wales.
More recent research has found that this gap is not simply down to lack of training, but also time pressures, organisational priorities favouring generalists over specialists and institutional culture within different police forces.
Sir Brian Leveson’s independent review of the UK criminal courts in England and Wales highlighted these concerns in 2025. He proposed greater use of AI in policing as a potential solution – including in investigative interviewing.
So, could AI make a positive difference – and what are the potential pitfalls?
Use of AI in investigative interviews
The UK government is keen to explore the full potential of AI in policing. In June 2026, it launched the country’s first AI centre for policing, promising to “accelerate responsible use of AI across all 43 forces in England and Wales”.
There are three key aspects of investigative interviewing where AI could, in theory, prove useful: information gathering, interview honing, and officer training.
In the initial stage of information gathering following a suspected crime, conversational AI agents (chatbots) capable of conducting initial interviews at scale could free up officers from this time-consuming work. Automated transcription tools can also help at this stage.
Then, as the investigation’s focus narrows, real-time prompts could help interviewers ask witnesses and victims open, non-leading questions. AI tools could also be used to generate fresh lines of inquiry.
In addition, AI-enabled avatars such as EchoMind and Innsikt could be used to train officers in the best approach to investigative interviewing.
However, all of these areas require further research. As the criminal psychologist Julia Shaw notes, AI can introduce post-event misinformation into witness memories. Our ongoing research shows that exposure to sycophantic AI sytems which validate and elaborate accounts, rather than passively receiving them, can lead to significant decreases in recall accuracy and increases in recall errors.
What we are seeing empirically is that AI can weaken memory. Witnesses who interact with AI remember differently and, potentially, inaccurately.
The effects on vulnerable witnesses including children, trauma survivors and people with cognitive disabilities represent a particularly urgent unknown.
We also don’t yet know how AI contamination scales across investigative workflows, or whether contamination compounds as information passes through multiple AI-mediated stages.
There is already a documented risk of automation bias – the tendency to defer to algorithmic outputs over independent human judgment – in criminal investigations. When AI systems present confident, well-formatted outputs, expert decision-makers rely more on AI rather than trusting their own assessments.
The AI evidence gap
AI systems used in live investigative interviews can produce unexpected behaviour and patterns of influence not easily observed in controlled laboratory settings.
Any issues in how AI-generated interviews with witnesses affect juror reasoning, for instance, will only show up in the courtroom, long after the interviews have been done.
In response to such concerns, our upcoming (as yet not-peer-reviewed) paper will present a systematic taxonomy (or playbook) of potential AI large language model errors in different aspects of law enforcement.
These include factual errors (the AI stating something that is untrue), faithfulness errors (creating material that does not accurately reflect the information given) and task-specific errors (failing at a specific task despite getting the facts right). Research has also modelled how such errors propagate across decision systems.
Science-based investigative interviewing took decades to develop, empirically validate and (albeit imperfectly) implement. This is reflected in England and Wales by Achieving Best Evidence, the Ministry of Justice’s official guidance on how victims and witnesses should be interviewed, so that their accounts hold up as evidence in court.
In contrast, we believe AI is now being integrated into investigative processes at a fast pace without acceptable scientific standards.
The concerns we raise are not arguments against innovation. They are a call for AI use consistent with the scientific standards that have long guided the field of investigative interviewing. Decades of research has shown us what robust evidence looks like. We should insist on it.
The authors do not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and have disclosed no relevant affiliations beyond their academic appointment.
