
AI in Criminal Investigations: Better Decisions, Not Just Faster Ones
Article written by:
Dr Ivar Fahsing
Chief Scientific Advisor at Davidhorn
Artificial intelligence is assisting criminal investigations, helping officers manage vast quantities of evidence and identify information that might otherwise be overlooked. But without the right safeguards, it could also reinforce the very human biases that contribute to investigative failure.
Drawing on the themes explored in Davidhorn’s white paper, ‘Aim Carefully: Human Bias, Moral AI, and the Future of Investigation’, this article examines how law enforcement can harness AI to strengthen investigative practice without compromising fairness, transparency or accountability.
The challenge for policing is not simply how to use AI, but how to ensure it helps investigators challenge assumptions, consider alternative explanations and reach more reliable conclusions. Because while technology can support the search for truth, responsibility for the decisions being made must always remain human.
Read the whitepaper
Criminal investigations have always involved making sense of incomplete and sometimes contradictory information. What has changed is the sheer volume of material that investigators must consider.
Digital communications, CCTV and other electronic records have created information overload, although this is not a new challenge. The investigation into the Yorkshire Ripper demonstrated how, even before today’s digital technologies, major enquiries could become overwhelmed by the volume of material being collected.
To illustrate the scale, a single terabyte of information, if represented as printed pages, could create a stack stretching for kilometres. For investigators tasked with identifying relevant information, the challenge is considerable.
Human beings are highly capable of recognising patterns, but our ability to consider multiple competing explanations simultaneously is limited. Under pressure, we naturally focus on information that appears most relevant, potentially overlooking evidence that challenges our initial assumptions.
This is where artificial intelligence offers significant potential. Its processing capacity could help investigators navigate large datasets, identify contradictions and examine alternative explanations.
However, introducing AI does not automatically make an investigation more objective or effective. Used incorrectly, it could reinforce existing assumptions, accelerate flawed reasoning and produce conclusions that have not been properly tested.
Responsible AI must be the starting point
Fairness, transparency and accountability provide essential foundations, but the real test is whether AI demonstrably improves investigative quality. The ambition must be better and fairer investigations, where technology strengthens professional judgement, challenges assumptions and reduces the risk of error.
Before implementing AI, police organisations must therefore be clear about the problem they are trying to solve. Introducing technology primarily as a cost-saving measure risks overlooking its impact on investigative quality.
Use AI to test hypotheses, rather than simply confirm them
One of the greatest risks in any investigation is confirmation bias: the tendency to favour information that supports an existing belief while giving less attention to contradictory evidence.
Investigators may begin with a reasonable suspicion, but as an enquiry progresses, the pressure to establish what happened can make it increasingly difficult to reconsider that initial theory.
The first challenge is to identify all reasonable explanations, including those consistent with a suspect’s innocence or the possibility that no crime has occurred. These explanations must then be systematically tested, with investigators actively seeking evidence that could exclude them.
Crucially, this must be a continuous process. As new information emerges, existing hypotheses may need to be reconsidered and alternative explanations explored.
AI could make this discipline more manageable. It could help to generate alternative hypotheses, highlight inconsistencies and identify evidence that challenges the prevailing explanation.
Investigators could ask what might disprove their theory. Which evidence contradicts the current thinking? What information might support the defence or establish that a suspect was not involved? This reflects the principle of falsification: actively attempting to rule out explanations.
However, AI-generated hypotheses must themselves be critically assessed, rather than assumed to represent every reasonable possibility.
AI must support reasoning, not simply generate more information
One of AI’s greatest potential benefits is its ability to help investigators manage information overload.
An AI system might rapidly generate summaries, identify connections and produce recommendations. But if those outputs simply create more material for investigators to assess, the technology may increase cognitive overload rather than reduce it.
The greater opportunity is to develop AI as cognitive infrastructure: technology that supports how investigators think, test evidence and make decisions.
AI could identify contradictions between witness accounts, highlight evidence that does not fit an existing hypothesis or draw attention to reasonable explanations that have not been adequately tested.
A system that processes thousands of documents in minutes may be technically impressive. But if it directs attention towards the wrong information or encourages premature conclusions, faster processing could simply accelerate investigative errors.
The importance of not confusing an AI-generated answer with an established fact
AI systems can produce outputs that appear authoritative, even when the underlying analysis is incomplete or incorrect. This creates particular risks in criminal investigations, where apparently objective recommendations may influence decisions with serious consequences.
Facial recognition technology, for example, has already been associated with cases of mistaken identification, demonstrating the potential consequences of relying too heavily on automated outputs.
A system might identify a pattern of behaviour or highlight a potential suspect. But those findings must be checked against the underlying evidence and assessed within the wider investigative context.
The critical questions remain: What evidence supports this finding? What might contradict it? What alternative explanations have been considered?
If investigators cannot answer these questions, the output should not become the basis for a consequential decision.
The relationship between AI and investigators offers a useful comparison with aviation. Modern aircraft rely on sophisticated automated systems, but pilots remain responsible for understanding how those systems operate, recognising when something is wrong and taking control when necessary. The same principle should apply to investigative AI.
Investigators must remain human-in-command
Investigators must retain the competence, authority and confidence to question the system, reject its recommendations and pursue alternative lines of enquiry.
There may also be a role for safeguards that alert users when they are relying too heavily on automated analysis or have not completed essential verification steps. However, these safeguards will only be effective if officers are properly trained.
As AI becomes more sophisticated, professional competence must include understanding its capabilities, recognising its limitations and knowing when not to use it.
AI should help investigators assess information more effectively. It must never become a shortcut around professional responsibility.
Preserve evidence integrity and make decisions open to scrutiny
Transparency is essential to a fair investigation. Investigators must be able to demonstrate how evidence was obtained, how it was assessed and why particular decisions were made.
AI should strengthen this accountability rather than make the investigative process harder to explain.
A fundamental requirement is preserving the integrity of original evidence. Source material must remain clearly distinguishable from AI-generated summaries, interpretations, hypotheses or recommendations. An AI-generated interpretation of a witness statement, for example, is not the statement itself.
If these distinctions become blurred, investigators risk treating automated interpretations as though they carry the same evidential weight as original material.
There must therefore be a reliable record of how AI was used, including the material analysed, questions asked, outputs generated and decisions subsequently taken by investigators. This information should be available for appropriate scrutiny.
Indeed, one potential benefit of AI is that it could make aspects of investigative reasoning more visible.
Traditionally, much of an investigator’s thinking takes place internally. Well-designed systems could help document the hypotheses considered, the evidence used to test them and the reasons particular explanations were retained or rejected.
This creates an opportunity to improve overall transparency and accountability, provided the technology is designed around established investigative methods and robust evidence-management principles.
Demonstrate improvements in investigative quality, not just efficiency
As police organisations consider adopting AI, considerable attention is being given to technical performance, efficiency and compliance.
A system may be accurate at extracting information, fast at processing documents and compliant with governance requirements, yet still fail to improve the decisions investigators make.
Responsible AI establishes the safeguards within which technology operates. Effective AI must demonstrate that its use produces better investigative outcomes.
Police organisations should examine whether AI-assisted investigations are more effective at identifying and testing alternative hypotheses, recognising contradictory evidence and avoiding premature conclusions.
They should also assess whether the technology reduces investigative errors, improves decision-making and helps investigators reach conclusions that are more reliable and defensible.
Evaluations should also consider unintended consequences. Does the system encourage overconfidence? Are investigators becoming less likely to challenge automated findings? Does it reduce cognitive overload or simply shift that burden elsewhere?
These questions matter because technology can change professional behaviour in ways that are not immediately apparent from performance statistics. The ultimate measure of success should not be the number of documents processed, hours saved or recommendations generated. It should be whether the technology contributes to better investigative decisions, fewer errors and fairer, more reliable justice outcomes.
A safer investigative future
The future of investigative AI should be a partnership between skilled professionals and technology designed to strengthen their reasoning and decision-making.
AI has the capacity to examine information at a scale that humans cannot. Investigators, meanwhile, bring contextual understanding, ethical judgement and accountability that technology cannot replace.
The opportunity lies in combining these strengths intelligently. That means using AI to help generate and test competing hypotheses, identify contradictions, challenge assumptions and manage cognitive overload. It also means maintaining meaningful human control, protecting original evidence and ensuring significant decisions remain open to scrutiny.
But these principles must be accompanied by evidence of effectiveness.
Responsible AI is essential, but responsibility alone is not enough. The ultimate measure of success must be whether AI helps investigators deliver more reliable outcomes.
Ultimately, the greatest risk is not simply that AI might produce an incorrect answer. It is that an investigator might accept that answer without adequately challenging it.
Responsible AI is where we must begin. Better, fairer and more reliable investigations are what we must achieve. Because a wrongful conviction is not an acceptable consequence of technological progress.