The UK’s policing AI lead, Alex Murray, has acknowledged that crime-fighting technology will contain bias, which is an inherent risk in AI systems, but pledged that it would be addressed through stronger safeguards and oversight.
This declaration comes as law enforcement agencies move forward with establishing a new national AI centre to coordinate and oversee the use of artificial intelligence across forces. Meanwhile, police forces are increasingly deploying data-driven tools to support operational decisions, from intelligence analysis to risk assessments and resource allocation.
At the heart of the issue is data bias. AI systems are trained on historical policing datasets that reflect not only recorded crime but also past enforcement patterns, reporting practices, and operational priorities. If that data is incomplete, inconsistent or shaped by historic inequalities, predictive models risk reproducing and potentially amplifying those distortions at scale.
Dr. Janet Bastiman, Chief Data Scientist at Napier AI, said: “Admitting that AI systems will contain bias is not the same as addressing it. Bias is not a minor technical issue that can be patched later, but stems directly from the data used to train these systems and the assumptions built into their design. If that data reflects historical inequalities or skewed enforcement patterns, the technology will inevitably reproduce them.
In policing and other high-stakes settings, this has real-world consequences. AI trained on historical crime data can end up codifying patterns of enforcement rather than patterns of crime. Once embedded in operational decisions, those distortions can scale quickly and become harder to detect.
The message is simple. You cannot build fair outcomes on biased foundations. Confronting data bias at its source, through rigorous auditing, transparency and meaningful human oversight, must be the starting point. But beyond this, deliberate approaches to tackling data bias must stem from tackling bias within teams as the data produced is only as inclusive and diverse as the teams responsible for it.”
The new AI centre is intended to improve coordination, governance and standards around the development and procurement of these technologies. Its role will include ensuring greater transparency, consistent evaluation frameworks and clearer accountability for how models are trained, tested and deployed in live policing environments.
As adoption accelerates, experts stress that robust data governance, independent auditing and ongoing monitoring will be essential to prevent feedback loops and unintended consequences. In high-stakes public sector settings such as policing, the integrity and representativeness of underlying data is not a technical detail; it is fundamental to fairness, public confidence and lawful decision-making.
Stuart Harvey, CEO of Datactics, said: “Policing is one of the most sensitive environments in which AI can be deployed, because decisions directly affect people’s freedoms and communities’ trust. When we talk about bias in police AI, we’re really talking about the quality of the data behind it, how applicable and appropriate it is to how it has been trained. Historical policing data does not just reflect crime patterns, it reflects where officers were sent, which communities were scrutinised and how incidents were recorded.
If those datasets are incomplete, inconsistent or skewed, predictive systems will mirror that imbalance. In a policing context, that can influence patrol allocation, risk scoring and investigative focus, potentially reinforcing disparities rather than improving outcomes.
The starting point has to be robust data governance. Forces need clear visibility of where their data comes from, how it has been shaped over time and how models perform across different groups and geographies. Without disciplined data management and ongoing oversight, AI risks undermining the very public confidence it is meant to strengthen.”
Leave a Comment