Presentations on the Policy Papers from Probable Futures
Participants in the workshop included tech industry experts, criminal and civil justice personnel and academics. The day began with Professor Marion Oswald and Professor Carole McCartney presenting the principles that guided the workshop, followed by the rest of the PROBabLE Futures team discussing the four policy papers regarding lessons for law enforcement from:
Workshop Breakout Discussions
The presentations of the case studies were followed by breakout discussions in which each table was asked two questions:
1) What other developments or domains might hold lessons for AI tools in law enforcement?
2) What other lessons might we learn from history/other domains that may mitigate risks of new AI tools in law enforcement?
The participants shared their brilliant ideas on which other sectors and technologies would be important case studies to consider:
Of these diverse examples, one analogy that stood out during the discussions was the Police Dog (K9 Unit). Participants and speakers noted that police dogs, much like AI are ‘probabilistic’ in nature. As one attendee noted, a dog might indicate the presence of drugs, but the dog does not make the arrest – the human handler interprets the signal and makes the decision. This serves as a powerful, pre-existing model for the ‘Human-in-the-loop’ approach to AI. Just as a handler knows a dog’s specific training and constraints, officers using AI must understand that the tool offers a probability, not a definitive answer.
Tech Panel Q&A
The workshop also provided the opportunity for tech industry experts to present their work, featuring representatives from Collaboraite, Palantir, OSIRT, Microsoft and PA Consulting.
Several themes (presented in the word frequency diagram below) emerged from the interaction between the audience and panellists.
A key theme of the panel presentations was challenging the standard terminology of AI governance. One of the panellists challenged the room to move away from the phrase ‘Human-in-the-loop’ and instead aim for ‘Tech-in-the-loop’. This was based on the argument that the human must always own the loop and the accountability; the technology is merely inserted to assist or triage information. This was illustrated with the complex challenge of transcribing ‘drill rap music’ for evidence, where semantic mapping is required to understand slang and context – a task where AI assists the human expert but cannot replace them.
- The legacy challenge: This was identified as a major hurdle for innovation with one panellist noting that while approximately £2 billion is spent on police technology, roughly 97% of that is spent on legacy systems, leaving a tiny fraction for new AI capabilities.
- Data connectivity: It was emphasised that the immediate value of AI often is not in ‘predictive policing’, but in connecting disparate systems so officers, for instance, do not have to ‘copy-paste data between ten windows’.
Large Language Models: Case Progression in Criminal Justice
The workshop concluded with a talk from Professor Dame Muffy Calder on Large Language Models (LLMs) in criminal justice and risks to case progression, including:
- 15 police tasks where LLMs-based tools could be employed across 4 stages of Criminal Justice System
- 4 categories of risk: Outputs, Inputs, Evaluation, Systems Engineering
- 17 risks to case progression
- 41 examples – of risks and impacts related to the tasks
One particularly striking example provided was the risk of Linguistic Homogenisation. The warning here is that if two officers attend a crime scene and use an LLM to help draft their statements, the resulting reports might be so linguistically similar that defence counsel could accuse the officers of collusion. In this sense, a list of known LLM risks effectively becomes a ‘crib sheet’ for defence lawyers to challenge police evidence.
Key Takeaways from the Workshop
The workshop highlighted that while the technology is new, the challenges are not.
- Learning from history: From the case studies presented, and additional ones identified, the consensus is that there already exists past and present technologies and processes for managing probabilistic tools. The challenge is applying these lessons to the speed and scale of AI in law enforcement.
- Triage, not decision: Whether it is a police dog or a complex algorithm, the consensus was that these tools function best as triage mechanisms, helping humans process vast amounts of data (like hours of video or thousands of emails) to find what matters, rather than making the final judgment.
- Fixing the data landscape: Before deploying advanced AI (this is applicable in many sectors), the underlying data landscape must be fixed. As noted during the tech panel, connecting systems is often the most impactful step.
About PROBabLE Futures
PROBabLE Futures, one of Responsible AI UK’s keystone projects, is a four-year interdisciplinary research initiative focused on evaluating probabilistic AI systems across the criminal justice sector. The project, working alongside law enforcement, third sector and commercial partners, is developing a framework to understand the implications of uncertainty and to build confidence in future Probabilistic AI in law enforcement, with the interests of justice and responsibility at its heart.