Digital Experimentation: Testing live facial recognition systems ‘in the wild’

Public sector organisations everywhere are embracing data-driven tools and systems inspired by their promised benefits, despite a dearth of evidence of social and economic benefits.  As AI Watch reported in 2020, ‘most digital transformations in the public sector seem to be guided by hopes and dreams, rather than confirmed by empirical evidence.’ 

For law enforcement authorities in particular, the use of live facial recognition technologies (‘FRT’) holds considerable promise: enabling them to catch terrorists and serious crime offenders and find missing children in busy, crowded yet open public spaces.  To acquire a better understand of whether, and to what extent, FRT can realistically and reliably deliver on its claimed benefits, law enforcement authorities (LEAs) in Europe have sought to trial these technologies ‘in the wild’, despite strenuous opposition from many quarters, particularly civil liberty organisations warning of the serious dangers they pose to basic rights and freedoms and calling for their prohibition. 

The ‘live’ testing of digital technologies on users in on-line environments (such as A/B testing) is commonplace, enabling developers to gather knowledge about which interventions elicit the desired response more successfully than others. These practices have already been criticised for violating basic principles of research ethics.  It is therefore hardly surprising that the conduct of live ‘trials’ of FRT by LEAs has attracted criticism, even though they can be understood as a welcome attempt by LEAs to acquire first-hand knowledge about the capability, usability and challenges associated with deploying live FRT in real-world contexts.  

The aim of this project is to subject the practice of ‘in the wild’ testing of live FRT by law enforcement to critical examination, guided by the following questions (a) what purposes were these tests intended to serve, and how were they conducted? (b) were those tests ‘successful’ and how was ‘success’ understood (and determined) and by whom? (c) What might we learn from these tests, particularly in relation to the EU’s proposed AI Act, and how do they compare with historical experience of risky new technologies and their legal regulation?