Report Summary
The Chair of the Scottish Police Authority launched a national conversation on the potential use of Live Facial Recognition (LFR) in Scotland in June 2024. This paper aims to provide a summary of the available information on the use of LFR in a policing context.
To access the full document please open the PDF document above.
To view as accessible content please use the sections below. (Note that tables and images are not available as accessible content).
Human Rights and Ethical Considerations - Biases in technology
One of the key concerns that has been raised in relation to Facial Recognition technology in general is the potential for bias in the algorithms used. These have historically been proven to be less accurate at detecting and matching faces of certain skin tones and genders compared to others, often resulting in discrepancies in true positive or false positive matches across different demographics. This bias can be the result of limitations in the AI model itself, often in cases where the data that has been used to “train” the model does not offer a truly representative sample of society. This can lead to the AI being inherently biased towards specific types of faces.
An independent audit by the University of Cambridge’s Minderoo Centre for Technology and Democracy in 2022 found that UK police deployments of LFR had failed to meet minimum ethical and legal standards. This was due to the deployments being too broad in scope and therefore infringing upon the privacy rights of the public. The audit also found that neither the MPS nor SWP had transparently evaluated the LFR tool they used for any bias and had failed to demonstrate that there was always a “human in the loop” ‑ a human operator who assesses and verifies any facial matches identified by the technology.
Following this audit, the MPS and SWP published the results of testing that was carried out on their Facial Recognition Technology by the National Physical Laboratory (NPL). This provided a better understanding of the demographic performance of the LFR system and established that there are settings that the algorithm can be operated at where there is no statistical significance between demographic performance. For more information on the findings from this report please see Appendix D.
It is accepted that the most effective mitigation to prevent any false identifications is to ensure there is a “human in the loop” to verify what the technology is providing as a potential match. However, some argue that human decision making will also bring challenge and bias. Continual testing and evaluation are also required to properly assess how well the technology performs in real-life situations.
