Facial recognition and rights

Why this matters

Facial recognition is one of the most discussed uses of AI in policing anywhere in the world. It promises to help identify suspects or find missing people by comparing images of faces. It also raises some of the sharpest questions about privacy, fairness and the relationship between the police and the public.

You may never operate such a system. But you may be shown a "match", asked to act on one, asked by a member of the public whether they were scanned or asked to help decide whether a tool should be used. This module explains what facial recognition can and cannot do and the safeguards that matter. It does not describe any system in Pakistan.

What facial recognition does

Facial recognition is a form of computer vision. In broad terms:

• Verification checks whether a face matches one particular person, such as unlocking a phone.
• Identification compares a face against many stored images to find possible matches.
• Live scanning compares faces in a crowd or camera feed against a watchlist as people pass.

These carry very different levels of intrusion. Live scanning of public spaces affects large numbers of people who are not suspected of anything.

Accuracy limits

A facial recognition system does not "recognise" a person the way a friend does. It produces a similarity score and someone sets a threshold above which a result counts as a match. That means:

• False matches happen. The system can link an innocent person to an image.
• Missed matches happen. The system can fail to find someone who is in the database.
• Image quality matters. Poor lighting, low resolution, camera angle, partial coverings and age differences all reduce reliability.
• The threshold is a choice. Lowering it finds more possible matches but also more wrong ones.

A system that performs well in testing may perform worse in real conditions such as a busy street at night.

Bias

Facial recognition systems have been found in independent testing elsewhere to perform less accurately for some groups than others, depending on how they were built and the images used to train them. Errors that fall more heavily on some groups are a form of bias. They can lead to some people being stopped or suspected more often for no good reason. Any proposal to use such a tool should ask how it performs across the population it will be used on and who checked.

Privacy and lawful authority

A face is personal data that cannot be changed like a password. Collecting, storing and comparing images of faces is a significant intrusion into privacy. Before any use, there should be a clear answer to these questions:

• What is the legal basis? Police powers come from law. Using a technology that affects people's rights needs clear lawful authority. Ask your legal adviser and do not assume.
• Is it necessary and proportionate? Is this the least intrusive way to achieve a legitimate aim?
• Whose images are stored, for how long and who can see them?
• Has an impact assessment been carried out covering privacy, fairness and risks to particular groups?

Oversight and human decisions

A match is a lead, not an identification. Good safeguards include:

• A trained person reviews every match by comparing the images themselves and looking for other evidence before any action.
• No arrest or charge rests on a match alone.
• Decisions are recorded, including who reviewed the match and why they acted or did not act.
• Independent oversight checks how the system is used, how often it is wrong and whether it treats groups differently.
• People can complain and seek redress if they believe they were wrongly identified.

This is human oversight in practice. It only works if the reviewer has the time, training and authority to disagree with the machine. Watch for automation bias: a high similarity score can make a weak match feel certain.

Try it

Read a published news report or public statement from any country about police use of facial recognition. Using the questions in this module, list what the report tells you about legal basis, accuracy, bias testing, human review and complaints. Note what is missing.

Check yourself

1. Why is a facial recognition "match" not the same as an identification?
2. Name three things that reduce facial recognition accuracy.
3. What four questions should be answered before any use?

Answers: (1) It is a similarity score above a chosen threshold; it can be wrong and must be reviewed by a trained person with other evidence. (2) Any three of: poor lighting, low resolution, camera angle, partial coverings, age differences, threshold settings. (3) The legal basis; whether it is necessary and proportionate; whose images are stored, for how long and who can see them; whether an impact assessment has been carried out.