The Capture season three: experts in facial recognition and AI decipher the fact from the fiction

The BBC’s conspiracy thriller drama The Capture is back for a third season. The first two series had viewers hooked with a story that intertwined police investigations, facial recognition and deepfake AI technology.

As experts in facial recognition and AI, we’re separating the fact from the fiction ahead of the new season.

Fans of The Capture will be familiar with scenes of investigators using facial recognition software to identify the people they are tracking around London – the soldier Shaun Emery (Callum Turner) in season one and the Russian mercenary Nikolai Mirsky (Jack Sandle) in season two.

Real facial recognition

Real facial recognition work involves several steps. An operator first uploads an image into facial recognition software, which searches for the presence of a face. Features are then extracted from each detected face and compared against the features of faces from a stored database. Features are things that the algorithm has identified as important for recognition decisions and are probably not describable attributes of a face as we know it (such as eyes, nose, mouth).

It is not feasible, nor indeed ethical, to run a facial recognition system against all images on the internet. Not least because of the technical limitations of searching for, storing and processing such a staggering amount of data. In police or intelligence settings faces are searched against specific databases and a human almost always reviews the output because algorithms are not perfect. It is crucial that human operators are trained on the strengths and limitations of these systems and have the necessary skills to review the output.

In season two of The Capture, the fictional facial recognition software company “Xanda” claims their system can “recognise a face with up to 100% accuracy … from every corner of the world with equal precision”. But they are opposed to independent testing.

The best facial recognition systems are now extremely accurate, under increasingly difficult image scenarios. However, accuracy and demographic differences vary widely across different systems and testing parameters. Racial bias is a genuine concern and independent testing is of paramount importance.

In season two, Xanda’s technology claims it can literally “unmask” a face. In reality, some algorithms compare masked to unmasked faces with high accuracy, but they cannot recover information from underneath the mask. Attempting to do so would produce inaccurate and dangerous results.

In the show, patrolling officers receive images of targets, and eyewitnesses perform identifications. This reflects a genuine use of human facial recognition in forensic operations. Most humans make errors on facial comparison tasks involving unfamiliar faces. Super-recognisers (people with a naturally high recognition ability) and trained forensic examiners are more accurate. Familiar humans (people who know the target) often make accurate identifications even in low-quality photos.

Real deepfakes

Deepfakes are digitally manipulated videos, images or voices created with AI to make it appear that someone did something that they didn’t. This technology already exists. Real examples of misuse include fake political videos, non-consensual intimate imagery, child exploitation images and fraud.

Characters in The Capture frequently mistake deepfakes as genuine sources of information. In season one, manipulated video footage causes chaos for Emery, and in season two politician Isaac Turner (Paapa Essiedu) is the subject of a rampage of manipulated footage which the public believes is real.

When asked how Turner sounded in a faked call, a character replied “like Isaac Turner”. This acceptance of deepfakes is realistic. Human detection accuracy is around a coin flip: a 50% chance of a correct detection in scientific studies.

Turner’s wife (Charlie Murphy) was confused by the content of her husband’s deepfake speech, but complimented the way he spoke: “You did good though … Not what you said but the way you said it. You sounded … authoritative.” Cloned voices are typically rated positively and as more dominant than real recordings. Though in practice, deepfake videos tend not to be effective for the types of political activity depicted in the show, and manipulated imagery is not a new problem, as it suggests.

Depictions of deepfakes in The Capture involve almost instantaneous video manipulation and live broadcast. This is far from the current state-of-the-art. While the technology has improved, it is prone to producing obviously fake videos. Creating a high quality video requires a large and diverse collection of photos for the target person, along with iterative tweaks to the output. This is time consuming, and impossible to do in real-time.

Moreover, a capability that can hijack any CCTV or live TV feed and instantly play any content does not yet exist. Camera networks exist over a variety of different technical protocols, ownership arrangements, and access models. That complexity is a roadblock for the kind of point-and-click hijacking depicted in the show.

What next?

In season three, detective Rachel Carey (Holliday Grainger) will encounter further deceptive footage. How can she trust what she sees?

Real strategies to detect deepfakes include training humans to detect artifice, familiarity with the person depicted, and liveness checks that measure natural human responses to various changes in the scene.

In season two, Carey recognised genuine footage of Turner because he held the cross on the necklace around his neck, which his deepfake never did – idiosyncratic mannerisms may be missing in a deepfake.

Detection algorithms developed by the media forensics community and digital watermarks (“invisible” codes in imagery detectable by algorithms) are technological countermeasures that also help.

There are new AI techniques, such as the creation of realistic fake faces and agentic AI, where autonomous software systems perform tasks independently of human control. Will we see the face of an identity that does not exist to throw off the investigators? Could AI get out of hand and create deepfakes for its own purposes, providing a dilemma for both the good guys and the bad guys in the show?

These are the challenges we’d like to see Carey tackle, because they’re ones society will probably face in the near future.

The authors do not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and have disclosed no relevant affiliations beyond their academic appointment.