Protecting identities without losing the humanity - anonymising interviewees with AI

Anonymising contributors is a vital part of documentary filmmaking, protecting people who are sharing sensitive personal stories or blowing the whistle on high-stakes issues. Traditionally, techniques like silhouetting, blurring, or hiring actors have been used to conceal identities. However, these older methods can inadvertently alter the tone of a film. Silhouettes, for instance, can make a subject appear suspicious or out of place alongside identified contributors.

To solve this, we have developed two AI-driven anonymisation methods that transform a contributor’s appearance while preserving their original facial expressions. These visual AI approaches allow audiences to fully connect with the emotion of a first-hand account without compromising the anonymity of contributors. We have developed two AI methods for anonymisation. The first uses a process which swaps the heads of our actor and contributor. The second approach uses character animation models to generate an artificial face that gets applied to the contributor.

In the head swap method, we train a series of bespoke models on an image dataset containing both the contributor and a hired actor. The contributor’s dataset comes directly from the interview footage, while we capture the actor’s dataset during a dedicated face capture session. By training on this paired data, the model learns to extract facial expressions and head poses from an image of the contributor. It can then generate a corresponding image of the actor’s head with identical expressions and head pose.

This allows us to input a video of the contributor and generate an anonymised version. This new version maintains the original head movements and facial expressions, producing a realistic recreation of the interview using the actor’s likeness. Crucially, we train our models exclusively on data captured from the specific interview sessions, which eliminates the risk of bias from data gathered elsewhere. We used this technique for a documentary called Matched With A Predator, where we worked with the Disclosure team to anonymise a contributor.

While the head-swap method has proven successful and delivers high-quality results, it presents several logistical and technical challenges that limit its use. One significant hurdle is the requirement to bring a hired actor to the interview location for paired data capture. This is often impractical in scenarios where the contributor requires absolute secrecy, such as clandestine meetings or filming in a private residence. There may also be high-risk environments where bringing an actor would be inappropriate. The head-swap method also relies on a highly controlled filming setup, which can be impractical when filming in public spaces or domestic settings.

There are also physical limitations to this approach. The process only replaces the head, so any other identifying features on the contributor’s body that clothing cannot obscure may remain a risk to their anonymity. Technically, training a model for every contributor-actor pair takes time to process the data and train the model - a minimum of a week, even in the best cases. However, an inference-only workflow that uses an existing model that doesn’t need extra training could be significantly faster. These limitations highlighted the need for a more flexible approach that doesn’t require physical actors or rigid environmental controls.

We therefore developed a second, experimental technique for AI-based interview anonymisation. While we have not yet used this in a broadcast production, our findings are promising. The first step in this method is to film the contributor’s interview. This could happen without strict constraints on the filming environment. We’d then move on to a collaborative design phase, where the production team could use a generative image editing model to design the contributor’s new visual identity. By providing text prompts, producers can specify detailed characteristics such as age, build, and hairstyle. They can also choose to modify only specific attributes, such as replacing the head while retaining the original body. The production team could choose to change the background or keep it identical to the original interview. We’d use this customised reference image and the original interview footage as inputs for a pre-trained, diffusion-based character animation model. This model would generate an anonymised video that faithfully maps the contributor’s original interview onto the newly designed persona, offering a highly flexible and inference-only alternative to the head-swap method.

The development of AI-driven anonymisation is a specialised solution for productions where the emotional weight of a contributor’s testimony is central to the narrative. These methods offer an alternative for high-stakes storytelling where the facial expressions, reactions, and own words of the contributor are important. Our successful use of head-swapping in a broadcast documentary and the potential of diffusion-based animation both demonstrate reliable, cost-effective ways of protecting vulnerable contributors without sacrificing the stylistic or emotional integrity of the film. Our work ensures that even in the most sensitive circumstances, the human element of a story remains intact.