Safeguarding anonymity at France TV
Last year, the French public broadcaster France TV was alerted that their voice masking methods could easily be reversed, making anonymized people suddenly recognizable. As a result, they swiftly implemented guides and regulations to protect their anonymous sources from artificial intelligence (AI).
Anonymity vs. AI
This incident raises a pertinent question: if voice anonymization can be compromised, what about anonymized images? Initially, simple editing techniques like blurring and pixelating seemed sufficient when tested against Google Images. However, further investigation revealed that blurred images can be reconstructed using mass-market generative AI applications like, for example, AVCLabs. While AI doesn't make the blur disappear, it analyzes and scrutinizes each clue in the image to create a new, artificial portrait of the person that resembles them closely enough to be recognizable. This discovery showed that anonymity was indeed compromised, prompting France TV to implement immediate changes.
To gain further insight, France TV consulted AI expert Professor Titus Zaharia, director of the ArtEmis division at Télécom SudParis. Professor Zaharia confirmed that the threat to anonymization is very real, and that AI can easily reconstruct original images from blurred ones. He also noted that this type of AI can evolve rapidly in the future. With these findings, it became clear that the risk posed by AI to anonymization is significant, urgent and global in scope.
In response, Medialab, a part of the innovation direction of information at France TV that is dedicated to providing the reliable and informed information of tomorrow, implemented large-scale actions to address this threat:
First, they created an anonymity charter, immediately applicable to France TV and its external service providers. This charter summarized responsibilities, main principles, and best practices for implementing appropriate procedures.
Second, they introduced an e-learning training course for all staff involved in anonymization processes, including reporters, journalists, editors, producers and more. This course was mandatory for 5,000 out of the approximately 8,900 employees at France TV.
Third, they sealed internal archives immediately by blocking all content for which it had been indicated that anonymized people were included. France TV also notified the National Audiovisual Institute, which keeps all of their historical archives, and they did the same thing.
Next steps
These rapid measures allowed France TV to secure both new and old productions and focus on exploring and creating solutions for removing identifying information. Despite discussions around replacing blurring techniques with the use of deep fakes "for good," caution is advised. Using artificial humans requires careful consideration, and it may still be preferable to use traditional blurring tools to ensure the audience always knows what is real and what isn't. A very interesting and socially relevant case for the News Hub that other media companies definitely should take note of.
