Toxic beauty advice and toxic influencers are causing a self-esteem crisis in teenage girls worldwide. 1 in 2 girls follow influencers that make them feel less confident, and 92% of teen girls want to change the way they look. Tragically, toxic advice has now become so common, girls don’t even realise it’s toxic anymore. Our idea was to use deepfake technology to put the words of toxic influencers – words heard every day by young girls – into the mouths of the one person they trust most in the world: their mom. This enabled us to expose the issue and alert parents to toxic trends across social platforms, and at the same time help girls recognize and avoid toxic advice online. The film ends with a rallying call for parents to support their girls and leads viewers to Dove Self-Esteem Project resources, to help them detoxify their daughters’ social feeds.

For our social experiment to succeed, we had to convincingly deepfake real moms so that the person who knew them better than anyone – their daughter – believed they were saying harmful and toxic beauty advice. For one of the first times in history, we created high-quality deepfakes of ordinary people, rather than celebrities, using a limited about of video and auditory data to build face and voice profiles from. We recorded the casting process, which allowed us to covertly capture data of the moms’ facial movements, expressions and speech patterns, and construct the deepfakes in secret.

Using large-scale social listening, we uncovered the most toxic beauty trends as they spiked in popularity in 2022: the growing ‘baby botox’ trend, for example, encourages teen girls to get “preventative” plastic surgery before they become adults. Trends were also broken down by age and demographic, reflecting the specific toxic beauty standards girls experience at certain ages in different cultures. Then, we analysed each girl’s social media feeds for toxic beauty trends. This ensured each girl saw their mom mentioning the exact trends they’d already seen on their social media. Ultimately, we tracked fifty toxic trends in real-time across multiple social platforms, inputting them into our deepfake models only two days before the experiment to guarantee they were as relevant as possible.

Harnessing natural language processing AI, we built transcripts based on the posts and language of hundreds of toxic influencers who were popular among the wider teenage demographic and were followed by our girls. The means that our deepfake moms used the exact phrases, language and syntax of toxic influencers. The result was five bespoke minute-long deepfakes, all of which fooled the daughters into believing their moms were giving them toxic beauty advice.