(CNN Business) — Instagram is testing new ways to determine the age of the youngest on the platform, including using artificial intelligence (AI) to analyze a photo and estimate the user’s age.
On Thursday, Instagram, owned by the company Meta, said in a blog post that artificial intelligence is one of three new methods it is testing to verify the age of users on the photo-sharing site. Users will need to use one of the options to verify their age if they change their date of birth on Instagram from under 18 to over 18. Instagram is testing these options first with its users in the United States. The social network already requires users that indicate their age when they start using the service, and uses AI in other ways to determine whether users are minors or adults.
This move is part of an ongoing effort to ensure that younger users of the photo-sharing app see age-appropriate content. It comes less than a year after revelations from a Facebook whistleblower raised concerns about the platform’s impact on younger users. Last year, Instagram came under fire when documents leaked by the whistleblower, Frances Haugen, showed that she was aware of how the social networking site can damage mental health and body image, especially among teenage girls.
The technology comes from a London-based company called Yoti. A video An animation that Instagram posted on its blog gives an insight into how Yoti’s AI age estimation works: a user is asked to take a video of themselves in selfie mode with their smartphone (Yoti says this step works to make sure there is a real person in the resulting image), and Instagram shares an image from that video with the company.
Yoti’s artificial intelligence first detects that there is a face in the photo and then scrutinizes their facial features to determine the person’s age.
Julie Dawson, Yoti’s head of policy and regulation, told CNN Business that its AI was trained on a dataset made up of images of people’s faces along with their year and month of birth. (The documentation that the company released in May to explain its technology said it had trained on “millions of diverse facial images”).
“When a new face appears, the AI performs a pixel-level analysis of that face and then reports back a number: the age estimate with a confidence value,” explains Dawson. Once the estimate is made, Yoti and Instagram delete the video and the still image taken from it.
Verifying a user’s age can be a vexing problem for tech companies, in part because many users may not have verifiable government-issued photo ID.
Karl Ricanek, a professor at the University of North Carolina at Wilmington and director of the Facial Aging Group Research Laboratory, believes Yoti’s technology is a good application of AI.
“It’s a worthwhile effort to try to protect children,” he says.
However, while this technology could be useful for Instagram, there are a number of factors that can make it difficult to accurately estimate age from a photograph, according to Ricanek, such as puberty, which changes a person’s facial structure, as well as skin tone and gender.
The recent documentation de Yoti indicates that its technology is, in general, slightly less accurate in estimating the age of children who are between 13 and 17 years old and have a darker skin tone than those with a lighter skin tone. According to Yoti’s data, her estimate of age deviates, on average, by 1.91 years for women aged 13 to 17 whose skin tones were classified as the two darkest shades on the Fitzpatrick scale. , a six-tone scale often used by tech companies to classify skin colors, versus a mean error of 1.41 years for women in the same age group whose skin tones were the two darkest shades scale gaps. For boys ages 13 to 17, the technology’s estimate of their age was wrong by an average of 1.56 years, according to the paper. (For adolescents in general, the average error is 1.52 years).
What this means, in practice, is that there will be a lot of mistakes, said Luke Stark, an assistant professor at Western University in Ontario, Canada, who studies the ethical and social implications of AI. “We’re still talking about an absolute mean error, either way, from a year to a year and a half,” he said.
Several CNN employees, all over the age of 25, tested a online demo of Yoti’s age estimation technology. The demo differs from the experience Instagram users will have in that a selfie is taken, rather than a short video, and the result is an age range estimate, rather than a specific age estimate, said the director of Yoti Marketing, Chris Field.
The results varied: for a couple of reporters, the estimated age range was correct, but for others it was off by many years. For example, one editor was estimated to be between 17 and 21 years old, when he is actually around 35.
Among other issues, Stark is also concerned that the technology is contributing to so-called “surveillance creep.”
“It’s certainly problematic, because it conditions people to assume that they’re going to be monitored and evaluated,” he said.