Instagram, which is banned for children under the age of 13, will use artificial intelligence to determine the age of its users. A measure that is still experimental, first tested in the USA.
The announcement was made this Thursday, June 23, 2022: Meta, Facebook’s parent company, will test a new way to verify the age of its users. “ If someone tries to change their date of birth from under 18 to over 18, we will ask them to prove their age ‘ the company said in a press release.
For this, users have 3 options:
- verification with an identity card,
- confirmation with relatives of the user,
- and shooting a selfie video in collaboration with Yoti, a company specializing in online identity verification using artificial intelligence.
The test phase will only take place in the USA for the time being. However, such a measure raises many questions, especially in terms of artificial intelligence bias.
How would that work exactly?
The reason Instagram wants to verify the age of its users is simple: the platform is banned for anyone under the age of 13. Besides that, ” In some countries the minimum age for registration is higher says Instagram. ” When we know that a user is a teenager between the ages of 13 and 17, we offer them age-appropriate experiences, e.g. B. by making his account private by default […] and we limit the ways advertisers can reach them with advertising. »
In the event affected users choose to provide a video selfie to prove their age, they would need to take a video of themselves moving their head. Once taken, the selfie is only sent to Yoti, says Meta, who can’t tell her identity — only her age. ” Yoti’s technology estimates age based on facial features ‘ explains Meta. However, the company does not specify which characteristics are examined. Once the analysis is complete, Meta and Yoti delete the images.
On its website, Yoti explains that its technology was developed using a type of artificial intelligence called a “neural network”. ” For AI, the image is simply a structure of pixels “, Yoti states on his website, ” and our technology has been trained to recognize structures “, so that she learned to distinguish the models that are usually attributed to people of a certain age. ” Yoti’s technology does not recognize the identity and the images provided are not reused to train the AI. »
Does age verification by facial analysis work?
Yoti is confident in its technology: for ” 6-12 year olds, the mean absolute error [l’écart moyen des prédictions par rapport aux valeurs réelles, NDLR] is 1.36 years and for 13 to 19 year olds it is 1.52 years.”. A total of 99.65% of users between the ages of 13 and 17 were correctly identified as under 23, according to the company.
But there are mistakes and, as with all AIs, prejudices. Yoti’s is no exception: it records more errors for dark-skinned users. While the margin of error remains fairly small, it is striking that it increases from 2.94 for the whitest women aged 25-70 to 4.79 for the darkest-skinned women.
But unfortunately that’s all the company communicates about it: There are no details on whether the tool will be improved or if the company wants to find the source of the bias. This could be partly explained by Yoti using Fitzpatrick’s classification.
This is a tool developed in the 1970s to classify skin types that are least sensitive to UV rays. However, this classification, widely used in engineering to train artificial intelligence, is increasingly criticized for its lack of nuance in relation to the darkest skins, which are underrepresented. On the contrary, it would be biased in favor of the lightest skin. Google has also announced that it will no longer use the Fitzpatrick classification for the development of its products in 2021.
Use in France?
Despite Yoti’s AI bias, Instagram shared that testing has already begun. They’re currently only taking place in the United States, and Meta hasn’t commented on a possible arrival in Europe. However, Yoti is already in use in Germany. “ Yoti’s technology has been approved by the German regulator ” explains the company, which also states that its app can be used to verify the age of people who are in Germany and want to access “. Adult content “.
Such an agreement has not yet been concluded in France. But the government has also been thinking about solutions for some time to restrict access to pornographic content for adults. Arcom therefore sent a formal notice to 5 pornographic websites in December 2021, asking them to ” concrete measures to prevent minors from accessing their content “. The paths taken by the government are for the time being nonetheless different : The purchase of a “porn pass” in store has been mentioned, like a check by bank card or even an age confirmation with… FranceConnect.
Linc, the CNIL’s innovation laboratory, has also just published a guide for ” a privacy-friendly age verification mechanism “. The latter recommends the facility a Protocol primarily based on a cryptographic concept called group signature “which would allow” a user within a group can sign anonymously “. No video selfie questions yet.