At the beginning of this year, the social platform Discord implemented a "teen-by-default" in response to global regulatory pressure, treating everyone as minors by default and directly locking some functions. To unlock adult mode, you have to pass age verification.There are two main verification methods: Either take a short selfie video with the camera and let AI estimate the age locally on our mobile phone; or honestly upload the ID card.
As a result, Internet buddies started various extreme tests. Some people used game screenshots to fool them, some used 3D models, and the most outrageous thing was using finger stick figures.
Someone drew two eyes and a mouth on their thumb, raised it in front of the camera, followed the system prompts to center, align, turn left, turn right... After the AI analysis was completed, the conclusion was:Aged 13-15, verified!
The Internet has never lacked imagination when it comes to "taking advantage of loopholes". There was also a 12-year-old boy who drew a beard with an eyebrow pencil and was judged by the system to be 15 years old and successfully passed the review.
Why is this happening?
In order to cope with supervision while minimizing privacy disputes, these platforms have chosen a device-side facial age estimation solution, in which the user records a video in front of the camera. The AI model is completely run on the local mobile phone or computer, and only the final age range judgment results are sent back to the server. The complete video or face data will not be uploaded.
This sounds like a low privacy risk, and the verification process is relatively light, but the trade-off is that available computing resources are limited, the model cannot be too complex, and it cannot rely on heavy-duty analysis on the server side.
These models mainly rely on visual features such as eyes, mouth, facial contours, and skin texture to infer age. A simple thumb drawn with two dots and a line unexpectedly conforms to the abstract feature pattern of a "young and smooth face" under specific light, angle, and low resolution.
When the system requires the "face" to be centered and turned left and right, the natural slight shaking and following movements of the fingers provide a basic motion signal, which is enough for the relatively lightweight liveness detection to consider this as a "live" object and directly release it.
AI age estimation models are mainly trained using massive amounts of real face data, and inherently lack sufficient robustness to such highly non-standard and abstract inputs.
In the early days of rollout, in order to prevent normal users from frequently failing due to poor lighting, wrong angles, or special looks, the platform had to set the threshold relatively loosely, which gave this low-tech evil method an opportunity to take advantage.
Similar situations have occurred on platforms such as Roblox. It can be seen that the current facial age technology is far from being watertight when faced with carefully designed deception methods.
Platforms are not sitting still.
Some time ago, Meta announced that it would use AI to actively scan photos and videos uploaded by users.Use general visual clues such as height and bone structure to determine whether the account owner is under 13 years old, and then remove it directly.
Meta specifically emphasizes: "This is not face recognition." It only makes approximate age range judgments. "
Moreover, Meta not only looks at pictures, but also combines the text content of the entire account with comprehensive portraits of interactive behaviors, such as birthday cakes, mention of school grade, etc., which will be included in the judgment.
Once the AI feels that it is underage, it will first kill the account. If you want to get it back, you have to go through the official verification process. This system is currently online in some countries and will be gradually expanded to more regions in the future. It is also planned to be expanded to Instagram Live, Facebook groups and other scenarios.
The platform is using AI to plug loopholes, and users are finding ways to exploit loopholes.
This cat-and-mouse game will most likely continue.