A Coin Flip at Eighteen: On-Device Facial Age Estimation with Open-Source Models Cannot Separate Teenagers from Young Adults

A growing number of U.S. state laws require online services to keep minors out or to verify users' ages, and facial age estimation is a common way to do it. Running the estimator entirely on the user's device is an attractive privacy design since no face image ever leaves the device. In this paper I ask whether freely available open-source models are accurate enough to support that design. I benchmark two widely used estimators, the InsightFace genderage model (1.3 MB) and the DeepFace age model (539 MB), on all 23,705 images of the UTKFace dataset, running locally on a consumer laptop (Apple M4) with a shared face detector. Both models fail as 18+ gates. At a threshold of 18, InsightFace admits 95.6% of 13–17 year olds and DeepFace admits 100%. The failure itself is not one of calibration: the area under the ROC curve for distinguishing 18–24 year olds from 13–17 year olds is 0.51 for InsightFace and 0.56 for DeepFace, close to chance, so no buffer age fixes it. Holding the share of admitted 13–17 year olds to 5% requires a threshold of 51 (InsightFace) or 34 (DeepFace), which refuses 75% and 73% of all adults respectively. At a 25-year buffer, error rates also differ by gender and by dataset ethnicity label: in both models, female teenagers are admitted more often than male teenagers. By contrast, the on-device cost is negligible: InsightFace runs in about 0.55 ms per face on CPU. Privacy-preserving deployment is therefore not the bottleneck; model accuracy in the 13–24 band is. I argue that age-gate evaluations should report separability in that band rather than overall mean absolute error, which hides the failure, and I also release my benchmark code.