AI Facial Age Estimation to enable age appropriate design and experiences - making the internet age aware

As online platforms increasingly cater to users of all ages, robust age assurance becomes critical to user safety, regulatory compliance, and digital inclusion.

In this session, we explore Yoti’s AI-powered facial age estimation — a technology that estimates age from facial images without identifying individuals. Distinct from facial recognition, this privacy-preserving tool deletes images post-estimation and is designed to comply with GDPR and age-appropriate design codes. Through regulatory roundtable insights, field trials, and youth research, we’ll examine how facial age estimation is being used across sectors — from retail and gaming to social media and adult services.

We'll delve into how it addresses key ethical challenges in AI: minimising bias across skin tones and genders, ensuring transparency in performance reporting, and building inclusive systems that respect autonomy and consent.

The talk will share Yoti’s innovative approaches to model training, bias monitoring, explainability, and spoof-resistance, as well as how the technology complements other age assurance tools in a layered “waterfall” approach. We’ll also present consumer attitudes, including findings from Play Verto's youth research, and explore the broader societal implications through the lens of consequences scanning and responsible innovation.

Key Takeaways for the Audience:

1. Privacy-by-Design and Ethical AI Principles Facial age estimation does not involve facial recognition: it doesn’t identify or store biometric data, and all images are permanently deleted after estimation .Complies with GDPR data minimisation and has been certified under PAS 1296, ISAE 3000, and approved by regulators such as Germany’s KJM and FSM

2. Accuracy, Fairness, and Bias Mitigation High true positive rates: 99.3% for 13–17 year olds estimated as under 21; 99.0% for 6–12 year olds estimated as under 13. No discernible bias across genders or skin tones, with transparent publication of Mean Absolute Error (MAE) by year, gender, and skin tone .Ongoing efforts to improve performance in underrepresented demographics through diverse training data collection and targeted fieldwork.

3. Explainability and Threshold Calibration Uses safety thresholds and a “Challenge 25”-like logic, allowing configuration of buffers to reduce false positives and ensure regulatory compliance. Illustrated "waterfall method" combines facial age estimation with other methods like digital ID for layered, inclusive age assurance

4. Innovation in Real-World Deployment Used by platforms across multiple sectors - social media, livestreaming, dating, gaming, adult content, CSAM age detection by law enforcement, retail, ecommerce - by circa one third of the largest global platforms including Meta, OnlyFans, Yubo, Sony Playstation, KWS, PMI, BAT. Proven scalability with over 850 million age checks globally and real-time deployment at 300 checks/second

5. Consequence Scanning and Regulatory Collaboration Yoti’s internal Ethics & Trust Committee has used frameworks like the Consequences Scanning Model to evaluate impacts. Ongoing regulatory engagement includes participation in the UK data protection office the ICO Sandbox to develop explainer materials and extend the age range and regulatory roundtables exploring unintended consequences and intersectional impacts on users

6. Consumer and Youth Attitudes Young people surveyed through Play Verto expressed support for the tech when it improves safety and emphasized the need for: Transparency and clear explanations Real-world trials with feedback loops Preserving autonomy and offering alternatives

7. Compliance as a Catalyst for Safer Design Facilitates alignment with: UK and international Age Appropriate Design Codes EU Digital Services Act (DSA) and AI Act US state-level legislation like California’s AADC 8. Benchmarking, Spoof-testing and Security Innovation Combines facial age estimation with liveness detection (iBeta NIST Level 2) and SICAP (Secure Image Capture and Processing) to defend against bots, deepfakes, and injection attacks Benchmarking by NIST and the Australia age assurance benchmarking Take-aways - links to 3 white papers - on facial age estimation, liveness detection, injection detection Explainer videos