Ask Sumsubers: How Do I Distinguish a Real Person from an AI-Generated Model?

Sumsub is delighted to announce its new bi-weekly expert Q&A series.

Pavel Goldman Kalaydin

Head of AI/ML at Sumsub

Sumsub keeps getting more and more questions from our social media followers about the specifics of compliance, automated solutions, and everything in between. Seeing this demand, we’ve decided to launch a new bi-weekly Q&A series, where our legal, tech, and other experts answer the most frequently asked questions. Check out The Sumsuber and our social media every other Thursday for new answers—and don’t forget to ask about the things that interest you.

This week, our Head of AI/ML, Pavel Goldman Kalaydin, will briefly discuss how AI can bypass facial recognition.

How does Sumsub distinguish between a real person and AI? What are the key AI patterns?

We should start by specifying that, behind every AI generated persona or deceived identity, there’s still a real human trying to pull the strings. And since there’s a person creating the AI, we can expect that they will make mistakes in the process. What Sumsub needs to do is to spot these inaccuracies. Depending on their skills, fraudsters can try to deceive the system in a variety of ways—from simply wearing a mask to creating complex deepfakes, which use machine learning to either generate a fake persona or impersonate an existing person using manipulated photos and videos of them. Our automated systems use data collected during the onboarding process to spot artifacts that indicate deepfake usage. Some of these red flags include inaccuracies in eyes, teeth, borders of the masks, and other features that a person may not spot right away.