Keep a history of the biometrics of the devices either on the account or in a cookie. Use a sufficiently secure and obfuscated language to capture and upload the data (maybe using steganography in an image, explaining the weird image uploads.) Prompt when the biometrics don't match a known user of the device/account (according to ml). Or if the tracking data didn't exist (incognito) then prompt using a backup model trained in the difference between the biometrics of known bots and known humans. Keep a very loose threshold here.
That's how I'd do it if I were Google (and no one else because only Google can afford that.) All the browser fingerprinting stuff mentioned is great but doesn't really work as much as you'd hope in practice.