AI Revives Criminal Physiognomy

An independent researcher trained a convolutional neural network on 160,000 mugshots (from a 1.2 million–record scrape) and claims 69% accuracy at identifying convicted pedophiles by face alone, noting offenders skew older, white, and overweight. Citing Kosinski et al., the post positions this as a natural extension of face‑to‑trait prediction that journals have shunned. Whether valid or flawed, the work shows how easy it is to build and publicize forbidden classifiers outside institutional review. — If physiognomic classifiers are trivial to build and circulate, policymakers, platforms, and law enforcement must plan for discriminatory screening, vigilantism, and governance beyond academic ethics boards.