AI Is Undermining Our Privacy. What Can We Do About It?

The talk opens with a robot vacuum maker found to upload photos, video and audio to train its AI, with consent sought through an app whose promised explanation link did not exist. The speaker treats such gaps as design problems and argues that anyone shaping a user experience is a designer. The talk then sets out six areas where AI erodes privacy: opaque data sharing, accidental exposure, reversal of anonymization, deceptive design, always-on listening, and malicious misuse.

Examples include companion chatbot apps that shared user data with advertisers and offered no way to delete it, a chatbot that leaked personal details when asked to repeat a word, and shared conversations that search engines indexed. The talk argues that anonymized datasets can often be re-identified from a few demographic fields, and that language models can match anonymized text to real people more accurately than humans. It also warns that AI could generate dark patterns, citing a pre-checked recurring donation box, and covers a wearable recorder that kept consent mode off by default and a deepfake video meeting that led to a large fraud loss.

The speaker closes with guidance for designers and developers: make privacy the default, seek explicit consent, disclose clearly what data goes where, warn users before they enter sensitive prompts, collect only justified data, distrust anonymization claims, and offer opt-out and deletion. The talk ends by calling for a federal privacy law and framing privacy as a matter of control rather than secrecy.