The AI privacy problem we’re opting into

TL;DR: Last week, OpenAI rolled out a new ChatGPT feature, Computer History, that tracks how you use your computer so it can better understand what you’re doing. It’s the latest example that more personalized AI does require more personal data. But before you say “duh,” consumers still don’t seem to understand what they’re giving up.

What happened: Sam Altman mused in a recent interview that within the next six months, a “descendant of ChatGPT” could watch your computer all the time to have “perfect context of your whole life.” Computer History is a step in that direction—it logs what you click and type on your computer (so far, it’s available on macOS), and then suggests workflow automations or finishes a task you started.

OpenAI says that while the events recorded in Computer History are stored temporarily on your machine, the memories created from them remain until you delete them.

Making you opt out is in: Computer History isn’t turned on by default, and it doesn’t capture images, video, or audio—which might help OpenAI avoid the controversy that the screenshot-happy Windows Recall faced.

But plenty of others leave it to users to say no. Google emailed customers in June that media they upload to Search, Maps, and other apps could be used for AI training unless they opt out—and the news didn’t even get attention until July. Last week, Twitch told creators they could opt out of Amazon training on their content, which elicited a big, “Wait, since when was it training AI?” reaction from the public. And since December of last year, Meta has used interactions with Meta AI to personalize content and ads, with no full opt-out for US users.

Oversharing era: While the opt-in or opt-out setting is one thing, the larger issue might be how willing some consumers are to hand over sensitive data. Last fall, a Washington Post analysis of almost 50,000 ChatGPT conversations found that users disclosed mental health and medical information, workplace and family disputes—and shared over 550 unique email addresses and 76 phone numbers.

Bottom line: People often share information with agents or chatbots without considering how it could be used later—or decide the trade-off is worth it anyway. —WK