Every The AI Downside article filed under Privacy — the failures, dark patterns, price hikes and privacy grabs we've logged here, newest first. 9 and counting.
Consumer ChatGPT uses your chats to train OpenAI’s models unless you opt out — a switch that lives in two places, doesn’t reach data already used, and which some users report finding turned back on.
Meta remotely disabled the cameras on thousands of Ray-Ban smart glasses after users covered the recording light. It is a fair privacy safeguard — and proof the camera you paid for switches off when Meta decides.
Instinct, the invite-only AI assistant now valued near $2.5bn, sent an email nobody approved, kept surfacing a tester's inbox after access was revoked, and claims a perpetual licence to your data. The demo feels like magic; the fine print is the deal.
From 18 August 2026, ChatGPT predicts whether you're under 18 from how you use it and quietly switches suspected minors to a restricted version. Opting out of the guess means verifying your age with a selfie or government ID. The safety goal is real; the mechanism puts the cost of a wrong guess on you.
From 17 August 2026, Atlassian uses your Jira and Confluence content to train its Rovo AI by default. Free, Standard and Premium users can switch off the content but not the metadata — the complete opt-out is an Enterprise-only setting.
Anthropic’s new watermark embeds an invisible statistical signal into every word Claude generates — and every word it edits. Text you wrote yourself and ran through Claude for a light proofread now carries a mark that says ‘processed by AI.’ The EU AI Act is the reason. Your own prose is the collateral.
Twitch has added a setting letting you opt out of Amazon training generative AI on your channel — streams, clips, chat and all. It is switched on by default, and Twitch’s own product chief admits opt-in wouldn’t work: nobody would choose it.
Your conversations, files and prompts are the raw material for the next model. Whether you knew you were donating them is a matter of default settings.
Persistent memory makes chatbots feel personal. It also builds a quiet, growing dossier you did not quite agree to and cannot easily see.
