A third confide in chatbots, half don't know it trains the model. That gap is a product spec

The number worth putting in a roadmap is not the confiding statistic. It is the arithmetic sitting next to it. If 53% did not know or were unsure that most chatbots train on conversations by default ^[6], then at most 47% did know ^[15]. Yet 58% said they were uncomfortable with the practice once it was explained to them ^[8]. At least eleven points of that discomfort therefore belongs to people who found out from the survey question ^[16]. Consent that only survives while nobody mentions it is not consent, it is a gap between the settings page and the text box, and the gap is measurable.

The legal exposure figure has the same shape and is larger. DuckDuckGo reports that 25% of respondents knew AI chats can be subpoenaed by law enforcement ^[7], which leaves three in four typing under an assumption the courts do not share ^[17].

Then there is the cohort question, which is where growth and privacy stop being separate teams. Among parents or guardians with children at home who use AI, 43% have told an AI something they had never shared with another person, against 25% of users with no children in the household ^[9][10] - roughly 1.7 times the rate ^[18]. The heaviest disclosure comes from the group defined by living with other people, none of whom clicked accept. Self-described AI enthusiasts report the same behaviour at 56% ^[5], about 23 points above the overall rate ^[19]. The users who like the product most are the ones writing the material that is hardest to defend holding.

DuckDuckGo's framing of why is worth reading with the source in mind: it argues AI companies have encouraged people to treat chatbots as confidants through advertising campaigns and media appearances, while those conversations are less private than users are led to believe ^[2]. DuckDuckGo sells private search ^[3], the survey is its own, and it covers nearly 2,000 US adults ^[1], so this is one self-interested data point rather than a literature. But the mechanism it describes does not need charity. A product that invites long, personal input gets thought processes and communication styles rather than keyword interests ^[13], and nothing about that upgrade is announced at the moment it happens.

The baseline for anyone shipping a confidant-shaped feature is already thin: 39% report no trust at all in large AI companies ^[11], which is better than the 48% who say the same of the US government ^[12] and still means two in five arrivals start below zero. DuckDuckGo's closing line is that the valuable data is your next question, not the ones you asked years ago ^[14]. Read from the other side of the table, that is a statement about how quickly an opt-out actually bites, and how little a corpus assembled under an unread default is worth defending.