The Dossier Your Chatbot Built Without Asking

There are plenty of ways to spend a Tuesday worrying about artificial intelligence — deepfakes, job losses, the data centers nobody wants next door. Here's a smaller and more immediate one: the chatbot you asked about your toe pain has drawn conclusions about you.

Not stored. Concluded. That distinction is the whole story.

The New York Times published a piece on Jul 23rd by tech columnist Brian X. Chen, who asked ChatGPT and Google's Gemini to disclose what they had figured out about him that he never actually told them. The answers were uncomfortably good. Gemini pegged him as "firmly in the peak 'Logistics Dad' phase of life," which it called "entirely wholesome, but... the definitive end of 'youthful spontaneity.'" It guessed at his income bracket from two data points: the German car he'd asked for repair help with, and the nanny contract he'd used AI to draft. It flagged a long-running foot problem from scattered questions about toe remedies over the course of a year. It even guessed he lived in a nicer neighborhood, reasoning that a motorcycle, a car being worked on and a rusted metal patio table together "requires a garage, a driveway or a dedicated outdoor workspace – amenities that are highly characteristic of single-family homes."

That is not a chat log. That's a dossier.

The mechanism, plainly. Two separate things are happening when you type into a chatbot. The first is retention — the company keeps the text. The second is inference — the model reasons from your text to facts you never volunteered. Privacy settings mostly address the first. Almost nothing addresses the second.

Lindsay Owens, the economist who runs Groundwork Collaborative and previously served as a senior economic policy adviser to Sen. Elizabeth Warren, has been studying the consumer side of this. "The comprehensive nature of the types of information and dossiers it's putting together is extraordinary," she said. "People have never been able to put this level of information together." The commercial worry she raises isn't embarrassment. It's pricing — a company that can infer what you'll pay is a company that can charge it.

Here's the part most people get wrong. Paying doesn't buy you out. On ChatGPT Plus, at $20 a month, your conversations are still used to train models unless you go into Settings, open Data Controls, and switch it off yourself. Policy reviewers who've compared the free and paid individual tiers side by side found the training columns essentially identical. Enterprise and workspace tiers are the ones with contractual privacy boundaries written in. Individual consumers get the toggle — if they find it.

Temporary Chat is the closer thing to a clean slate: no history, no memory, no training. OpenAI says it may still keep a copy for up to 30 days for safety purposes.

On the Google side, Gmail and cross-app permissions now sit inside a Personal Intelligence dashboard in the Gemini interface. Critics have argued Google leaned on the pre-existing "Smart Features" toggle to switch on deeper data access without a clean second consent prompt. If you're weighing Gemini for anything sensitive, default-on access to your full email history is worth knowing before you start.

How many of us are we talking about? Roughly 45% of full-time employees report using ChatGPT for work tasks in the past month, and 62% of college students use it weekly. Among employed U.S. adults, workplace use has gone from 8% in 2023 to 28% — a 3.5x jump in two years. Rhode Island workers, students and parents are inside those numbers. And 82% of users describe their conversations as sensitive or highly sensitive.

The catch is that we're bad at guessing what gets inferred. In a 2025 study posted to arXiv, researchers found users trying to predict which personal attributes a model could deduce performed only slightly better than chance. Nearly half were troubled once shown the inference.

Rhode Island has started legislating here, narrowly. Gov. Dan McKee signed three AI measures on June 22nd: a ban on offering therapy or psychotherapy in the state unless a licensed professional provides it, a requirement that chatbot operators maintain protocols for users expressing suicidal ideation or self-harm, and a rule that health care providers using AI to transcribe in-person or telehealth visits notify patients and review the records for accuracy. Beginning July 1, 2027, chatbot operators must file annual reports with the attorney general counting safety protocol activations, published in aggregate on the AG's website. Violations carry penalties up to $15,000 per day, with fines routed to suicide prevention programs.

What didn't pass matters too. A broader bill covering workplace surveillance and automated employment decisions cleared the Senate and died in the House. "We are just at the beginning of the artificial intelligence revolution," Patrick Crowley, president of the Rhode Island AFL-CIO, said of the effort. "Rhode Island has a chance to be a national leader by creating some common-sense guardrails, so workers do not get exploited by this new technology."

So: no state law tells a chatbot what it may conclude about your income from your car. Until one does, the toggle in Data Controls and a blunt prompt — tell me everything you've figured out about me that I never actually stated — are what you've got. Run it. You may find out you're in the peak Logistics Dad phase of life.