How to Avoid Leaving a Digital Trail

Why this is important

Sensitive disclosures carry real-world risk. People naturally talk to AI assistants the way they might talk to a trusted advisor — sharing health concerns, financial stress, relationship problems. But the "cascading inference" scenario is sobering: asking about heart-healthy recipes could get you algorithmically flagged as health-vulnerable, and that classification could eventually influence insurance pricing or targeted advertising. The AI interaction feels private; the data pipeline behind it is not.

Your words can outlive the conversation. When you type something into an AI chatbot, it doesn't just disappear after you close the window. That data can be retained indefinitely, read by human reviewers, merged with your activity on other platforms, and used to train future models — often without you realizing it.

Most people don't read the fine print. The gap between how private these tools feel and how private they actually are is significant.

The practical upshot is simple: treat an AI chatbot the way you'd treat a postcard, not a sealed letter. Useful for many things, but not the place for anything you'd be uncomfortable having read by a stranger.

What NOT to share with an AI chatbot

Based on guidance from the University of Kentucky's ITS office (January 2025) and the Mozilla Foundation's privacy guide:

  • Social Security numbers, driver's license or passport details

  • Usernames, passwords, or login credentials

  • Phone numbers, home addresses, email addresses

  • Medical records, health scans, or personal health details

  • Financial information: bank details, credit card numbers, pay stubs, investment accounts

  • Work-related confidential information

  • Intellectual property or creative work you haven't published

  • Other people's personal information

How to opt out of training data collection

The key setting for each major platform:

  • ChatGPT: Settings > Data Controls > toggle off "Improve model for everyone"

  • Claude: Settings > Privacy > toggle off model training

  • Gemini: Google Account > Data & Privacy > Gemini Apps Activity > toggle off

  • Copilot: Settings > Privacy > toggle off model training for text/voice

Even with training turned off, conversations may still be stored temporarily (typically 30 days) for safety monitoring, and human reviewers may still sample flagged conversations. The Mozilla Foundation's advice: "Don't share anything you wouldn't want another person to see."

Key Research Findings

The landmark Stanford HAI study (October 2025)

Jennifer King, Privacy and Data Policy Fellow at Stanford's Institute for Human-Centered AI, led a study comparing the privacy policies of six leading U.S. AI companies: Amazon (Nova), Anthropic (Claude), Google (Gemini), Meta (Meta AI), Microsoft (Copilot), and OpenAI (ChatGPT). The study analyzed 28 documents (privacy policies, subpolicies, FAQs) and found:

  • All six companies use consumers' chat data by default to train their models.

  • Some developers keep this information in their systems indefinitely.

  • Some companies allow human reviewers to read users' chat transcripts for training purposes.

  • For multiproduct companies (Google, Meta, Microsoft, Amazon), user chat interactions routinely get merged with data from other products the consumer uses on those platforms (search queries, purchases, social media activity).

  • Most companies are not taking steps to remove children's input from training.

King's essential consumer advice: "If you share sensitive information in a dialogue with ChatGPT, Gemini, or other frontier models, it may be collected and used for training, even if it's in a separate file that you uploaded during the conversation."

The "cascading inference" scenario. King illustrates a realistic example: asking a chatbot for low-sugar or heart-friendly recipes could lead the algorithm to classify the user as health-vulnerable. That classification can flow through the developer's ecosystem, potentially influencing targeted advertising or reaching insurance companies over time.

Psychiatric Times analysis (January 2026)

An article titled "Chatbot Privacy Is an Oxymoron" highlighted several developments:

  • In 2024, Italy's privacy regulator fined OpenAI 15 million euros for lack of transparency in ChatGPT's data collection.

  • A 2025 court order in a New York Times copyright lawsuit required OpenAI to retain all conversations, including those deleted by users, until the case is resolved.

  • OpenAI CEO Sam Altman acknowledged in June 2025 that OpenAI is legally required to share "private" conversations if subpoenaed.

  • Once data enters a company's systems, it can be reused without renewed consent.

The Help Net Security report (October 2025)

This report noted that shared ChatGPT conversation links appeared in Google search results after being publicly shared. Hundreds of thousands of Grok conversations were also found indexed by Google. Even "incognito" or "private" modes mostly hide things from the user, not from the company.

Sources