ChatGPT and Gemini can infer personal details fast, so test what they know

It can be unsettling to realize how much Gemini and ChatGPT can infer about you, and how easily privacy can be punctured. The New York Times Tech guide is blunt about the starting point: you should not assume the systems are clueless. You should find out what they have figured out about you, with a method you control.

The core of the article is simple: use prompts to probe the model and see what information it surfaces or implies. Think of it like a security audit, not a vibe check. If the responses reveal personal details, patterns, or associations you did not explicitly provide in the chat, that is a signal that your data footprint may be more exposed than you assumed. The unsettling part is not just the content of any one answer, but the speed and ease with which a general-purpose model can respond.

Why this matters now is that both ChatGPT and Gemini sit at the center of a broader arms race in intelligence and distribution. These models are built to be helpful. In practice, that means they are optimized to interpret context, connect dots, and produce coherent answers. When you prompt them, you are not just asking for information. You are providing a setting where the model can infer. For executives, this turns privacy from an abstract compliance checkbox into a measurable operational risk: if the system can plausibly infer something about you, it can plausibly infer something about your employees, customers, or partners.

There is also a regulatory and governance angle that decision-makers cannot ignore. Privacy law in many jurisdictions is moving toward stronger expectations around how personal data is collected, processed, and protected. Even when a specific model behavior is hard to map to a single legal theory, regulators and boards tend to ask the same questions: what is being learned, what is being stored or reused, and what controls exist to limit unintended disclosure. A key second-order implication here is that governance teams will likely want evidence, not anecdotes. The Times guide, by focusing on how to test, supports that mindset. It is the difference between saying, “This feels risky,” and demonstrating whether the risk shows up.

Then there is the incentive problem. Organizations want these tools to be broadly usable, and users want convenience. That often leads to prompts that include personal or sensitive details, sometimes accidentally. In a workplace setting, that can happen in everyday workflows: drafting emails, summarizing customer conversations, creating internal documentation, or troubleshooting issues. The models can only respond with what they see in your prompt and what they can infer from patterns. But from a privacy standpoint, the result can still be the same: answers that reveal more than intended.

For boards and senior leaders, the stakes expand beyond individual discomfort. If a model can infer or surface personal details, that behavior can become a reputational issue and a customer trust issue. It can also turn into an internal risk issue if staff do not understand what is appropriate to share. In many organizations, the “model policy” is written down, but the actual practice is driven by what employees think is safe. A testing approach, like the one described in the Times piece, can help close that gap.

So what should an executive do with this? Start with the same mindset the article implies: treat the model as an unknown surface area until you probe it. Use prompts to check what comes back when you provide certain details, and observe whether the output changes in ways that suggest personal knowledge. Then decide what training, guardrails, and escalation paths are needed based on what you find. The second-order question is not just “Does it know something?” but “Does it know enough to matter for our people, our customers, and our obligations?”