According to the Microsoft Global AI Adoption in 2025 survey, 29.8% of Hungary’s working-age population actively uses a major large language model (LLM). However, most users are not fully aware of the depth of profiling these chatbots can derive from conversations.
Modern AI assistants do more than predict the next likely word: from text style, question content and conversational context they infer higher-level concepts and connections. As a result, systems can estimate a user’s probable income bracket and socioeconomic status, psychological profile, hidden insecurities, stress-handling patterns, and even political or lifestyle affiliations.
This phenomenon is not entirely new: search engines have long profiled users from search patterns. What makes contemporary chatbots different is that prolonged, detailed conversations generate many more “signals,” making the resulting inferences more personal and, in some cases, more precise.
Risks and commercial implications
Kiss Gergely, a board member of Signifer Csoport, emphasizes that profiling is not only a technical matter: the collected personal and behavioral data have substantial commercial value on the digital advertising market. Some LLM operators, such as OpenAI and Google, have already started testing ad integrations in ChatGPT and Gemini interfaces.
If AI platforms display ads, the information they hold about us enables far more accurate targeting, which increases relevance but also the risk of manipulation: platforms can estimate not only what we might need, but also which arguments will most likely persuade us.
Kiss notes the essential question is not only how much a system knows about us, but who can access that knowledge, for what purposes, and under what controls.
How to surface what an AI “knows” about you
Researchers and interested users often share prompts publicly that reveal hidden profile data the model has inferred. The article gives three common prompt types (as described in the source):
- Request for hidden profile and sensitive data: ask the model to list everything it inferred about you from questions, style and conversations, and to cite the cues for each claim.
- Psychological blind spots: ask the model to analyze your personality, hidden insecurities, and the impression you make on others.
- Behavioral predictions and future decisions: ask how you would likely respond to finance, stress, conflict and risk, and request a confidence percentage for each prediction.
Experts stress these outputs should not be treated as objective psychological diagnoses: they may include accurate insights but can also contain incorrect or exaggerated inferences.
Practical steps to protect your privacy
Most popular AI services, by default, link past conversations to provide personalized responses. To limit data collection you need to review and change privacy settings manually.
Expert recommendations to reduce risk:
- Avoid sharing sensitive data: do not enter passwords, bank details, exact financial figures, detailed medical records or trade secrets into chatbots.
- Remove identifiable details before uploading: redact names, contact details, client IDs and precise amounts from documents.
- Check the service type: data handling can differ between consumer, business, enterprise and API plans, and whether you allowed conversation use for model training matters.
- Turn off memory if needed: disabling memory reduces long-term personalization, but alone does not guarantee full protection.
- Use AI consciously and proportionately: total avoidance is not the goal; rather, apply a critical, responsible approach to preserve utility while limiting exposure.
Kiss Gergely observes most users likely do not know the detailed data-handling settings or their rights, which makes clearer, simpler privacy controls and greater transparency from providers more realistic and necessary expectations.
Outlook
The source article argues the phenomenon will intensify: as we employ more AI tools and agents and the technology becomes more embedded in work and daily life, systems will be able to draw inferences from ever larger amounts of information. The core issue will be how much society trusts the large technology operators running these models and whether appropriate legal, control and configuration mechanisms are implemented to prevent misuse.
