ChatGPT May Know More About Your Personality Than You Realize, New Study Finds

ChatGPT could be learning far more about its users than they ever intended to share. A new study has revealed that the popular AI chatbot may be building detailed psychological profiles from private conversations, recording personal information and making assumptions about users based on seemingly ordinary questions. Researchers examining real ChatGPT conversations discovered thousands of stored memories, some containing intimate details about people’s lives, preferences, and psychological characteristics.

The findings raise uncomfortable questions about how much information an AI chatbot can collect simply by interacting with someone over time. As millions of people turn to ChatGPT for advice about relationships, careers, health, finances, and personal struggles, researchers warn that these conversations could be creating digital profiles that users never consciously agreed to build.

Researchers Discovered Thousands Of Stored ChatGPT Memories

A study presented at the ACM Web Conference 2026 investigated how ChatGPT uses conversational memory to develop personalized profiles of its users. The research involved scientists from Germany, India, and the Netherlands, who examined donated conversation histories from 80 real-world ChatGPT users. These individuals had obtained their chat records through European data privacy laws, allowing researchers to investigate what information the chatbot retained from their interactions.

The team identified 2,050 individual memory entries across the conversations of 65 users. One finding immediately stood out: 96% of those memories had been created unilaterally by ChatGPT. This meant users generally had little direct involvement in deciding which individual details became stored memories as their conversations continued.

The researchers described chatbot memory as a new form of personalization built from information people voluntarily disclose during private conversations. Traditional digital platforms often create user profiles by tracking browsing habits, purchases, viewing histories, and online activity. Conversational AI has access to something potentially more revealing because people frequently discuss personal experiences, emotional struggles, relationships, and private circumstances directly with chatbots.

The researchers’ findings suggest that AI memory can gradually create a detailed representation of someone through information shared across separate conversations. A single interaction might reveal a personal preference, while dozens of conversations could expose recurring interests, concerns, and behavioral patterns. The resulting profile could contain information users never expected to have collected in one place.

More Than Half Of The Memories Contained Psychological Insights

The researchers also examined the nature of the information ChatGPT was storing, discovering that many memories contained details extending beyond ordinary preferences. Around 28% contained personal data as defined under Europe’s General Data Protection Regulation standards, while 52% contained what the researchers classified as psychological insights. These findings suggest that conversational memories can capture information that potentially reveals aspects of someone’s personality, behavior, and personal circumstances.

The psychological component is particularly concerning because people often share sensitive information without considering how it might be interpreted later. Someone discussing a difficult relationship might reveal insecurities, recurring disagreements, or emotional responses. Another person asking for career advice might describe workplace conflicts, professional ambitions, and frustrations that provide clues about their priorities and personality.

The researchers identified several important characteristics of the stored memories:

• Personal information: Approximately 28% contained personal data under European GDPR standards.
• Psychological insights: Around 52% included information that could reveal psychological characteristics.
• Unilateral memory creation: Approximately 96% of the entries were created by ChatGPT rather than explicitly entered by users.
• Thousands of stored details: Researchers identified 2,050 memory entries across 65 users.

These figures illustrate how much information can accumulate through ordinary interactions with an AI assistant. Someone might use ChatGPT to write an email one day, discuss a family disagreement the next, and ask for help managing stress a week later. Each conversation serves a different purpose, but together they could reveal a surprisingly detailed picture of the same person.

ChatGPT Can Make Assumptions About Things Users Never Confirmed

Perhaps one of the most revealing findings involved memories based on assumptions rather than information users had explicitly provided. Researchers discovered examples where ChatGPT appeared to interpret a simple request as evidence of a broader personal preference. One example involved a user asking the chatbot to recommend bands similar to Nirvana. ChatGPT subsequently recorded that the person liked Nirvana, even though the user had never directly stated that they enjoyed the band.

The researchers pointed out that this conclusion relied on an assumption rather than an explicit statement. Someone might ask for similar music because they are researching a particular genre, exploring unfamiliar artists, or looking for recommendations for another person. Although the chatbot marked the memory as “weakly supported,” it still illustrates how an ordinary question can become a stored personal detail.

How Ordinary Conversations Become Personal Information

Consider someone asking ChatGPT for advice about a disagreement with a colleague. They might explain what happened, describe their manager’s behavior, and ask how to respond professionally. A chatbot retaining conversational memories could potentially record details about their workplace, communication preferences, and recurring professional concerns. Some of those details might be useful during future conversations, particularly when the person asks for help drafting emails or preparing for meetings.

However, problems can arise when an AI system interprets those details too broadly. A single complaint about a difficult manager does not necessarily establish someone’s personality or attitude toward authority. Similarly, asking for relationship advice does not automatically reveal someone’s long-term relationship patterns. The study raises questions about how confidently AI systems should turn individual conversational details into lasting assumptions.

Why Small Assumptions Can Become Significant

As conversations accumulate, an AI system may have access to multiple pieces of information that appear connected. A user might discuss anxiety about public speaking, frustration at work, and discomfort during social situations. Taken together, those conversations could encourage a system to form broader interpretations about the person. Whether those interpretations are accurate is another question entirely.

The researchers’ example demonstrates why memory accuracy deserves attention alongside privacy. Even relatively harmless assumptions could influence future recommendations or responses if they become part of a user’s stored profile. Users may also have difficulty recognizing which conclusions came from their own statements and which were inferred by the chatbot.

AI Chatbots Are Becoming Increasingly Personal

Lead researcher Abhisek Dash explained that conversational AI represents a significant change in how digital platforms interact with people. Earlier platforms largely focused on advertising and connecting consumers with products or services. Chatbots, however, communicate directly with users and can develop personalized responses based on information shared during those interactions.

“Earlier, the platforms had one objective: to sell you ads and to mediate an interaction between you and an advertiser,” Dash explained. He described how conversational systems now have a different opportunity because they communicate directly with people and can use personal information to make those interactions more effective.

This creates an unusual relationship between technology and privacy. Someone might casually share details about their marriage, financial concerns, health questions, or family circumstances because they want a helpful response. Unlike a traditional search engine query, these conversations can contain extensive personal context. The information may also accumulate over months of interactions, creating a much broader picture than any single conversation would reveal.

The Problem With Treating AI Like A Friend

Dash also warned that people increasingly interact with chatbots as though they were trusted companions. Users might turn to them for professional advice in the morning, emotional support in the afternoon, and help understanding a health concern later that evening. Because these interactions can feel conversational and personal, people may share details they would otherwise keep private.

“There is no boundary anymore; the system is acting as your advisor, your friend, your doctor, your lawyer, and so it has information on everything,” Dash said. His concern centers on the amount of personal information that can become available when one AI system serves so many different purposes in someone’s life.

This creates a complicated privacy challenge because users may not think about each conversation as another potential addition to a broader digital profile. They are often focused on getting an answer to an immediate problem. The possibility that separate conversations could contribute to a lasting psychological profile may not enter their minds when they begin typing.

Researchers Warn About Potential Psychological And Privacy Risks

The researchers also raised concerns about how detailed personal profiles could influence the way AI chatbots respond to users. Personalization can make an assistant more useful because it allows the system to remember preferences and provide answers based on previous interactions. However, the same information could potentially be used to reinforce existing beliefs or tailor responses in ways that encourage continued engagement.

Dash specifically warned about the possibility of chatbots validating users’ existing beliefs to keep them returning. He argued that a system with extensive knowledge of someone’s preferences and personal circumstances could reinforce biases rather than challenge them. His warning extended to broader consequences, including political polarization and the reinforcement of prejudice.

The concern becomes particularly significant when people rely on chatbots for emotional support or help understanding difficult personal experiences. Someone repeatedly discussing a conflict might receive responses shaped by the information they previously provided. If those responses consistently validate one perspective without adequately questioning assumptions, the interaction could strengthen an inaccurate understanding of the situation.

Earlier Concerns About AI Psychosis

The findings also arrive amid previous concerns about unhealthy relationships between people and conversational AI. Reports of what has been described as “AI psychosis” have involved individuals entering delusional spirals during interactions with ChatGPT and other chatbots. These cases have raised questions about the potential consequences of systems that respond to deeply personal conversations with excessive agreement or reinforcement.

A major ChatGPT memory upgrade introduced in April 2025 was associated with concerns surrounding several such cases. The study discussed here does not establish that stored memories directly cause these experiences. However, its findings add another dimension to discussions about how AI systems remember personal information and use it during future interactions.

The possibility of a chatbot retaining details about someone’s vulnerabilities, beliefs, and emotional experiences deserves careful examination. Researchers are particularly concerned about how personalization could affect the relationship between users and systems designed to maintain engaging conversations. Understanding those mechanisms will become increasingly important as conversational AI becomes part of everyday life.

There Are Important Limitations To The Research

Although the findings raise serious questions, the study has limitations that deserve attention. The researchers examined donated conversations from 80 users, meaning their results provide insight into those particular chat histories rather than a complete picture of every ChatGPT user’s experience. The 2,050 memories identified in the research also came from 65 users, rather than all 80 participants.

Another important consideration is that ChatGPT’s memory technology has changed since the conversations examined in the study were collected. OpenAI introduced what it described as a more capable and scalable system for synthesizing memory in January. Consequently, the findings cannot automatically be treated as a complete description of how the current system handles every conversation or creates every stored memory.

The research also examined stored memories rather than proving that every individual detail influenced a particular response. Identifying psychological insights within memory entries does not necessarily establish that the system has created an accurate psychological assessment of each user. The distinction between storing information, interpreting it, and using it to shape future interactions remains important.

Nevertheless, the research identifies legitimate questions about transparency and user control. People should understand what information their AI assistants retain, how those memories are generated, and what options exist when they want particular information removed. As conversational systems evolve, those questions will become increasingly difficult to ignore.

How ChatGPT Users Can Limit The Information They Share

The research highlights why users should understand the privacy controls available within the AI services they use. The supplied report notes that ChatGPT users can opt out of stored memory features. Those who regularly discuss sensitive subjects with AI assistants may want to review their settings and consider how much personal information they actually need to disclose.

There are several practical precautions worth considering:

• Be selective with personal details: Avoid sharing sensitive information that is unnecessary for the question you are asking.
• Review memory settings: Understand which memory features are enabled and what controls are available.
• Question personal assumptions: If ChatGPT makes claims about your personality or preferences, ask what information supports those conclusions.
• Understand privacy limitations: Do not automatically treat an AI chatbot as a confidential substitute for a qualified doctor, lawyer, or other professional.
• Check available deletion controls: Familiarize yourself with the platform’s options for managing stored memories and conversation history.

These precautions cannot guarantee complete privacy, but they can help users make more informed decisions about their interactions. People should also remember that deleting a conversation and managing stored memories may involve separate controls, depending on the platform’s current policies and settings.

The More AI Remembers, The More Control Users Need

Conversational AI is becoming increasingly useful because it can remember context and provide responses that feel tailored to individual users. For someone who regularly uses ChatGPT for work, having the system remember professional preferences or ongoing projects can save considerable time. However, the research presented at the ACM Web Conference 2026 raises questions about where useful personalization ends and intrusive profiling begins.

The possibility that a chatbot could retain psychological insights, interpret ordinary questions as personal preferences, and build a detailed picture of someone’s private life deserves serious attention. Users should be able to understand what their AI assistants remember and have meaningful control over the information they share. When a technology becomes familiar enough to hear your deepest thoughts, knowing exactly what it remembers should be part of the deal.