AI Chatbots Are Evolving to Remember More About You, Raising Privacy Concerns
As AI assistants gain memory capabilities and connect to more user data, understanding and managing privacy settings becomes increasingly critical.
Artificial intelligence chatbots are evolving to retain more information about user interactions, leading to enhanced personalization but also raising significant privacy questions. While previously the primary concern was whether conversations were used for model training, AI companies are now giving assistants the ability to remember past exchanges and even draw context from activity across multiple services.
This increased memory and connectivity allow AI assistants to build a more detailed understanding of users over time. The extent to which this impacts privacy depends on the specific service and the user's chosen settings. This information can personalize future responses, influence recommendations, and even shape the advertisements users see.
AI Memory and Personalization Beyond Training
Turning off AI model training does not necessarily disable memory functions or all forms of personalization. Companies like OpenAI and Meta have separate controls for model training, memory, and advertising personalization. For instance, OpenAI offers users controls for model training, memory, and advertising personalization independently. Meta also outlines its own rules regarding how interactions with Meta AI influence content and advertisements.
Specific Company Approaches to AI Privacy
OpenAI's ChatGPT introduced 'Computer History' for macOS on August 13, allowing the AI to use selected app and website activity as context. This optional feature, available to Pro, Business, and Enterprise users, records interaction events instead of screenshots. It is off by default, and users can choose which apps and websites contribute, pause collection, inspect, and delete their history. OpenAI states that temporary interaction-event files are retained on the Mac for up to 48 hours for processing and are not used for training, while generated memory files are stored as unencrypted Markdown files on the user's device until deleted.
Google Search's 'Search Services History' can now save media uploaded or created during search interactions if the 'Save Media' setting is enabled. Google states this helps personalize user experiences and improve AI models, but claims media selected for AI training is disconnected from the Google Account first. Saved media might remain in training systems for up to four years after being disconnected from an account. Users can disable the 'Save Media' setting to stop future media from being saved or used for training, but previously saved media must be deleted separately.
Meta announced in 2025 that interactions with its generative AI features could influence the content and ads users see across its products, effective December 16, 2025, in most regions. For example, discussing hiking with Meta AI could lead to seeing related ads or recommendations. Meta states it does not use AI conversation data related to health or political views for ad personalization, nor does it use religious views or other sensitive categories for this purpose.
Meta also introduced 'Incognito Chat' for Meta AI on WhatsApp and its standalone app on May 13, 2025. This feature is designed to process conversations in a secure environment that Meta cannot access, with messages not being saved and disappearing by default.
Apple's 'Apple Intelligence' system processes many AI requests on-device. For more complex tasks, it uses 'Private Cloud Compute,' which Apple states receives only the necessary information, does not store it, and does not retain it after the request is completed. If users opt to send requests to third-party services like ChatGPT through Apple's integration without logging in, OpenAI reportedly does not receive IP addresses, store requests, or use them for training. However, connecting a ChatGPT account means OpenAI's existing account settings apply.
OpenAI began testing ads in ChatGPT in the U.S. on February 9 for eligible Free and Go users. The company states that ads can consider the context of the current conversation, and with personalized ads enabled, past chats and memory can contribute to ad relevance. OpenAI asserts that advertisers do not receive user conversations, chat history, or memories and that the company does not sell user data to advertisers. Advertisers receive aggregated performance information. Users can disable ad personalization, though ads may still use the immediate chat context.
The Patchwork of AI Privacy Regulation
The United States currently lacks a comprehensive federal privacy law governing data collection and use by all companies. Existing sector-specific laws and a growing number of state privacy regulations apply, and consumer protection laws are relevant to AI companies. The Federal Trade Commission has warned AI companies about the legal consequences of privacy and confidentiality promises, investigating deceptive or unfair data practices.
Managing AI Privacy Settings
Users can take steps to enhance their privacy when using AI tools. These include being cautious about sharing sensitive information, utilizing temporary or private chat features like ChatGPT's Temporary Chat or Meta's Incognito Chat, and regularly reviewing AI memory settings. For ChatGPT, users can review and delete remembered information. Opting out of model training is also an option. For Google users, reviewing and adjusting Search Services History and the 'Save Media' setting is recommended. Periodically reviewing connected apps and services and checking advertising controls separately from training settings are also advised steps to manage the data AI services can access and utilize.