Key Details
The OECD’s Enhancing User Control and Privacy Protections in the Context of AI follows personal data through two routes into AI systems—directly from chatbot interactions and indirectly from public websites—and considers what happens after collection.
Chatbot interactions: Prompts and responses can contain sensitive information. Some providers offer an account-wide choice to exclude interactions from model training, but not a way to exclude a particular conversation or file.
Web scraping: Publicly visible personal data can be collected without direct contact with the people concerned. Notices, limits on collection, filtering and ways to object all matter.
User rights: Requests to access, correct or delete personal data can concern both information supplied to a system and personal information it generates.
Data already learned: Removing a record from a training dataset may not remove information a model has memorised, particularly if versions of that model have been fine-tuned and redistributed.
New forms of access: AI tools can also retrieve information while answering a question or pass data between agents, creating uses beyond the original training dataset.
The brief reviews guidance and emerging practices; it does not establish new binding AI rules.
Chatbot Privacy Choices Need to Match the Conversation
People use chatbots for everyday queries as well as sensitive discussions. They may not realise that what they type could be retained or used to improve a model. Providers’ account-wide training opt-outs offer a choice, but the sensitivity of one exchange may differ sharply from another.
The OECD therefore examines more precise controls for individual conversations, files or categories of information. It also points to just-in-time notices: short explanations shown when someone is about to enter information, rather than left solely in a privacy policy. Whether these measures provide meaningful control depends on clear language, usable settings and the defaults people encounter.
Web Scraping Needs Safeguards Before and After Collection
Information on a public website is not automatically free of privacy obligations when collected for AI training. The brief reviews guidance on avoiding sensitive sources, limiting what is gathered and filtering personal information before it enters a model. Website instructions that tell bots not to scrape content can signal a preference, but their effectiveness depends on compliance.
Protection does not end at collection. Developers can document where data came from, test whether models reproduce personal information and provide routes for people to object or seek redress. The OECD also distinguishes training data from retrieval-augmented generation (RAG), in which a system pulls information from external sources while answering a query. Training-stage filters alone cannot address every risk arising from that live access.
A Right to Redress Must Work Across Outputs and Model Versions
Personal information can sometimes be memorised during training and later reproduced. An AI system may also generate plausible but false information about a person or harmful content involving their data. Effective redress must therefore consider what appears in outputs, not only what remains in a developer’s database.
Deleting a source record does not necessarily erase what a model has learned. Full retraining can be impracticable, while machine unlearning — targeted removal of a data point’s influence — is still difficult to implement and verify at scale. Accountability becomes more complicated when other developers fine-tune or redistribute a model: a request handled by the original developer may not reach every later version.
The same question extends to AI agents that act through other tools or agents. Limits accepted by a user at the first interaction can become harder to maintain as information passes between systems.
Privacy Protection Can Also Support Useful AI Development
The OECD does not frame privacy and access to useful data as opposing goals in every case. Filtering, traceability and privacy-enhancing technologies may allow developers to use or share data more responsibly while reducing exposure of personal information. These measures complement legal obligations and user choice; none substitutes for the others.
What Is Machine Unlearning?
Machine unlearning aims to remove the influence of particular data from an AI model without training the entire model again. It could help address deletion requests, but reliably establishing that the information has been removed remains a technical challenge.
Policy Relevance
The brief does not assess Indian AI services or Indian law. For India, it identifies three linked questions for AI governance:
Choice at the point of use: Can a person understand and control whether a sensitive chatbot exchange is used to improve a model?
Accountability for collected data: Can a developer show where scraped personal data came from, what was filtered out and how objections are handled?
Redress beyond the original system: If personal information appears in an output or a redistributed model, is it clear who can receive the complaint and what remedy is possible?
These are also questions of data quality and trust. Access to more information will not necessarily improve AI if its origins, permitted uses and risks cannot be managed.
Follow the Policy Brief Here: Enhancing User Control and Privacy Protections in the Context of AI.