The Organisation for Economic Co-operation and Development published a policy brief on Oct. 15 on strengthening user control and privacy protections in AI. It finds that AI agents will make long-standing data protection principles harder to apply, while also suggesting agents could help people manage their privacy.
The brief, “Enhancing user control and privacy protections in the context of AI,” looks at how personal data enters AI systems and what regulators and companies are doing about it.
Where agents come in
As AI systems move beyond chat toward agents that act on a user’s behalf, the OECD says regulators and industry will need to work out how user control applies to them. These systems “may challenge the application of existing data protection concepts, such as consent and purpose limitation, designed for direct user interactions,” the brief says. Purpose limitation is the rule that data collected for one use shouldn’t be repurposed for another.
The brief singles out handoffs between agents. “Consent is particularly at risk of degrading across agent-to-agent handoffs,” it says, since the scope a user originally approved gets harder to enforce at each step.
It lists other risks: agents inferring or exposing confidential information across services, and reusing personal data from prompts in ways users don’t expect. Deleting data “will be particularly challenging in agentic/multi-agent systems, where personal data flows across models, tools, and orchestration layers,” the OECD writes.
The brief also sees an opening. AI agents could act as “privacy-supporting tools” that help users enforce their own preferences and apply automated safeguards across services. It suggests future OECD work could study both sides, possibly through use cases.
The rest of the brief
Most of the document covers two sources of training data. One is people’s conversations with chatbots. The other is web scraping, the automated copying of public web pages, according to the brief and a summary by the Digital Watch Observatory.
The brief notes that many chatbot providers let users opt out of having conversations used for training, but those controls usually apply to all activity at once. Users generally can’t exclude a single conversation or file, it says.
It also addresses data memorized inside a model. Retraining a large model from scratch to remove personal data is often unfeasible, the OECD says, which has drawn interest in machine unlearning, techniques that remove specific data without full retraining. The brief calls that research promising but still hard to carry out and evaluate at scale.
The Digital Watch summary notes that the brief reviews guidance from privacy regulators in Canada, the Netherlands, Australia, Singapore, the EU, France and the U.K. France’s CNIL has built a model genealogy tool to trace fine-tuned versions of open models that may hold memorized personal data, according to the brief.
The OECD closes with recommendations for policymakers. They include regulation that accounts for the technical limits of AI systems and research into unlearning and other privacy-enhancing technologies. The OECD also calls for simple no-code controls that let everyday users manage how systems use their data, along with closer dialogue among regulators, researchers and civil society.
One item specifically calls for exploring AI agents “to serve as privacy defenders that dynamically enforce protections.”
Source: OECD, “Enhancing user control and privacy protections in the context of AI”
Additional sources: OECD (full brief, PDF), OECD, “Enhancing user control and privacy protections in the context of AI”; Digital Watch Observatory, “OECD publishes Policy Brief on user control and privacy in AI”
