Tell an AI chatbot something personal and you have deliberately placed information into a conversation. Put a photograph or biographical detail on a public webpage and a different system may collect it without ever speaking to you. A new international policy paper argues that both routes pose serious questions about AI data privacy, because being volunteered or publicly visible is not the same as granting unlimited permission for every future use.
The OECD policy brief published on 15 October 2026 examines how personal information reaches artificial intelligence systems and what individuals can realistically do to control it. It covers information people submit directly, material gathered from websites and other data-acquisition routes. The paper does not announce a new law, order companies to delete their models or claim that every chatbot trains on every conversation. Its subject is the gap between existing privacy ambitions and the practical ways complex AI systems are built and used.
AI data privacy begins before the model gives an answer
A chatbot interaction can feel private because it resembles a one-to-one conversation. A person may ask for help with a difficult family situation, a work problem or a letter containing details about someone else. The provider’s actual retention, training and sharing practices will depend on the product and settings, not on the conversational appearance of the interface. For that reason, the moment information enters a service is already a privacy decision, whether or not the resulting answer seems harmless.
The second route is less visible. Public websites can contain names, photographs, posts, employment histories and fragments of personal circumstances. Some of that material may be collected into datasets for AI development. Public access does not automatically eliminate data-protection rules or every reasonable expectation about reuse. Context matters: something published to help neighbours organise an event was not necessarily written with automated profiling in mind.
The OECD paper maps these routes rather than collapsing them into a single accusation. Systems can also receive information through licensed datasets, partnerships or connected services. Each path raises different questions about notice, consent or other lawful grounds, purpose, security and the ability to object. A useful policy response must distinguish them. Telling everyone to avoid chatbots would not address indiscriminate website collection, while blocking a web crawler would not protect details a user willingly uploads elsewhere.
Why deleting a conversation is not always a simple technical answer
People reasonably expect a meaningful way to remove information they no longer want held. With ordinary stored records, the task may be relatively straightforward: locate the record and delete it under an applicable policy. AI systems can create harder questions when some information has been used during model development, incorporated into other datasets or reproduced in outputs. Researchers and regulators are examining when an individual record can be removed and how the result can be verified.
That does not mean deletion is impossible in every system, or that every prompt has become permanent model knowledge. Products differ widely. The important distinction is between deleting an accessible chat history, stopping future use of retained content and changing a model that may have learned from training data. Those are separate technical and legal tasks. A control that addresses one should not be advertised as though it necessarily covers all three.
The brief discusses emerging approaches including privacy-enhancing technologies, clearer data governance and ways to express preferences about collection. A machine-readable signal asking a crawler not to use a page is a potentially valuable tool, but it works only if the relevant parties recognise and honour it and if there are appropriate enforcement mechanisms. Technical controls cannot by themselves decide every question of legal fairness.
European regulators have already been examining the same tension
The issue is not appearing in a regulatory vacuum. The European Data Protection Board's December 2024 opinion considered questions about personal data in AI models, including circumstances in which a model may be regarded as anonymous and the grounds on which personal data might be processed. It is a regulatory opinion, not a finding that all models are unlawful. The detail of the assessment matters, including whether information can be extracted and which safeguards are in place.
For readers in Europe, there is a crucial difference between a general international recommendation and enforceable national or European obligations. The OECD can draw attention to common problems and suggest approaches; it does not administer the EU’s General Data Protection Regulation. Conversely, a company cannot assume a public webpage is an unrestricted supply of personal information merely because collecting it is technically easy.
LiveAIWire has previously examined how the UK's privacy regulator approaches AI agents. Systems that take actions across accounts can amplify the issue because data do not necessarily stop at the first application. A personal assistant may pass information between email, calendar, shopping and other tools to complete a task. That can be useful while creating more points where access permissions and the original purpose of collection need to be understood.
What ordinary users can actually do now
The most useful first step is to treat a chatbot like an outside service, not a confidential notebook by default. Check its data controls before entering identifiable or sensitive information, particularly details belonging to another person. When a task allows it, describe the problem without including names, account numbers or unnecessary documents. A request to improve wording rarely needs every personal fact from the original correspondence.
For information already on public websites, examine privacy settings and the audience for older posts. Removing a page from ordinary visibility can limit further casual exposure, though it cannot guarantee that every prior copy or dataset disappears. Website operators can investigate available crawler controls and the terms under which material is reused. These steps reduce some risks without supplying a universal solution to an international data market.
LiveAIWire has also explored the privacy implications of AI agents retrieving personal details. That reporting offers a practical reminder that system behaviour depends on permissions and context. The most effective defence is often to reduce what an assistant can access in the first place. A powerful tool does not require unrestricted entry to every part of someone’s digital life simply because it can connect to those services.
The OECD contribution is to make the two halves of the problem visible together. Information we offer in a conversation and information gathered about us from a webpage can both help an AI system produce useful results. They can also both travel beyond the context in which they were supplied. Better privacy will mean clearer explanations of those paths, controls that achieve what they promise and obligations that still operate when data are technically easy to collect.
That is less dramatic than an announcement that chatbots have suddenly been banned from reading the internet. It is also closer to the decisions facing people and policymakers. The practical question is how someone can discover where their information went, understand the reason it was used and exercise a meaningful choice about what happens next.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.