Can an AI forget you?

Short answer: only partly. You can delete your chats, wipe an assistant’s saved memory, stop your future conversations being used for training, and ask for your personal data to be filtered out of its answers. What you cannot do, with any major provider as of September 2026, is have your data removed from a model that has already been trained on it. The information is not stored anywhere you could point to. It has been absorbed into the model itself.

That gap matters because the law says you should be able to ask. The GDPR’s right to erasure applies in the EU and, through the UK GDPR, in Britain. European regulators now say a trained model is not automatically anonymous. So the right to be forgotten runs into a system that was never built to forget.

The five layers of an AI system, and which you can clear

Most of the confusion comes from treating an AI system as one thing. It is at least five, and they forget in completely different ways.

The first three are ordinary data, and ordinary deletion works. The fourth is a filter: the information is still in the model, but a layer in front of it stops it being said. The fifth is the hard one. The UK’s Information Commissioner’s Office put it plainly after its consultation on generative AI: output filters “may not be sufficient, as they do not actually remove the data from the model”.

What the right to be forgotten actually says

Article 17 of the GDPR gives you the right to have personal data about you erased “without undue delay” on six grounds, including where the data is no longer necessary, where you withdraw consent, where you object and there is no overriding reason to keep processing it, and where it was processed unlawfully. It has exceptions, the most important being freedom of expression and information, legal obligations, public health, research and legal claims.

The phrase “right to be forgotten” comes from a 2014 case, Google Spain v AEPD. The Court of Justice of the EU ruled that even accurate information can become “inadequate, irrelevant or no longer relevant, or excessive” over time, and that a search engine must then stop showing it. The remedy there was delisting: the page stayed online, it just stopped being found.

That is almost exactly what AI companies offer today. OpenAI’s own removal page uses the same vocabulary, asking whether information is “inaccurate, excessive, irrelevant, or no longer appropriate”, and the result of a successful request is that the information is prevented from appearing in ChatGPT’s responses. The right to be forgotten, applied to AI, has become a right to be filtered.

Is a trained model personal data?

This is the question everything turns on, and European regulators have not all agreed.

• The Hamburg data protection authority argued in July 2024 that the mere storage of a large language model is not processing of personal data, because no personal data is stored in it. On that view, your rights apply to what goes into an AI system and what comes out, but not to the model.
• The European Data Protection Board disagreed in its Opinion 28/2024, adopted in December 2024. It said AI models trained on personal data “cannot, in all cases, be considered anonymous”, and that personal data “may still remain ‘absorbed’ in the parameters of the model”. A model only counts as anonymous if extracting personal data from it, directly or through prompts, is insignificantly likely.
• France’s CNIL said in February 2025 that people can exercise their rights over training data and the model itself, unless the model is anonymous, and accepted output filtering as a fallback where erasing from the model is too hard.
• The UK’s ICO expects developers to have clear and effective processes for people’s rights across training data, outputs “and the model itself”.

The technical evidence sides with the board. In 2020, researchers extracted verbatim names, phone numbers and email addresses from GPT-2, including text that appeared only once in its training data. A model that can recite something has, in any meaningful sense, remembered it.

What each AI company lets you do

Every major provider now lets you delete conversations and control whether they are used for training. None offers removal from a model that has already been trained.

The fine print is revealing. Anthropic says that if you turn training off, your data “will still be included in model training that has already started and in models that have already been trained”. Google keeps Gemini chats for 72 hours even with activity switched off, and chats already read by human reviewers are kept for up to three years, disconnected from your account. Deleting is real, but it only ever works forwards.

Meta is the odd one out, because it trains on what people post publicly, not only on chats. After pausing in 2024, it resumed training on public content from EU users in May 2025, with an objection form that Ireland’s regulator required it to make easier to use. Objecting keeps your posts out from then on.

When the AI gets you wrong

Erasure is only half the problem. The GDPR also says personal data must be accurate, and gives you the right to have mistakes corrected. Chatbots make things up about real people.

In April 2024 the campaign group noyb complained on behalf of a public figure whose date of birth ChatGPT kept getting wrong. OpenAI’s answer, according to noyb, was that correcting it was technically impossible, and that it could only filter or block answers. In March 2025 noyb filed a second complaint for a Norwegian man whom ChatGPT had described as having murdered two of his children, a story mixed with true details about his family and hometown. Both complaints went to the Irish Data Protection Commission, which leads on OpenAI in Europe. As of September 2026, neither has been decided.

Enforcement has been slow and fragile. Italy’s data protection authority fined OpenAI €15 million in December 2024 over how ChatGPT was trained and run. In March 2026 a court in Rome annulled the fine, not on the merits but on jurisdiction, because OpenAI had set up in Ireland, making the Irish regulator its lead authority.

Why a model cannot simply unlearn you

The obvious fix is to retrain the model without your data. For a large model that means months of computing time and costs no company will pay for one request. So researchers have spent years on machine unlearning: ways to make a trained model behave as if it had never seen a piece of data.

• Exact unlearning splits training into pieces so only the affected piece needs retraining. The 2019 SISA method did this, but its speed-up fell to 1.36 times on the largest task tested, too little for a modern language model.
• Approximate unlearning nudges the model away from the information. In 2023, researchers made a model that took more than 184,000 GPU-hours to train effectively stop recalling Harry Potter after about one GPU-hour of fine-tuning.
• It does not hold. The TOFU benchmark, built from 200 fictional authors in 2024, found that none of the methods tested showed effective unlearning. Later in 2024, researchers showed that fine-tuning an unlearned model on loosely related public text could bring the forgotten material back. Their conclusion was that these methods obscure knowledge rather than remove it.

That leaves filters, which suppress rather than delete. A filter can be worked around, it can fail, and it has to know what to look for. It is a promise not to repeat something, not a promise to forget it.

How to reduce what AI holds about you

1. Turn off training in each assistant you use, so your future conversations stay out of future models.
2. Delete old chats and review saved memory. Memory is stored separately from chats, so deleting a conversation does not always delete what the assistant remembered from it.
3. Use temporary or private chat modes for anything sensitive. OpenAI says temporary chats are not used for training and do not create memories.
4. If an AI says something false or harmful about you, file a privacy request with the provider, and keep a copy of what it said and when. OpenAI’s removal form works without an account.
5. Complain to the company first, then the regulator. In the UK, organisations have had a legal duty since 19 June 2026 to handle data protection complaints and acknowledge them within 30 days. After that, the ICO. In the EU, your national data protection authority.

Forgetting was never designed in

The right to be forgotten assumes information lives somewhere: in a record, a file, a row in a database you can find and delete. A language model breaks that assumption. What it learned about you is not stored in one place but spread thinly across billions of numbers, mixed in with everything else it knows. There is nothing to point to.

That is the thread running through this whole site: every memory system has to forget, and the ones that do not forget on purpose forget by accident. Human memory fades on a curve. Computer memory leaks unless it is refreshed. AI models were built to accumulate and never to let go, and the law is now asking them for a capability nobody engineered. Until someone does, the honest answer to “can an AI forget you?” is that it can be told to stay quiet.

This essay explains the law and the technology. It is not legal advice. Every provider policy was checked against the provider’s own pages in September 2026, and they change often.