What is machine unlearning? And can you get your data out of an AI model?

The Morse teamPublished November 2, 20268 min read

Machine unlearning is the research field that tries to remove what an AI model learned from specific data after training, without retraining the model from scratch. As of November 2026, none of the four chatbot companies we checked offers it to you: OpenAI, Anthropic, Google, and Meta describe deleting a chat or opting out as keeping your data out of future training, and Anthropic and Meta say plainly that a model already trained on it stays as it is.

We read the companies' own help and privacy pages on November 2, 2026: three OpenAI Help Center articles, two from Anthropic's Privacy Center, Google's Gemini Apps Privacy Hub, and Meta's explainer on how its generative AI models work. We also read the GDPR's erasure article, the European Data Protection Board's 2024 opinion on AI models, a German regulator's notice on Meta, and the research that defines the field. Anthropic deserves credit for saying in one sentence on its retention page what the others leave you to work out.

Machine unlearning: key facts

What is machine unlearning?

Machine unlearning is a way to make a trained model forget particular data, so it behaves as if it had never seen it. Google Research, announcing the first machine unlearning challenge on June 29, 2023, called it "an emergent subfield of machine learning" that aims to remove the influence of a set of training examples it calls the "forget set" (Google Research). The challenge ran in the NeurIPS 2023 competition track, hosted on Kaggle.

The sure way to forget is to retrain the model without the data. Google's post calls that approach straightforward but says it isn't always viable, because "retraining deep models can be computationally expensive." Unlearning methods try to get close to the retrained model at a fraction of the cost.

Can you remove your data from an AI model?

Not in any way the big chatbot companies offer you. Anthropic says it most directly. If you delete a Claude chat, "it will not be used to train future models," but "your data will still be included in model training runs that are already in progress, or in models that have been trained" (Anthropic). That page is dated July 1, 2026.

Meta says the same about public posts. "If you delete a piece of information a model learned from (like a public post), the model won't change in that moment," its explainer reads, though the information itself is deleted and "won't be used for future training" (Meta). What Meta trains on from your Instagram account is in does Instagram use your photos to train AI.

A data protection regulator put it more bluntly. When Meta began training on European users' public posts on May 27, 2025, Hamburg's data protection commissioner told users that if their data has already been used, "this cannot be reversed by subsequently objecting," because training data is "irrevocably incorporated into artificial intelligence models" and "given the current state of technology, its influence cannot be removed from the model" (HmbBfDI). An objection, the notice adds, "is effective with regard to future training."

What does deleting a chat do to each company's models?

Anthropic and Meta say deleting changes nothing in models already built, and OpenAI's and Google's pages don't say either way. What happens to the stored copy of a chat in each app is in are AI chats private. The table sets each company's own words on training side by side, as read on November 2, 2026.

Anthropic's switch reaches back further than OpenAI's or Google's, covering "previous or new chats" for any training run that hasn't started, and which AI companies train on your data lists each app's switches.

What does a "right to be forgotten" request to OpenAI do?

It keeps information out of ChatGPT's answers and leaves the model as it is. OpenAI's help article says that under laws like the GDPR, you can ask it "to stop certain personal information about you from appearing in ChatGPT responses," through a Remove my personal data from ChatGPT responses request in its Privacy Portal (OpenAI).

The request has conditions. OpenAI asks for "a government-issued ID or alternative proof," and says each request involves "a contextual balancing of the individual's privacy rights against other relevant interests, including freedom of expression and the public interest." It weighs whether you have a public role, whether the information is accurate, how old it is, and how sensitive it is.

OpenAI explains why the request works on answers rather than weights. Its article on how its models are developed says they "do not store or retain copies of the data they are trained on," and that individuals in certain jurisdictions can object or make other requests through the Privacy Portal or dsar@openai.com (OpenAI). Anthropic makes the same point about Claude, whose models "do not have access to or pull from the original training data once the models have been trained," and warns that rights requests about its training data are "limited" and that the process is "complex" (Anthropic).

Does the GDPR's right to erasure reach a trained model?

European regulators haven't settled it, and this isn't legal advice. Article 17 of the GDPR gives you the "right to erasure ('right to be forgotten')," and requires a controller to erase personal data "without undue delay" on grounds that include an objection under Article 21 or unlawful processing (EUR-Lex).

Whether a model holds personal data at all is the open question. The European Data Protection Board's opinion of December 18, 2024 says whether an AI model is anonymous "should be assessed on a case by case basis," and that it counts as anonymous only when it is very unlikely that anyone could identify the people in its training data or "extract such personal data from the model through queries" (EDPB). It adds that training on unlawfully processed data "could have an impact on the lawfulness of its deployment, unless the model has been duly anonymised."

The companies say their models keep patterns rather than copies, and the regulators say that has to be shown model by model. Until that's settled, Hamburg's answer holds in practice. An objection works for future training.

Does machine unlearning work on chatbots?

In research, partly. In 2023, Ronen Eldan and Mark Russinovich of Microsoft made Meta's Llama 2 7B model lose its ability to produce or recall Harry Potter content "in about 1 GPU hour of finetuning," against more than 184,000 GPU hours for the original pretraining, while its scores on common benchmarks stayed "almost unaffected" (arXiv). The paper's own title calls this "approximate unlearning."

Checking that a model really forgot is the hard part. Google's challenge post says the evaluation of forgetting algorithms "has so far been highly inconsistent," and that as a result "we are unable to make direct comparisons between different unlearning methods in the literature" (Google Research).

None of the OpenAI, Anthropic, Google, or Meta pages we read on November 2, 2026 mentions unlearning as something a user can ask for. What they offer are deletion, opt-outs for future training, and, at OpenAI, a filter on what ChatGPT says about you.

How Morse keeps your chats out of training

Once your words are in a training set, no setting, form, or regulator above can promise to pull them back out, so the only fix that works is for them never to get there. Morse doesn't train AI on your content, sell your data, or show ads (privacy policy), so Morse never adds your words to a training set.

There's no opt-out that only counts from today, because there was never an opt-in. There's no de-identified copy left in a training pipeline after you delete, because your chat history is stored in your browser, encrypted, not on Morse's servers. Deleting a chat or clearing your browser removes it from your device for good. Morse can't recover it either, so keep anything you'll need.

Each chat has its own privacy mode, chosen by how far you want the words to travel:

• Private, for open-weight models and Grok, on hosts that keep nothing. It's the default for a new chat. Morse has zero data retention agreements with its inference, search, and other providers, so DeepSeek, Kimi, GLM, and Grok answer without the host keeping your words. Where a provider confirms zero retention on each reply, Morse checks it and refuses a reply that doesn't. What is zero data retention explains the arrangement.
• E2EE, for anything no one should ever read. Your browser checks a secure enclave's hardware attestation before anything leaves your device, then encrypts the message so only that enclave can read it. Morse's servers relay text they can't read. This rests on the enclave hardware and attestation working as designed. How end-to-end encrypted AI chat works covers each check.
• Anonymous, for GPT, Claude, and Gemini without the provider knowing it's you. Morse's proxy removes your account, email, and IP address and sends the message under Morse's own account. The provider sees the words without knowing they're yours and may keep them for a limited time to check for abuse, under its own policy, so leave out names and details that identify you.

Because we don't store chats, we usually have very little to give authorities.

What to do before you type

• Something you'd never want in a model: don't send it to an app that trains by default. In ChatGPT, Claude, and Gemini, turn training off or use a temporary or incognito chat first, since deleting afterwards doesn't reach models already trained. On Morse, start the chat in E2EE.
• Something ChatGPT already says about you: file the Privacy Portal request, with links to the chats and the reason the information is inaccurate, excessive, or out of date.
• Public posts on Facebook or Instagram: object as early as you can, since an objection counts only for future training. Use the forms in the Hamburg notice or the steps in how to turn off Meta AI.
• Everyday questions you'd rather keep: ask them in a Private chat on Morse, where there's no training to opt out of.

Try a chat from the Morse home page without an account, or start free with no card. End-to-end encrypted chats come with Pro, at $19 a month.