Can You Make AI Forget You? Removal, Corrections and the Right to Be Forgotten

Most of this journal is about persuading machines to say your name. This essay is about the opposite request, and the people who make it: the wrong fact, the old story, the address that should never have been public. Can a machine that learned about you be made to unlearn?

The question deserves a more honest answer than either the privacy optimists or the fatalists give it. There are real levers, real laws, and real wins. There are also hard technical limits that no form on any website can overcome. Let us take the levers one at a time, and be precise about what each actually does.

What are you actually asking a machine to do?

First, a distinction that determines everything downstream. An AI system holds you in two places. There is the model itself: patterns about you compressed into billions of parameters during training, a process that ended months before you asked. And there is retrieval: the live web pages and search indexes the system consults while answering. These respond to removal completely differently. Retrieval can forget you quickly, because deleting or delisting a page removes it from what the machine reads tomorrow. The trained model cannot, in any practical sense, because your data is not a file inside it that someone can locate and delete. It is a diffuse statistical influence spread through the weights. Researchers call the attempted cure machine unlearning, and the honest state of the art is that fully retraining a large model without your data is the only rigorous method, at a computational cost so enormous that no company does it in response to an individual request. What companies actually deploy are filters: the model still holds the pattern, but a layer above it declines to output your details.

Can you remove information from ChatGPT?

You can formally ask, and the route is more established than most people realise. OpenAI operates a privacy portal that accepts, among other requests, "remove my personal data from ChatGPT responses," alongside data subject requests by email. Its own help documentation on the right to be forgotten and personal data removal from ChatGPT sets out the deal plainly: under laws like GDPR you may request erasure or object to processing, particularly where information is inaccurate, excessive or no longer appropriate; requests are assessed case by case, balancing your privacy against freedom of expression and public interest; and the right is not absolute, so OpenAI may decline where it judges there is lawful reason.

Read the small print of what success buys you, because OpenAI states it candidly: a granted request stops your personal data appearing in ChatGPT responses. It does not remove you from external websites or search engines, and it is suppression at the output layer rather than surgery on the model. Two practical additions. If you run a website, blocking OpenAI's GPTBot in robots.txt keeps your pages out of future training crawls, a decision with real trade-offs we weigh in Should You Block AI Bots From Your Website? And expect the public-figure threshold to matter: the more legitimately public your professional life is, the more the balancing test tilts toward the information staying.

What can Google actually remove?

Google offers two very different doors, often confused. The first is "Results about you," a tool for everyone: it monitors Search for your contact details, phone number, email, home address, and since a February 2026 expansion, government-issued ID numbers, and lets you request removal of results exposing them (Google's documentation). It works, it is fast by removal standards, and its limit is stated in its own help pages: it removes results from Google Search, not the pages themselves. The information stays on the source site for anyone, or any crawler, that arrives another way.

The second door is European and older: the right to be forgotten delisting process, born of a 2014 Court of Justice ruling and now anchored in GDPR's Article 17 right to erasure. Through Google's dedicated process, people in Europe can ask that results about them be delisted from name searches where the information is inaccurate, inadequate, irrelevant or excessive, again subject to a public-interest balancing test. Two structural limits matter. Delisting applies to name queries, not to the page's existence in the index generally. And following a 2019 Court of Justice decision, Google applies these delistings within Europe only; searchers elsewhere see the result as before. A right that stops at a border is a real right, but a regional one.

Does the right to be forgotten apply to AI at all?

Increasingly, yes, and regulators have drawn blood. GDPR's Article 17 gives people in Europe a right to erasure of personal data, with exceptions for freedom of expression, journalism, legal obligations and public interest. European authorities have made clear they consider AI companies bound by it: Italy's data protection authority, the Garante, fined OpenAI 15 million euros in December 2024 after finding, among other things, that it had processed personal data to train ChatGPT without an adequate legal basis (Euronews' report). The direction of travel is unmistakable: personal data in AI systems is regulated data.

But a right is only as strong as its enforcement mechanism, and here the technical reality bites back. A regulator can order erasure; it cannot make retraining cheap. So compliance, in practice, converges on the same toolkit: output filters, exclusion of your data from future training runs, and deletion from the retrieval layer. Whether that constitutes genuine erasure under Article 17 is a live legal argument that will run for years. For you, today, the operational summary is: the law gets you suppression and future exclusion, reliably in Europe, unevenly elsewhere, and the word "forget" in every headline, including ours, is a metaphor.

Three different things you might actually be trying to erase

People say make AI forget me as if it is one request aimed at one thing, but it usually splits into three genuinely separate targets, and confusing them is why so many removal attempts feel like they half-worked. The first target is a search engine's index, like Google Search, which is really just a giant, regularly refreshed pointer list to pages that exist elsewhere. The second target is a browsing-enabled chatbot's live retrieval, the kind of assistant that checks the current web before answering, similar in spirit to how a search engine works, just wrapped in conversation. The third target, and the hardest by far, is a model's frozen training weights, the compressed statistical patterns baked in during a training run that already finished. Each of these three responds to a removal request on a completely different timeline, through a completely different mechanism, and understanding which one you are actually dealing with changes what a realistic outcome looks like.

Table: the same word, removal, means three very different things depending on which layer you're actually asking to change.

Imagine someone named Farid who discovers an old, inaccurate blog post claiming he was involved in a lawsuit he was never part of. He successfully gets Google to remove the page from search results in his region and, separately, persuades the original site to take the post down entirely. Within a couple of weeks, a browsing-enabled assistant like a search-connected chatbot stops surfacing the claim too, because the source it would have cited is simply gone. But if a general-purpose model without live browsing was trained on a web snapshot that included that old post before it was removed, that specific model may still repeat the false claim from memory, because deleting the source after the fact does not reach back into a training run that already happened. Farid's fix worked completely at the search and live-retrieval layer and not at all inside that one older model, until a future retraining eventually happens to exclude it, which is not something he can request or schedule directly. The same frozen-weights mechanic, running in reverse, is what makes a model slow to learn new true facts about you too, a pattern explored in this piece on knowledge cutoffs and your name, since a training cutoff cuts both ways: it is exactly as reluctant to add new correct information as it is to drop old incorrect information.

The five levers, and what each one honestly does

Notice the pattern running down the right-hand column: every lever is partial, regional, or forward-looking. Stack them and you get meaningful suppression. No stack gets you oblivion.

When correction beats removal

Here is the strategic turn most removal-seekers miss, and where this essay rejoins the rest of this journal. If what the machines hold about you is wrong, rather than private, erasure is usually the weaker play. Suppose you win: the false claim is filtered, the page delisted. What remains is a vacuum with your name on it, and generative systems abhor a vacuum. The next answer about you gets assembled from whatever survived, which may be older, thinner, or wronger, and when nothing survives, models have a documented habit of inventing plausible filler, the failure we anatomised in When AI Invents Your Credentials.

Correction, by contrast, gives the machine something better to say instead. Fix the fact at its source, publish the canonical version on your own site, and let the corrected record propagate through the next crawl. The step-by-step version of that play is in What To Do When AI Gets Facts Wrong About You, and running exactly this triage, deciding what to suppress, what to correct and what to out-publish, is part of what we do for clients. The rule of thumb we use: removal for genuinely private data, safety risks and doxxing; correction for everything a professional actually needs fixed.

So, can you make AI forget you?

You can make it stop repeating you, in specific products, in specific regions, with persistence and paperwork. You can keep your future self out of training sets you control the gates to. In Europe, the law will increasingly take your side. What you cannot do, and should stop expecting, is press delete on a distributed memory spread across model weights, mirrored pages, and a dozen companies' indexes. Digital forgetting is not an event; it is a maintenance discipline, the same discipline as visibility, pointed the other way. The machines' memory of you was always going to need a keeper. The only question is whether it gets one.