The right to erasure — sometimes called the right to be forgotten — is among the most substantive rights in modern privacy law. Under GDPR, you can ask a company to delete your personal data, and they’re generally obligated to comply. Under CCPA, California residents have similar rights. More than twenty US states now have privacy laws with deletion provisions.
The question is whether these rights work when your data has already been used to train an AI.
In April 2026, the FTC resolved a case in which OkCupid had shared three million user photos with a facial recognition company called Clarifai. The settlement required Clarifai to delete the photos and any AI models trained on them. The case offers a useful window into what deletion actually means in the context of trained AI systems — and what it doesn’t.
What the Right to Erasure Says
GDPR Article 17 gives individuals the right to request that a data controller erase personal data concerning them “without undue delay” when the data is no longer necessary for its original purpose, consent is withdrawn, or the data was unlawfully processed.
The regulation applies to personal data — information that relates to an identified or identifiable person. Your name, your email address, your photos if they depict you — all clearly qualify.
The right isn’t absolute. There are exemptions for freedom of expression, legal obligations, public health, archiving purposes, and legal claims. But for most consumer data held by apps and platforms, the right is enforceable.
What the right says, on its face, is clear enough: delete the data. What it means in the context of machine learning is considerably less clear.
How AI Training Works (and Why Deletion Is Hard)
Training an AI model is not like copying a file. When a model is trained on a dataset of photos, it doesn’t store those photos. It extracts statistical patterns from them — relationships between pixels and labels, features that distinguish one class of image from another, and in the case of facial recognition, the geometric relationships between facial features across many examples.
These patterns are encoded in the model’s parameters: millions or billions of numerical weights that together define how the model responds to inputs. The original photos aren’t in the weights. The information extracted from the photos is.
This creates a conceptual problem for the right to erasure. You can delete the original training photos. You can delete the model file — the weights. But you cannot, with current technology, surgically remove what a model “learned” from specific training examples.
This isn’t a hypothetical concern. Machine learning researchers have investigated techniques like “machine unlearning” — methods for updating a trained model to remove the influence of specific training examples without full retraining. These techniques exist in research settings but remain technically difficult, computationally expensive, and imperfect.
What Actually Happened in the Clarifai Case
The FTC’s settlement with OkCupid and Clarifai required:
- Deletion of the approximately three million photos transferred by OkCupid
- Deletion of any AI models trained on those photos
Clarifai complied. The photos were deleted. The models were deleted.
But “deleting the model” means deleting the file containing the model’s weights. Clarifai could, in principle, train a new model on different photos — photos it holds legitimately — that learns similar facial recognition capabilities. The deletion of the model file doesn’t erase the capability the company developed by building that model.
The OkCupid settlement doesn’t require Clarifai to be incapable of facial recognition. It requires the deletion of specific assets derived from improperly obtained data. The boundary between “knowledge extracted from your photos specifically” and “general capability Clarifai has developed” doesn’t map onto anything that can be deleted.
How Regulators Are Approaching This
European data protection authorities have been the most active in testing the right to erasure against AI systems.
The Italian Data Protection Authority (Garante) suspended ChatGPT’s operation in Italy in March 2023, citing among other concerns the question of how OpenAI handles erasure requests for data used in training. The suspension was temporary, but it established that AI training data practices fall within GDPR’s scope.
The Spanish AEPD has issued guidance indicating that AI models trained on personal data may themselves constitute personal data processing, with implications for erasure rights. The French CNIL has investigated AI training practices and issued guidance on the conditions under which training data use is lawful.
The emerging regulatory consensus is that:
- Using personal data to train AI models requires a lawful basis under GDPR
- Where erasure rights apply to the training data, companies must delete that data
- The question of whether trained model weights constitute personal data — and whether they must be modified or deleted on erasure request — remains contested and under active regulatory development
The Technical Proposals
Several approaches have been proposed to address the gap between “delete the training data” and “remove the influence of the training data”:
Full model retraining. The most robust approach: delete the training data and retrain the model from scratch without it. Computationally expensive. For large models trained on billions of parameters, potentially prohibitively so.
Machine unlearning. Targeted algorithms that update model weights to reduce or remove the influence of specific training examples. Promising in research settings, but unproven at scale and difficult to verify.
Data minimisation. Avoiding the problem by not training on personal data that may later be subject to erasure requests. The GDPR’s data minimisation principle points in this direction — collect and use only the data you need.
Differential privacy. A mathematical framework for training models in a way that limits what can be inferred about any individual training example. Strong differential privacy provides formal guarantees about how little influence any single example has on the model. Not universally applicable and involves accuracy trade-offs.
None of these solutions is fully satisfactory for the general case. Full retraining is the clearest path to genuine compliance with erasure requests, but it’s also the most expensive. Machine unlearning is promising but unverified. Data minimisation prevents the problem but limits what AI systems can be built.
What You Can Actually Request — and Verify
If you suspect your photos or personal data were used to train an AI without your consent, you have several avenues:
File a subject access request. Under GDPR, you can ask any company operating in the EU to tell you what personal data they hold about you and how it’s used. This includes whether your data was used for AI training. Companies operating under CCPA have similar obligations for California residents.
File an erasure request. You can formally request deletion of your personal data. If the company trained AI on it, the response should address what steps they’ve taken to address that use — including whether the training data has been deleted and what has been done about models trained on it.
Lodge a complaint. If a company fails to honour a valid erasure request, you can complain to the relevant data protection authority — in the EU, to the national DPA in the country where the company is established or where you reside.
What you can verify is limited. You can confirm that a company has deleted your data from its servers (to the extent you can verify anything about a company’s internal systems). You cannot verify that a trained model no longer reflects patterns learned from your data. The technical infeasibility of that verification is one of the core unsolved problems in AI compliance.
Why This Is Getting More Urgent
The volume of personal data being used to train AI is growing. Every photo you upload to a consumer app, every document you share with a productivity tool, every voice recording processed by a transcription service — all of this has potential value as training data, and many of the companies that hold your data have strong incentives to use it.
At the same time, regulatory pressure is increasing. The EU AI Act creates new obligations for AI systems, including requirements related to training data transparency. National data protection authorities are actively investigating AI training practices. The OkCupid case in the US demonstrates that even without strong financial penalties, enforcement actions can result in mandatory deletion.
The right to erasure isn’t a complete solution to the problem of AI training on personal data. But it’s the most concrete legal tool available, and using it — or supporting the legal frameworks that make it enforceable — matters for establishing norms around how companies treat personal data in AI development.
What Privacy-First Storage Changes
One of the most practical responses to the AI training data problem is choosing platforms whose business model doesn’t create incentives to use your data for training.
daftei stores your files — photos, documents, voice notes — in an encrypted vault. Every file is encrypted at rest with AES-256 and in transit with TLS 1.3. daftei doesn’t use your content to train AI models for third parties. There’s no advertising business that creates an incentive to extract value from your data.
When you delete files from daftei, there’s a 30-day grace period followed by permanent, irreversible erasure — not a perpetual backup retention period like many cloud services maintain. daftei is GDPR and CCPA compliant, meaning your erasure rights are enforceable.
The 5 GB free tier lets you evaluate the service without a financial commitment. Unlimited storage is available on Pro at $5.99/month or $44.99/year.
The Honest Summary
The right to erasure works reasonably well for traditional data storage: you ask a company to delete a file, they delete it, it’s gone. It works less cleanly in the context of trained AI models, where “deletion” can mean deleting training data without fully addressing what the model learned from it.
Regulatory frameworks are catching up, and the pressure on AI companies to implement genuine technical solutions to erasure compliance is increasing. But for now, the most meaningful protection is upstream of the erasure right: choosing carefully which platforms get access to your personal data, and preferring platforms whose business model doesn’t turn your data into training material in the first place.
The right to forget is important. The ability to not be used in the first place is better.