The right to erasure is one of the clearest promises in the GDPR: ask a company to delete your personal data, and it must comply within 30 days. The problem is that AI systems don't work like databases. Deleting a file is straightforward. Unlearning what an AI model learned about you is something else entirely.
This article covers what right to erasure actually means for AI systems, where the legal obligation and technical reality diverge, and what compliance looks like in practice. If you run a business using AI, or you've asked a company to delete your data from their AI model, this explains what is and isn't actually happening.
How Right To Erasure Applies To AI Systems
Right to erasure (also called the right to be forgotten) is a GDPR clause that lets individuals request deletion of their personal data. The regulation assumes data lives in a file, a database, a spreadsheet. You delete the file, the data is gone. With AI systems, that assumption breaks. When an AI learns from your data, it doesn't store your name and address in a neat row anymore. It absorbs patterns from your behaviour, your preferences, your choices, and encodes them into millions of numerical weights across a neural network.
Under GDPR Article 17, a business must erase personal data if the individual requests it, unless the data is needed for a legitimate reason. But GDPR was written before large language models and recommendation engines became standard. The regulation does not distinguish between storing data and using data to train a model. It treats them as equivalent obligations. That's the legal pressure point. From a compliance perspective, if someone asks you to stop using their data to train your AI, you must respect that request.
The complication: complying with that request often means retraining your entire model from scratch, without that person's data. For a small business using a third-party AI service, that might mean paying a data processor to rebuild the model. For a large company with a custom AI system, it could mean weeks of engineering work and significant infrastructure costs. The obligation is clear. The mechanism for fulfilling it is not.
Why Deleting Data From An AI Model Is Not The Same As Deleting A File
A traditional database stores your data as discrete records. A customer record contains your name, email, purchase history, address. Delete the record, and that data is gone. An AI model, particularly a machine learning model or neural network, does not store your data in discrete chunks. Instead, it compresses information from thousands or millions of records into weights and biases. Your data has been dissolved into the model itself.
Consider a recommendation engine that powers a retail website. It learns from millions of customer browsing sessions and purchases to predict what you might want to buy. Your behaviour contributes to the model's accuracy, but your specific actions are not stored anywhere you can point to and say "delete this.". They are baked into the probability calculations the model makes every time someone visits the site. To remove your influence from that model, the company would need to retrain it using only data from everyone else.
Some companies have experimented with "machine unlearning," a technique that attempts to remove a person's influence from a trained model without retraining from scratch. The idea is promising, but the technology is still experimental. No unlearning method has been proven to be foolproof across all model types. Industry benchmarks suggest unlearning can reduce the accuracy of a model by 2 to 5 percent compared to a clean retrain, depending on how much influence the erased individual's data had on the original model. For a business relying on model accuracy, that trade-off can be expensive.
GDPR Delete My Data AI: What Companies Actually Do
In practice, most businesses respond to a right to erasure request involving AI in one of four ways. The first is the simplest: they delete the raw data from their storage systems but do not retrain the model. The data is gone, but the model still reflects what it learned. This technically complies with GDPR in the narrow sense that they have erased the data, but it does not actually remove the person's influence from the AI.
The second approach is to retrain the model without the individual's data. This is the most thorough response but also the most expensive. A company running a machine learning model on millions of customer interactions might need to spend thousands of pounds and days of engineer time to rebuild it. Most businesses only do this if they receive multiple erasure requests or if the person requests it explicitly.
The third approach is to use a data processor that specialises in handling erasure requests for AI systems. Companies like Clearview AI have faced legal challenges over their practices, but the principle exists: a third party manages the technical work of removing data from the model, usually through retraining or unlearning. The cost is passed to the business, typically as a per-request fee or as part of a compliance service contract.
The fourth approach, increasingly common, is to simply apply the right to erasure to the raw data and metadata while explicitly excluding the trained model. Companies disclose this in their privacy policies: "We will delete your data from our systems, but your data may have contributed to AI models we use. Those models will continue to operate as trained." Some regulators accept this as compliant. Others are beginning to challenge it.
The Legal Reality: How Regulators Are Interpreting Right To Erasure
GDPR enforcement is fragmented, which creates confusion. The EU's data protection authorities (there are 27, one per member state) do not always agree on how right to erasure applies to AI. However, some patterns are emerging. The European Data Protection Board (EDPB) has suggested that erasing raw data while leaving the trained model intact may not satisfy the spirit of the regulation, even if it satisfies the letter.
In 2023, Italy's privacy regulator fined OpenAI 15 million euros partly over concerns that ChatGPT could not demonstrate it had removed individuals' data from its training. The company's response was illustrative: they could not actually show that the data was removed, because their model had absorbed it. They could only show that future training did not include that data. This forced a distinction between historical training and future training that the regulation did not originally contemplate.
National regulators are also starting to impose technical requirements on erasure requests. Some now require companies to produce evidence that the individual's influence has been removed from the model, not just that the raw data has been deleted. This pushes businesses toward retraining or unlearning, both expensive. The trend suggests that "data deletion without model retraining" will become increasingly difficult to defend as a compliance strategy within the next 18 months.
Practical Costs Of Complying With Right To Erasure Requests
What does complying actually cost a business? The answer depends entirely on the size and type of the AI system. For a small business using a third-party AI service like a voice agent, the cost is relatively low. Services like those offered by vendors include data deletion as part of standard compliance. The expense is absorbed by the provider and baked into subscription costs. If your business uses a Voice AI system with a built-in CRM, deletion requests are typically handled automatically: the call recording and transcript are removed from the system within 30 days, and the model does not retain information about that individual from future interactions.
For a mid-sized company running a custom AI system, the cost of honouring a right to erasure request can be substantial. Retraining a recommendation engine or predictive model typically costs between 5,000 and 50,000 pounds, depending on model complexity and the amount of data involved. If a company receives dozens of erasure requests per month (which is realistic for a large retailer or financial services firm), they either retrain frequently at enormous cost, or they invest in machine unlearning technology, which carries its own development and licensing costs.
For a large enterprise, the costs become operational. Financial services companies, particularly those using AI for credit risk assessment or loan recommendations, need to ensure that erased individuals no longer influence lending decisions. This often requires quarterly or bi-annual retrains, building erasure compliance into the model lifecycle. Industry surveys suggest that large financial institutions spend 2 to 10 million pounds annually on AI model retraining driven partly by compliance requirements.
Right To Erasure AI Systems: The Trade-Offs You Need To Know
Before you assume that right to erasure compliance is straightforward, understand the genuine trade-offs. The first is accuracy loss. Every time you retrain a model without a person's data, you lose the signal that data provided. For some models, this barely matters. For others, it erodes accuracy noticeably. A recommendation engine that loses 3 percent accuracy might see click-through rates drop by 8 to 12 percent. That matters to revenue.
The second trade-off is speed. If you use a third-party AI service, erasure requests are usually handled within the service level agreement, typically 30 to 90 days. If you manage your own model, retraining takes time. During that time, your system still runs on the old model, so the erased individual's influence persists. Complete erasure takes weeks, not days.
The third trade-off is that right to erasure does not apply universally. If you have a legitimate legal obligation to retain data (such as anti-money-laundering requirements in financial services, or medical record retention laws), you can legally refuse an erasure request. Similarly, if the data is not actually personal data (it has been anonymised beyond recovery), erasure does not apply. But proving that data is truly anonymised is difficult. Regulators are increasingly sceptical of anonymisation claims.
The honest assessment: right to erasure is a real obligation with real costs. If your business model depends on training increasingly large AI models and you receive frequent erasure requests, compliance will become expensive. You should factor retraining costs into your budget now, or choose AI systems that handle erasure compliance automatically as part of the service.
How To Respond To An Individual's Right To Erasure Request
If someone sends you a right to erasure request involving an AI system, follow this process. First, acknowledge the request in writing within 14 days. Under GDPR, you have 30 days to comply, but acknowledge immediately to show good faith. Second, identify where their data lives: in your database, in your CRM, in training data for your AI model, in call recordings or transcripts, in email archives, anywhere.
Third, delete data from all non-model systems first. This is quick and unambiguous. Delete their records from your database, remove their files from your servers, purge their calls from your system logs. If you use a outbound campaigns platform or any other tool that stores their information, delete it there too. Document what you deleted and when.
Fourth, make a decision about the AI model itself. If it is a third-party service, check your data processing agreement. Most SaaS platforms handle this automatically. If it is a custom model you manage, decide whether to retrain without their data or apply machine unlearning techniques. Document your choice and your reasoning. If you cannot practically retrain (due to cost or technical constraints), document that too, along with the controls you have in place to prevent future influence from their data.
Fifth, communicate clearly with the individual. Tell them exactly what you deleted and what you did not, and explain why if applicable. Tell them when the work will be complete. If retraining takes three weeks, say that. Vagueness erodes trust and invites regulatory scrutiny.
The Growing Complexity: Multiple Models And Data Sources
Most businesses do not run a single AI model. They run multiple systems across different departments. A bank might have a credit risk model, a fraud detection model, a chatbot, and an email classification system. Each one may have learned from the customer's data. Honouring a right to erasure request means identifying and updating every single model, not just one.
This compounds the cost and complexity. A single customer's erasure request might require updates to four or five different models, each with its own retraining schedule and technical requirements. Some of these models may be managed by different teams or even external vendors. Orchestrating the erasure across all of them is non-trivial. Companies dealing with this at scale are starting to invest in data governance tools and AI model registries just to track where each person's data flows.
Additionally, data flows through multiple stages before it reaches a trained model. Raw data is collected, cleaned, transformed, enriched with other signals, and then used to train. Each stage can introduce copies and derivatives of the original data. For a true erasure, you need to trace the data through the entire pipeline and ensure it is removed at every step. This is where most compliance efforts actually fail. Companies delete the raw data and assume the job is done, but derivatives persist in intermediate databases or cached processing systems.
Emerging Technologies For AI Data Deletion
The technical challenges of right to erasure have spurred research into new approaches. Machine unlearning, mentioned earlier, is one. Companies like Clearview AI and others are investing in algorithms that attempt to reverse a model's learning without full retraining. The promise is compelling: delete someone's influence in hours instead of weeks. The reality is less mature.
Most unlearning methods only work reliably on certain types of models. Deep neural networks are particularly difficult. A 2023 benchmark by researchers at Stanford found that no unlearning method achieved perfect accuracy across different model architectures. Some methods lost 5 to 8 percent of model accuracy, which is unacceptable for high-stakes decisions like credit approval or medical diagnosis. The technology will improve, but it is not production-ready for most enterprises.
An alternative is federated learning, an approach where models are trained on data held locally rather than centralised. If a user erases their data from a local device, their influence is automatically not included in future model updates. This is how some mobile phones handle on-device machine learning. For a centralised business, however, federated learning is impractical. It requires significant architectural changes and is most viable for consumer devices and edge computing, not business data systems.
What Compliance Actually Looks Like For Businesses Using AI
Practical compliance with right to erasure requires documenting four things. First, document your data flows: where data enters your system, where it is stored, where it is used, and where it is deleted. Create a data map. This is the foundation of everything else. Without it, you cannot answer a right to erasure request accurately.
Second, document your AI systems: what models you run, what data they trained on, how often they are retrained, and how long it takes to retrain them. Be specific. "We use machine learning to personalise user experience" tells you nothing. "We run a gradient boosting model updated weekly on 12 months of user behaviour data, taking 6 hours to retrain" is actionable.
Third, document your process for handling erasure requests. Who receives them? How do they trigger action? What systems need to be updated? How do you verify deletion? A good process is repeatable and auditable. An auditor should be able to follow your documented steps and confirm that you actually deleted the data.
Fourth, include erasure requirements in your data processing agreements with vendors. If you use third-party AI services or cloud platforms, your contract should specify their obligations to delete data and provide evidence of deletion. Many vendors still do not take this seriously. Push back. A contract that says "deletion happens automatically" is better than a contract silent on the issue.
Real-World Example: A Customer Data Deletion Scenario
Imagine a mid-sized e-commerce company that uses AI to recommend products. The model trains on three years of customer behaviour data. A customer requests deletion of all their data under GDPR. Here is what compliance looks like. Step one: delete their customer record from the database. This takes minutes. Step two: delete their transaction history, browsing history, and email marketing records. This takes an hour.
Step three is the difficult part: update the recommendation model. The model trained on millions of interactions from thousands of customers. This particular customer contributed maybe 500 interactions (orders, clicks, searches). Retraining without those 500 interactions takes about 8 hours of computation time and costs roughly 2,000 pounds in cloud infrastructure. The company must decide: spend 2,000 pounds to comply fully, or spend 500 pounds to delete the raw data and argue that future model updates will exclude them.
Most companies choose the latter for a single request. But if 10 customers request deletion per month, that is potentially 20,000 pounds per month in retraining costs. Over a year, that changes the economics entirely. The company might then choose to retrain less frequently (quarterly instead of monthly) and batch erasure requests, which means the deleted individual's influence persists in the model for up to three months after their request. That is technically non-compliant but practically common.
Preparing Your Business For Right To Erasure Compliance
If you use AI systems, prepare now rather than react when the first erasure request arrives. Start with a data audit. Map every dataset, every model, every integration. Understand retention periods. Identify what data is truly necessary to keep and what is not. Delete unnecessary data proactively. This reduces your erasure compliance burden immediately.
Second, choose AI systems with erasure compliance built in. If you are evaluating new platforms, ask vendors directly: "How do you handle GDPR right to erasure requests? Can you demonstrate deletion? What is the cost and timeline?" Vendors who have clear answers are further along in their compliance thinking. Those who squirm are risks.
Third, budget for retraining. If you run custom AI models, include quarterly or bi-annual retraining cycles in your cost projections. This is not a hypothetical future cost. Regulators are enforcing erasure, and costs will increase over time. Build it into your model now.
Fourth, invest in data governance tooling. Tools that track data lineage, automate deletion, and provide audit trails are no longer nice-to-have. They are becoming necessary. Services that help you manage caller memory and contact data with compliance in mind, for instance, reduce your manual work and lower your legal risk. If you work with custom solutions, your vendor should offer these capabilities.
The Future: Regulatory Tightening And The Cost Of Non-Compliance
Expect enforcement to tighten. Right now, many businesses get away with partial compliance (deleting raw data but leaving models untouched). Over the next two to three years, regulators will crack down on this. Recent fines against OpenAI and other companies are just the beginning. The pattern is clear: regulators are taking AI compliance seriously.
Fines for GDPR violations can reach up to 20 million euros or 4 percent of global annual revenue, whichever is higher. For a company with 500 million in annual revenue, that is 20 million euros per violation. A single unresolved erasure request, if discovered by a regulator, could trigger that fine. The cost of compliance is far lower than the cost of non-compliance.
Additionally, customers are beginning to care about this. Surveys show that 60 to 70 percent of adults want the ability to delete their data from AI systems. If your competitors offer clear data deletion and you do not, it becomes a competitive disadvantage. Being forthright about your data practices builds trust. Being evasive erodes it.
Frequently Asked Questions
If I delete someone's raw data, is my AI model compliant with right to erasure?
Not necessarily. Deleting raw data satisfies part of GDPR, but regulators increasingly expect the individual's influence to be removed from trained models too. If you retain a model that still learned from their data, you have only partially complied. Document your approach and be prepared to justify it.
How long does it take to retrain an AI model after an erasure request?
It depends on model size and complexity. Small models retrain in hours. Large neural networks can take weeks. Expect retraining to take anywhere from one day to four weeks, plus the time to identify and delete the data in the first place. Budget for full compliance to take 30 to 90 days per request.
Can machine unlearning fully remove someone's influence from an AI model?
Not reliably yet. Unlearning techniques can approximate removal but do not guarantee it. Some loss of model accuracy is typical. For high-stakes applications like credit decisions or medical diagnosis, full retraining remains the safest approach.
What should I include in my data processing agreement with AI vendors?
Require the vendor to specify: their obligation to delete data on request, the timeline for deletion, how they will verify deletion, any costs involved, and what happens to trained models. Get it in writing. Vague clauses about "best efforts" are not sufficient.
Does right to erasure apply if the data has been anonymised?
Only if it is truly anonymised beyond recovery. GDPR defines anonymisation strictly: the data must be irreversibly stripped of identifying information. Most data companies claim as anonymised is actually pseudonymised and remains subject to erasure rights. Be conservative in your interpretation.
How should I acknowledge an erasure request?
Acknowledge it in writing within 14 days. Specify what you will delete and your timeline. If you need to retrain a model, explain that and give an expected completion date. Silence or delays invite regulatory complaints.
What if my business model requires keeping AI training data indefinitely?
You will need to either retrain models frequently to honour erasure requests, invest in unlearning technology, or switch to a federated learning architecture. All three are expensive. This is a real constraint on business models that depend on large centralised datasets.
