The Algorithmic Right to Be Forgotten in Insurance

II.AI Absorbs Instead of Archiving

Data dissolved into the model

A machine-learning model absorbs and transforms its training data instead of storing it. During training, personal information spreads across millions of interconnected parameters. There it mixes, combines and reconfigures until it can no longer be separated from the rest³. The data is metabolized. It feeds what the model knows without surviving in an identifiable form, like food that can no longer be traced in the body it helped build. A database keeps content you can point to. A model keeps only weights, abstract coefficients in which each person's contribution is diluted. Asking such a system to forget a person means asking it to find something that, by design, has no location, and to remove a contribution now fused with all the others.

Everywhere and nowhere

This dissolution leaves personal data in an odd position, everywhere and nowhere at once. Nowhere, because it sits in no specific location that could be erased. Everywhere, because research on model inversion and membership inference shows that training information can be rebuilt from the model, so no one can claim it is gone⁵. The data is neither present, as in an archive, nor absent, as after an erasure. It survives in a diffuse form that can be rebuilt. This third state disarms the right to be forgotten. The right planned for two states, presence and absence, and has no answer to diffusion.

Generative models make this harder still. A storage system returns what it was given. A generative system transforms and recombines what it absorbed into new outputs. A person's trace can therefore resurface in a derived, recombined form, sometimes unrecognizable, that still carries the original influence. No simple procedure erases data that no longer exists as such yet whose effects spread into new model outputs. The right to be forgotten was built for stable content. Here it meets a shifting material that dissolves the data and then regenerates it in unpredictable forms, which pushes complete erasure even further out of reach.

III.Machine Unlearning and Its Limits

The promise of machine unlearning

About a decade ago, a field emerged to address this problem. Machine unlearning covers techniques that let a model unlearn specific data without full retraining. The idea is to adjust the model's internal parameters to strip out the influence of the targeted data, which saves the time and computing power a full rebuild would take⁶. Early research showed that, under certain conditions, an artificial intelligence can be made to forget part of its data. Certified unlearning frameworks go further. They use probabilistic bounds to guarantee mathematically that the influence of the removed data has been reduced enough⁸. Machine unlearning looks like the technical answer to the legal problem, a form of erasure suited to how diffuse data is.

The limits of an imperfect forgetting

The promise runs into serious limits. No universally recognized method yet exists to tell whether a model has actually forgotten. For the largest and most complex models, guaranteeing total forgetting is close to impossible⁷. Machine unlearning also faces entanglement. One person's data is never isolated. It is mixed with the data of thousands of others, so removing it without degrading the model or affecting other people's data is a hard balance to strike⁴. Full retraining remains an option, wiping everything and rebuilding from scratch, but its computing cost is often prohibitive and grows with model size⁹. Machine unlearning therefore delivers real but partial erasure. It is approximate, hard to prove, and far from the clean deletion the law assumes.

A challenge that goes beyond Europe

The problem goes beyond European law. It arises wherever a statute grants a similar right. The California Consumer Privacy Act provides a right comparable to erasure. New rules on artificial intelligence, including the EU AI Act ramping up from August 2025, add governance requirements that assume control over the data models absorb¹⁰. Together these frameworks set a worldwide expectation of erasure that the technology struggles to meet everywhere. The tension is structural. It sits between what the law aims for and how machine learning is built, and no single text is to blame. A simple regulatory fix will not solve it. Lawmakers will have to rethink what forgetting can mean once data lives inside models.

IV.An Absolute Right, an Approximate Forgetting

Where the law and the technology part ways

The result is a basic gap between the law and the technology. The law promises binary, absolute erasure, where the data either exists or does not. The technology can only offer gradual, probabilistic forgetting, where the data's influence shrinks with no certainty that it is gone. Lawyers think in all-or-nothing terms. Engineers think in degrees and error margins. The two views rarely meet. Applied to models, the right to be forgotten turns from a right to disappear into a right to be attenuated. The shift is quiet but consequential. It drains part of the right's substance while the wording of the text stays the same, and the person concerned is not necessarily told.

The impossible verification

Verification makes the gap worse. A right is worth something only if compliance can be checked, and here it is extremely hard to confirm that a model has forgotten a person. Neither the individual nor the regulator has a simple, reliable way to confirm the erasure happened. In practice, an absolute obligation becomes a best-efforts obligation. Some have concluded that the right to be forgotten is dead and that data now lives forever in artificial intelligence. The phrase is provocative, but it points to something true². The right survives on paper, but enforcing it in a verifiable way keeps getting harder. For models, it becomes a right to be forgotten approximately and without proof.

That is no reason to give up, but it calls for clear thinking about what the right can deliver. Promising erasure the technology cannot provide weakens the law's credibility and breeds distrust. Admitting the limits of forgetting openly lets effort go where results are possible. The worst stance is to keep up the fiction of an absolute right while tolerating, in practice, that it goes unenforced. That devalues the right and protects no one. Clear-eyed regulation would accept that machine unlearning works by degrees. It would require the best erasure achievable, documented and auditable, instead of an impossible total disappearance, along with safeguards on how the model can still be used. The right to be forgotten will keep real force if the requirement matches technical reality instead of clinging to an unreachable ideal.

V.The New Exposure for the Firm and Its Insurer

A compliance risk that no control closes

This gap creates a concrete exposure for any organization that deploys models trained on personal data. If a firm cannot technically honor an erasure request, it faces regulatory sanctions, with fines that can reach a significant share of global turnover, plus litigation and injunctions. What sets this risk apart is that no control fully closes it. Machine unlearning is still imperfect and retraining is prohibitive, so the organization keeps a residual risk it cannot remove. Exposures that no technical measure can eliminate at the source are the ones that end up retained on the balance sheet or transferred to insurers.

The insurer, exposed on both sides

The insurer is involved on two counts. First, it may carry the risk through D&O, cyber or tech E&O cover, which can respond when a failed erasure leads to a sanction or a dispute. Second, it is a data controller itself. Insurance is one of the heaviest users of models trained on personal data, for pricing, risk scoring, fraud detection or claims forecasting, topics this series has covered from other angles. The insurer may have to pay for a failed erasure and is exposed to the same failure in its own operations. That uncomfortable position forces it to take the question seriously on both sides of the contract.

VI.From Erasure to Data Minimization

Don't absorb what you may have to forget

The real answer probably lies upstream, in data governance, more than in endlessly refining after-the-fact erasure. Since no one knows how to reliably remove what a model has absorbed, the sensible course is to avoid absorbing what you may have to forget. That means data minimization, pseudonymization and aggregation of training data wherever possible, and privacy-preserving learning techniques that limit the personal footprint in the model¹¹. Systems should be designed by asking from the start which data may legitimately be used and whether people's rights can be technically enforced, especially for the most sensitive models covering health, behavior or risk scoring. Restraint on input makes forgetting possible on output. The best protection against an erasure request is never to have absorbed the data in question.

Rethinking erasure for models

This shift also calls for rethinking what erasure means. Instead of requiring an impossible physical deletion, the law could move toward a functional view of forgetting. The test would be whether the data has been made ineffective, so that no decision can be based on it anymore, whether or not it has disappeared. Favoring unusability over destruction would likely fit how models work better than the ideal of total disappearance. It brings us back, from the other side, to the provenance question raised in the article on fraud. There, the aim was to authenticate data at the source. Here, it is to neutralize its influence at the source. In both cases, control shifts from the content to its origin and life cycle.

The symmetry is no accident. It points to a broader change in how data is governed under artificial intelligence. Attention used to focus on content, on what a piece of data said and who could access it. Now it is moving to the data's path, where it comes from, how it was captured, what it fed and how its influence can be traced or neutralized. The algorithmic right to be forgotten, like the authentication of evidence, is part of this move toward life-cycle governance, where value and risk sit less in the information than in its traceability. For the firm and its insurer, compliance can no longer mean the ability to delete after the fact. It has to be a control architecture built in from the start, from the data collected to the models it feeds.

With models, the right to be forgotten is less about erasing the past than about not capturing it. Its future will be decided upstream, before absorption¹².