Skip to main content

The retention decision: what you must keep, what you must destroy, and proving both

The short answer

Retention is a decision, not a default

The natural state of systems is to keep everything forever, and that is the failure. Governance replaces the default with a per-dataset decision that fixes a purpose, a lawful basis, a clock, and a trigger.

What you will be able to do

  • Frame retention as an active decision tied to purpose and lawful basis, not a storage default, and explain why "keep everything" and "delete everything" are both governance failures.
  • Apply the storage-limitation principle (keep personal data no longer than necessary for the purpose) to a real dataset, naming the purpose, the lawful basis, and the trigger that ends the retention period.
  • Separate the two obligations that pull in opposite directions. Name the duty to destroy data past its purpose or without a lawful basis, and the duty to preserve data you are legally required to keep (legal holds, records rules, audit and reproducibility needs).
  • Locate every place a dataset actually lives in an AI system (databases, backups, logs, caches, embeddings, model weights, vendor systems, analytics copies) so that "delete" means all of them, not one table.
  • Design a proof-of-destruction record that would survive challenge: what was destroyed, when, by whom, by what method, and who witnessed it.
  • Resolve the common collisions: a legal hold that blocks an erasure request, a minimization duty that fights a reproducibility need, and a "we deleted the database but the model still remembers" problem.
  • Choose among the three retention exits (delete, anonymise, restrict) for a given dataset, and explain why naive anonymisation is not a lawful exit for rich data.
  • Treat storage location and jurisdiction as part of the retention decision, and handle a conflict where one legal system orders you to keep data another orders you to destroy.
  • Produce a retention and destruction schedule for one real dataset in your own organization that a data protection officer (DPO), an auditor, and a court could all follow.

The ideas, one by one

Two duties pull opposite ways, and both are real

You must destroy data past its purpose or without a lawful basis, and you must keep data under legal holds, records mandates, and AI audit obligations. "Keep everything" and "delete everything" are both governance failures; the skill is knowing which duty governs which data and documenting why.

In an AI system, "delete" means eight places, not one

The same data lives in backups, logs, caches, warehouses, embeddings, model weights, and vendor systems. A deletion that touches only the primary store is theater. Map where data actually lives before claiming it is gone.

The model may remember after the dataset is gone

Models memorize training data, and reliable machine unlearning is not yet deployable (emerging, 2026). If you must destroy training data, budget for retraining, and know this before you train, not after.

Proving destruction is the half everyone skips

Anyone can produce data they kept; almost nobody can prove data they destroyed. A proof-of-destruction record (what, when, who, how, verified) must be designed into the destruction process, because it cannot be honestly reconstructed later.

Over-retention is accumulated liability, not caution

The most secure record is the one you no longer hold. Every dataset kept past its purpose is a larger breach surface and a standing storage-limitation violation. Deleting on schedule is a security control.

Collisions are resolved by partitioning and documenting, never by silence

When an erasure request meets a legal hold, erase what you can, retain the narrow set the law requires with its basis, restrict it, and tell the person. Doing nothing is the only indefensible option.

The retention schedule is a load-bearing artifact

One row per field-group, tied to purpose and basis, reaching every copy, producing proof at every destruction. It becomes the retention layer of your provenance file and the retention evidence in your conformity file, and it is what answers "prove that data is gone" when your file is attacked.

There are three exits, not one

A dataset leaves its retention period by delete, anonymise, or restrict. Naive anonymisation is not a real exit, because weakly de-identified rich data is re-identifiable, as AOL and Netflix both proved. Anonymise only when it is genuine and provable; pseudonymised data is still personal data with a clock.

Location is a retention decision

Where data lives and under whose law determines whether you can force deletion and prove it, and whether a foreign keep order will collide with your destroy duty. Decide storage jurisdiction at collection, not at deletion, or inherit the Worldcoin cross-border proof problem.

Retention is a design-time constraint on building AI

The cheapest moment to exclude data you might have to destroy is before it enters the model weights. Check lawful basis and destruction exposure before training, because afterward the only reliable remedy is a costly retrain.

Retention is a joint decision with a single owner

Legal sets the periods, engineering knows where the data lives, and the business owns the purpose. A schedule written by any one of them alone has a predictable hole, so the artifact has one named owner who convenes the others and keeps it current.

Listen to it

The same lesson, talked through by two hosts as episode 13 of the podcast.

Real cases

These examples span jurisdictions and show the retention decision from both sides: what happened when organizations kept what they should have destroyed, destroyed what they should have kept, or could not prove either.

Example 1: Worldcoin in Kenya (the anchor). Worldcoin, a project co-founded by OpenAI chief executive Sam Altman and operated by Tools for Humanity, offered people cryptocurrency tokens to have their irises scanned by an "orb," creating a biometric-backed digital identity called World ID. In Kenya, tens of thousands enrolled through 2023. On 2 August 2023 the government suspended the operation. The Office of the Data Protection Commissioner (ODPC) and the courts found the processing unlawful: no Data Protection Impact Assessment (DPIA, the mandatory risk assessment for high-risk processing) had been carried out before collecting sensitive biometric data, and consent obtained through the inducement of free tokens was not the freely given consent the Data Protection Act, 2019 requires. On 5 May 2025 the High Court in Nairobi ordered Tools for Humanity to permanently delete all the biometric data it had collected, within seven days, under ODPC supervision. Because the data had been transferred out of the country, verifying destruction meant reaching across borders; the ODPC confirmed publicly, in January 2026 after a compliance audit, that all the Kenyan biometric data had been deleted. The case is a complete retention lesson: data collected without a lawful basis had to be destroyed, not merely paused; the destruction had to be proven, not asserted; and the proof required a court order, formal supervision, and an audit because assurances were not trusted. (Sources: Business Daily Africa; the-star.co.ke; CIPIT, Strathmore University; the Data Protection Act, 2019, Kenya.)

Example 2: The right to be forgotten in the European Union. In 2014 the Court of Justice of the European Union ruled in Google Spain SL v Agencia Española de Protección de Datos and Mario Costeja González (Case C-131/12) that individuals can, in defined circumstances, require a search engine to remove links to personal data that is inadequate, irrelevant, or no longer relevant. The ruling established, before the GDPR codified it in Article 17, that "no longer necessary" is a legally operative condition and that continued retention or availability of personal data can itself be the harm. It reframed deletion from an act of courtesy to an enforceable obligation. (Source: Court of Justice of the European Union, judgment in C-131/12.)

Example 3: Failing to keep records in United States finance. In September 2022 the SEC announced charges against fifteen broker-dealers and an affiliated investment adviser for widespread and long-standing failures to maintain and preserve electronic communications, with penalties exceeding USD 1.1 billion; the CFTC brought parallel actions. Employees had conducted business over personal messaging apps that the firms did not capture, so records the firms were legally required to keep and produce simply did not exist. This is the keep side made concrete: destroying or never capturing records you are mandated to retain draws penalties that dwarf most privacy fines. (Sources: US SEC and CFTC press releases, September 2022.)

Example 4: Retention rules as fixed obligations. Under HIPAA in the United States, covered entities must retain required documentation for six years from creation or last effective date (45 CFR 164.316(b)(2)); under SEC Rule 17a-4, broker-dealers must preserve specified records for defined periods in a non-rewriteable, non-erasable format. These are not privacy rules; they are affirmative keep-and-prove mandates. They show that a real retention schedule almost always contains rows where the correct period is measured in years and the correct action at the trigger is "retain and be able to produce," not "delete." (Sources: 45 CFR 164.316; 17 CFR 240.17a-4.)

Example 5: Backups and the meaning of "delete." The UK Information Commissioner's Office has long guided that the right to erasure extends to backups. Where deleting a single record from a backup is not immediately feasible, the controller may put the backup data "beyond use" (isolated, secured, not used for any other purpose, and deleted when the backup cycle overwrites it) and must tell the individual. This established position is the honest answer to the "but it is in our backups" objection: beyond-use plus a documented deletion-on-cycle is compliant; pretending backups do not count is not. (Source: UK ICO guidance on the right to erasure.)

Example 6: Anonymisation that was not. Two cases separated by two years made the same point. In 2006 AOL released about 20 million search queries from more than 650,000 users for research, having replaced usernames with numbers; The New York Times re-identified user 4417749 as a named individual within days from her own searches, and AOL pulled the data, apologised, and saw executives depart. In 2008, researchers re-identified subscribers in the "anonymised" Netflix Prize dataset by matching it against public IMDb ratings, and a planned sequel to the contest was cancelled after regulatory and privacy concerns. Both organisations believed they had taken data out of scope by stripping obvious identifiers. Both were wrong, because rich, high-dimensional data is re-identifiable from a handful of external facts. The retention lesson: anonymisation is a valid exit from a retention obligation only when it is genuine and robust against realistic re-identification, and calling weakly de-identified data "anonymous" turns a retention shortcut into a worse breach. (Sources: The New York Times, 2006; Narayanan and Shmatikov, "Robust De-anonymization of Large Sparse Datasets," 2008.)

Example 7: When keep and destroy orders collide across borders. Organisations operating in both the United States and the European Union routinely face a direct collision: a US court's duty to preserve evidence (a litigation hold, enforced through Federal Rule of Civil Procedure 37(e)) meets an EU data subject's right to erasure or a regulator's deletion order over the same records. One legal system commands retention; the other commands destruction. As of 2026 this tension is managed case by case through legal advice in both jurisdictions, narrowing the data in scope, restricting rather than freely retaining, and documenting the competing obligations, rather than resolved by any single clean rule. It is the clearest example that retention is not a private engineering choice but a negotiation among legal systems, and that where you store data determines how painful that negotiation becomes. (Source: US Federal Rule of Civil Procedure 37(e); GDPR Article 17; cross-border discovery and blocking-statute commentary.)

Example 8: The model that remembers. Extraction and membership-inference research (including work led by Nicholas Carlini and collaborators from 2021 onward) has demonstrated that large models can memorize and, under the right prompts, regurgitate specific training examples, including verbatim personal data. The governance consequence for retention is direct: destroying a training dataset does not guarantee the trained model has forgotten it. As of 2026, reliable machine unlearning is emerging research, not a deployable guarantee, so an organization that must destroy training data should treat retraining as the likely remedy and factor that cost into the decision before training. (Source: Carlini et al., extraction and membership-inference literature, 2021 to 2024.)

Where people go wrong

  • "We deleted it" means the record is gone. Almost never true on the first pass. The data typically survives in backups, logs, caches, warehouses, embeddings, model weights, and vendor systems. A deletion that touches only the primary database is a gesture, not a destruction. Map the eight copy locations before you claim anything is deleted.
  • Keeping everything is the safe choice. Over-retention is not safety; it is accumulated liability. Every dataset kept past its purpose is a larger breach surface, a bigger discovery burden, a standing storage-limitation violation, and a temptation for future misuse. The most secure record is the one you destroyed and can prove you destroyed, because the proof itself is a new record that must be secured and kept minimal, or you have just built a smaller version of the same liability.
  • Deleting everything is the safe choice. The opposite error. Legal holds, records-retention mandates, anti-money-laundering rules, tax obligations, and AI audit needs create hard duties to keep specific data. Destroying data you were required to preserve can be spoliation or a regulatory violation with penalties larger than most privacy fines.
  • A retention policy is a document you write once. A policy that is not tied to the systems that enforce it, not owned by a named person, and not reviewed on a schedule is decoration. Retention is enforced by automation and evidenced by proof, or it is not enforced at all.
  • Consent withdrawal and legal holds cancel each other, so do nothing. They do not cancel; they partition. Erase everything the hold does not cover, retain the narrow set it does, restrict that set to the retention purpose, and tell the individual what you kept and why. Doing nothing is the one indefensible option.
  • Destroying the training dataset clears the model. A model can memorize its training data. Deleting the dataset does not reliably remove that influence, and reliable machine unlearning is not yet a deployable guarantee (emerging, 2026). If you must destroy training data, retraining without it may be the only sound remedy, and you should have known that before training.
  • Proof of destruction can be assembled after the fact if anyone asks. It cannot be honestly reconstructed later; the timestamped record either was created at destruction or does not exist. Design the proof-of-destruction record as part of the destruction process, not as a scramble when the regulator writes.
  • Retention is one clock per record. A single record holds fields with different purposes, bases, and obligations. The verification selfie and the tax-relevant transaction amount in the same customer row have completely different retention answers. Schedule by field-group where the fields diverge, not by record.
  • Anonymising the data means we can keep it forever with no obligation. Only if the anonymisation is genuine and robust against re-identification, which for rich data is hard. Stripping names and inserting numbers is not anonymisation; AOL and Netflix both learned that "anonymised" high-dimensional data was re-identified from a few external facts. Pseudonymised data (identifiers swapped for a key you hold) is still personal data with a retention clock. Claiming a protection you do not have is worse than admitting you kept the original.
  • Cross-border storage is just a security question. Where data physically lives, and under whose law, is a retention question. It determines whether you can compel deletion, how you prove it, and whether a foreign court's keep order will collide with a local destroy order. The Worldcoin deletion needed a court order and months of verification precisely because the data and the proof lived across a border.
  • When two legal systems give opposite orders, we can wait until it resolves. A US preservation duty and an EU erasure duty over the same data will not resolve themselves, and waiting breaches whichever obligation is most enforceable. Surface the conflict, take advice in both jurisdictions, narrow and restrict the data, and document the reasoning. Inaction is the one universally indefensible choice.
  • Backups make deletion impossible, so backups are exempt. They are not exempt. Where immediate surgical deletion from a backup is infeasible, put the backup data beyond use (isolated, secured, unused, deleted on the backup cycle) and document it. "It is in our backups" is not a defense; "it is beyond use and expires on this date" is.
  • The eight copy locations are the whole map. They are the known map. Dark data (forgotten exports, an old spreadsheet, a shared drive nobody audits) and shadow AI (an engineer piping data into a third-party foundation model or tool that was never approved or inventoried) create copies outside any of the eight named locations. A schedule built only from what the team already knows about will miss exactly the copy most likely to surface in a breach.

Questions people ask

What is retention decision?
The governance choice, made per dataset, of how long to keep data, what to destroy, when, and how to prove both, replacing the system default of keeping everything forever.
What is storage limitation?
The principle that personal data must be kept in identifiable form no longer than necessary for the purpose it was collected for. Codified in GDPR Article 5(1)(e), Kenya's Data Protection Act, 2019 Section 39, and equivalents worldwide. More on Storage limitation
What is retention period?
The length of time a specific dataset or field-group may be kept, derived from its purpose and any overriding keep obligation, not chosen for convenience. More on Retention period
What is destruction trigger?
The defined event that starts the clock toward destruction, such as purpose fulfilled, consent withdrawn, account closed, or a fixed period expiring.
What is lawful basis?
The legal ground on which data is processed (for example consent, contract, legal obligation, or legitimate interest). Retention is only lawful while a valid basis persists; when the basis collapses, the data usually must be destroyed. More on Lawful basis

Keep going

Watch the lesson and prove it

Proving it happens inside the program: watch, explain it in your own words, and it goes on a record employers can check.

Watch the lesson and prove it →