Skip to main content

Consent archaeology: what this data was collected for versus what you want to do

The short answer

The governing question is permission, not possession

Having the data is not being allowed to use it. Consent archaeology asks what the data was collected for and whether your new use is one people actually agreed to, and that question decides the matter regardless of how good or how convenient the data is.

What you will be able to do

  • Distinguish the purpose data was collected for from the purpose you now want to use it for, and state why the gap between them is the governing question, not the size or quality of the data.
  • Trace a dataset back to its original collection point and locate the actual promise made to people: the consent text, the privacy notice, the terms, or the legal basis recorded at the time.
  • Analyze whether a new AI use is compatible with the original purpose, using the concrete factors real privacy law applies (the link between old and new purpose, the context and relationship, the sensitivity of the data, the consequences for people, and the safeguards in place).
  • Classify each dataset feeding your intended AI system into one of three verdicts: the new use is covered, the new use needs a fresh legal basis or fresh consent, or the new use should not happen with this data at all.
  • Identify the special cases that change the answer: sensitive categories of data, data about children, data collected under one company that moved to another, and public or scraped data that still carries obligations.
  • Produce a written consent-and-purpose record for each dataset that names the original purpose, the promise made, your intended new use, your compatibility judgment, and the decision, so it can drop straight into the data provenance file you build later in this module.
  • Determine which legal basis actually governed a dataset (consent, contract, legal obligation, or a documented legitimate interest) and judge what a new use would require under each, rather than assuming a flexible basis that was never the real one.
  • Defend that record against challenge: explain to a skeptic exactly why a given reuse survives scrutiny or why you cut it.

The ideas, one by one

Dig to bedrock

Recover the actual promise made at collection (the real notice, consent text, or legal basis in force that day) and then the reasonable expectation underneath it. Stopping at "we have a broad clause" is the exact mistake that ended the Lee Luda chatbot; the clause did not reach what a normal person would have foreseen.

Judge compatibility with the law's own factors

Run the link, the context, the nature of the data, the consequences, and the safeguards explicitly, write the reasoning for each, and do not average them. A single decisive factor, such as special-category data with no valid basis or a real re-identification risk, can sink a reuse on its own.

Land on one of three verdicts, in writing

Covered, needs-a-fresh-basis, or do-not-use. Forcing a written verdict per dataset (and per slice when promises differ) is what turns a vague worry into a governable decision your organization has actually made.

Check the four special cases every time

Sensitive categories, children's data, data that changed hands, and public or scraped data each carry heightened obligation and can flip a comfortable "covered" into a "do not." Free-text and conversational data almost always hide special-category content.

A "do not use" is governance succeeding

The value you add is the discipline to cut or gate data your organization is excited about, with reasoning that survives a challenge. The record you produce feeds directly into your data provenance file, your conformity file, and your evidence annex, so the dig you do today is the proof you rely on when the audit comes.

Purpose limitation protects trust, so ask whether reuse breaks the deal

The principle exists because people share data on a bounded understanding of what it is for, and a surprising reuse breaks that understanding. When you judge a reuse, the sharpest question is not "could we technically use this" but "would using it this way break the deal the person thought they were making." That framing gets you to the right verdict faster than any clause-parsing.

Find the legal basis, do not assume consent

Data can rest on consent, contract, legal obligation, or a documented legitimate interest, and each stretches differently to a new use. Part of the dig is establishing which basis actually governed the dataset, because a reuse a contract cannot carry might need fresh consent, and a legitimate interest is only as good as the balancing assessment behind it.

The promise must come before the use

You cannot use data for a new purpose and write the permission afterward. Updating terms after launch, as alleged in the LinkedIn 2024 case, does not retroactively authorize processing that already happened, and "we will fix the notice later" is a sequence error, not a plan.

Listen to it

The same lesson, talked through by two hosts as episode 12 of the podcast.

Real cases

These examples show consent archaeology applied to real, documented situations, with the reasoning stated. The Scatter Lab case is the anchor and is treated in depth; the others are supporting illustrations. Details are as verified against primary and reputable sources; where a matter is still developing it is labeled.

Example 1 (anchor): Scatter Lab and the Lee Luda chatbot, South Korea (2021). Scatter Lab ran two apps, one giving feedback on your texting and one analyzing relationships, that collected users' KakaoTalk messages. The company then reused those messages to build Lee Luda, a chatbot launched in December 2020 and presented as a young woman you could chat with. It reportedly drew on around 9.4 billion messages from about 600,000 people. Within weeks Luda was generating hate speech and sexual content and, worse, occasionally surfacing real names, nicknames, and addresses from its training data. On 28 April 2021 the Personal Information Protection Commission (PIPC) imposed a penalty of about 103.3 million Korean won (roughly USD 92,900) and corrective orders, finding eight violations of the Personal Information Protection Act. The core finding is a consent-archaeology finding: the "New Service Development" clause users had agreed to was not enough for them to anticipate that their private messages would be used to develop and operate a chatbot, so the reuse exceeded the original collection purpose. The PIPC also found unlawful collection of more than 200,000 children's records without parental consent and unlawful exposure of 1,431 real messages on a public code repository. The company later pseudonymised its data pipeline, deleted the disputed database and model, and relaunched a reworked version of the service. The dollar figure was trivial for a lesson this large: the data was excellent, the permission was missing, and nobody dug for it before shipping. (Sources: Future of Privacy Forum, "South Korea: The First Case Where the Personal Information Protection Act Was Applied to an AI System"; The Register, 29 April 2021.)

Example 2: Meta's paused EU AI training (2024). In 2024 Meta announced plans to train its AI systems on the public posts of its European users, relying on a "legitimate interest" basis rather than fresh consent. After pressure from the Irish Data Protection Commission (DPC) and eleven complaints filed by the privacy group noyb arguing the legal basis was invalid, Meta paused the plan. This is consent archaeology playing out at frontier scale: the question was not whether Meta held the posts (it did) but whether the original context in which people posted covered reuse as AI training material, and whether "legitimate interest" could carry that reuse. The pause shows Verdict 2 in the wild, a reuse stopped pending a firmer basis. (Source: TechCrunch, "Meta pauses plans to train AI using European users' data," 14 June 2024.) Treat the ultimate lawfulness of such training as still contested; the European Data Protection Board's Opinion 28/2024 (below) is the current reference point.

Example 3: The EDPB's Opinion 28/2024 on AI models (December 2024). On 17 December 2024 the European Data Protection Board (EDPB), the body that harmonizes GDPR interpretation across the EU, adopted Opinion 28/2024 at the Irish regulator's request. It addresses three questions squarely relevant to this topic: when an AI model can be considered anonymous, whether "legitimate interest" can be a valid basis for developing and deploying AI models, and, most strikingly, what happens if a model was built on personal data that was processed unlawfully. The opinion's answer to the third question is the sharpest tool this topic can hand you: if the underlying data was collected or processed unlawfully, that unlawfulness can taint the resulting model, which is exactly the "quiet poison" logic Module 2 is built on. Legitimate interest can work, the EDPB says, but only where the processing is strictly necessary and the balance of rights is respected, judged through a documented three-step assessment. This is established EU guidance as of the date given. (Source: EDPB, Opinion 28/2024, 17 December 2024.)

Example 4: LinkedIn's default opt-in for AI training (2024 to 2025). In September 2024 LinkedIn updated its settings so that member data could be used to train generative AI models, with the setting switched on by default rather than requiring members to opt in. Reporting (404 Media, TechCrunch, 18 September 2024) indicated the company had begun using member data for training before its privacy terms were updated to describe it, and LinkedIn said it would update the terms "shortly." In January 2025 a proposed class action was filed in California alleging that private messages of Premium subscribers were used to train models. This is consent archaeology at the scale of a major platform, and it illustrates two distinct failures cleanly. The first is the sequence failure: using data for a new purpose and updating the promise afterward reverses the correct order, in which the promise must exist before the use. The second is the default-on failure: profile and message content was collected in the context of professional networking and private messaging, and switching on AI training by default asserts a compatibility that the original context does not obviously support. Whatever the litigation ultimately decides, the case is a live demonstration of the collection-purpose-versus-intended-use gap and of why "we will fix the terms later" is not consent archaeology. (Sources: TechCrunch, "LinkedIn scraped user data for training before updating its terms of service," 18 September 2024; The Washington Post, 23 September 2024.) Treat the specific legal outcome as still developing.

Example 5: DeepMind and the Royal Free NHS Streams project (pointer). The most famous purpose-mismatch case in health data, in which a UK hospital trust shared 1.6 million patient records for a purpose the patients had not been told about, is the deep case study for Topic 2.6, where you assemble the provenance file. (see Topic 2.6) It is named here only to show that the same collection-purpose-versus-intended-use question governs the most sensitive data there is; the full treatment belongs to that topic.

Example 6: Public image datasets and the source-taint problem (pointer). The discovery of unlawful content buried inside a very large public image training set, and what that does to everything trained on it, is the anchor for Topic 2.1. (see Topic 2.1) It is the sibling failure to this one: 2.1 is data that was poisoned at the source, 2.2 is data that was clean but never permitted for this use. Together they define the two halves of "your data is not automatically usable."

What the pattern across these cases shows. Read together, the cases share one shape. In every one, the organization held data it was entitled to hold and then used it for a purpose the people had not agreed to, and in every one the defense would have been the same document: a written record showing the original purpose, the intended use, and why the second sits inside the first. Scatter Lab had a broad clause and no such record. Meta and LinkedIn asserted a basis the original context did not obviously support. The EDPB opinion tells you that when this analysis is skipped, the defect can follow the data all the way into the model. The lesson is not that reuse is forbidden; the telemetry case in Section 3H reused data cleanly. The lesson is that reuse is a decision that must be made and recorded before the build, not assumed because the data is convenient.

Where people go wrong

  • "We own the data, so we can use it however we want." Ownership of a database is not permission to use the personal data inside it for any purpose. Purpose limitation binds the holder regardless of who "owns" the storage. The question is never possession; it is permitted purpose. This single confusion is the root of most consent-archaeology failures.
  • "There's a broad clause that covers it." A broad clause such as "to improve our services" or "for new service development" is exactly what failed for Scatter Lab. Regulators test whether the clause was specific enough for a reasonable person to anticipate the actual use. Breadth is a weakness, not a strength: the vaguer the clause, the less it can be said to have informed anyone about this particular reuse.
  • "If we anonymize it, none of this applies." Sometimes true, usually harder than it sounds. Genuine anonymisation (data that can no longer be linked to a person, even by combination with other data) does take data out of scope. But free text, chat logs, and audio are extraordinarily hard to truly anonymize because they are saturated with self-identifying detail and, for audio, voiceprints. A light scrub that removes obvious names is pseudonymisation, not anonymisation, and pseudonymised data is still personal data. Do not let "we'll anonymize it" end the analysis; verify it against a real test (draft EDPB guidance from July 2026 frames this as no isolation, no linkage, no inference, all three, not just names removed), and remember the bar is high.
  • "It was public, so consent isn't needed." Public visibility is not consent and does not switch off purpose limitation, special-category protection, or children's protection. Scraped and public data still describes people and still carries obligations you must establish, not assume.
  • "The data changed hands, so we get to write the rules now." When data arrives through acquisition or a vendor, it carries the promise made at its original collection, not a fresh promise you author today. You inherit the consent debts along with the asset. Dig for the original terms.
  • "Sensitive content is rare in our data." In free-text and conversational data it is not rare; it is everywhere. Ordinary messages routinely reveal health, relationships, beliefs, and legal situations. Assuming your chat logs are "ordinary" data is one of the fastest routes to an Article 9 violation you never saw coming.
  • "We have a legitimate interest, so we're fine." Legitimate interest is a real basis, but naming it is the start of the analysis, not the end. It requires a documented balancing test showing the use is genuinely necessary and does not override people's rights, and it does not carry special-category data on its own. Treating the phrase as a conclusion is a shortcut regulators specifically look for.
  • "Transcripts are fine even if the audio is not." The words and the voice are two different things. Even where transcript text might be covered, the underlying voice recording carries a voiceprint, which is biometric and can identify the person, so training a model on voice needs its own basis. "You agreed we could read your text" is not "you agreed we could train on your voice."
  • "A quick legal sign-off is the same as consent archaeology." A one-line "legal says it's fine" is not a record and does not survive a challenge. The deliverable is a written finding that names the original purpose, the intended use, the factor-by-factor reasoning, and the verdict, so that a hostile reviewer can check each step. Prose reassurance is not proof.
  • "If the analysis says do-not-use, governance failed." The opposite. Reaching Verdict 3 on data your organization wanted to use is governance working. A function that always returns "covered" is not protecting anyone; it is rubber-stamping. The willingness to cut a dataset, with reasons, is a sign of a mature process, not a broken one.
  • "Consent archaeology is a one-time gate at the start." Purposes drift. A dataset cleared for one AI use gets quietly repurposed for a second, and the original clearance does not stretch to cover it. Each materially new use needs its own dig. The record you write is dated and scoped for a reason.
  • "Legitimate interest lets us use anything for AI." Legitimate interest is a real legal basis, but it is not a universal key. It requires a documented balancing test showing the use is genuinely necessary and does not override people's rights, and it does not carry special-category data on its own. Treating "we have a legitimate interest" as the end of the analysis rather than the start of a documented assessment is exactly the shortcut regulators scrutinize.
  • "We can fix the terms after we launch." Updating a privacy notice or terms of service after you have already used the data for the new purpose reverses the correct order. The promise has to exist before the use it is meant to cover. Using data first and describing it later, as the LinkedIn 2024 reporting alleged, does not retroactively create the permission for the processing that already happened.
  • "The data is old and the people are long gone, so nobody will mind." Age of data does not dissolve the original promise or cure an unlawful collection. Old data collected under a narrow purpose is still governed by that narrow purpose, and old data collected unlawfully (children's data with no parental consent, for instance) stays unlawful. Time launders nothing.

Questions people ask

What is consent archaeology?
The disciplined practice of tracing a dataset back to the moment it was collected, recovering the actual promise or legal basis in force at that time, and judging whether an intended new use, especially an AI use, is one the people it describes actually agreed to. Named for the excavation involved: the permission is invisible in the data and must be dug up from the record around it.
What is collection purpose?
The reason personal data was gathered, as stated to people at the time of collection through a privacy notice, consent text, terms of service, or a recorded legal basis. The collection purpose is the reference point against which any later use is judged.
What is intended use?
What an organization now wants to do with data it already holds. Consent archaeology is the comparison of the intended use against the collection purpose. More on Intended use
What is purpose limitation?
The privacy principle that personal data collected for one specified purpose may not be freely reused for an unrelated, incompatible purpose. Written into the GDPR at Article 5(1)(b), the CCPA/CPRA at California Civil Code section 1798.100(c), South Korea's PIPA, and most modern privacy regimes. More on Purpose limitation
What is legal basis?
The lawful ground on which personal data is processed. Consent is one basis, but data can also rest on a contract, a legal obligation, or a documented legitimate interest, among others. Part of the dig is establishing which basis actually governed a dataset, because that determines what a reuse requires.

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 →