Data Minimisation and Large AI Training Datasets: the Tension

“Data minimisation is incompatible with machine learning” is one of the most repeated claims about the GDPR and it is wrong on the text. The principle constrains relevance to a purpose, not volume. That correction is right, and it is used to carry far more weight than it can bear.

The argument, in one paragraph

The argument runs: Article 5(1)(c) requires personal data to be adequate, relevant and limited to what is necessary in relation to the purposes for which it is processed. The measure is fit to purpose. Where the purpose is to train a model that generalises, a large and diverse corpus is precisely what is necessary, and a smaller one would be inadequate — note that “adequate” is a floor in the same sentence as the ceiling. So a hundred million documents can satisfy minimisation where a thousand would not. Volume is not the test; necessity is.

Everything in that paragraph is correct. The essay’s position is that it is also the beginning of the analysis rather than the end of it, and that it is routinely deployed as though it were the end.

What Article 5(1)(c) actually says

The full text in Regulation (EU) 2016/679 is that personal data shall be “adequate, relevant and limited to what is necessary in relation to the purposes for which they are processed”. Three words, each doing separate work. Adequate — enough for the purpose. Relevant — connected to the purpose. Limited to what is necessary — no more than the purpose requires.

The 1995 Directive said “adequate, relevant and not excessive”. The change to “limited to what is necessary” tightened rather than loosened the requirement, which is worth knowing because the purpose-fit reading is sometimes presented as a modern relaxation. It is not. It is a correct reading of a provision that has always been about the relationship between the data and the purpose.

Note also what minimisation is not. It is not the necessity limb of Article 6(1)(f), though the analyses overlap; it is not storage limitation under 5(1)(e); and it is not a rule about the number of data subjects as opposed to the data held about each. Those distinctions matter for the argument below.

Where regulators have made this argument

The purpose-fit reading is not a defence invention. France’s CNIL published a set of recommendations on the development of AI systems from 2024 onwards, addressing data minimisation directly and taking the position that the principle does not prevent the use of large training datasets, provided the data is selected and, where possible, cleaned and filtered by reference to the objective — published by the Commission nationale de l’informatique et des libertés. The European Data Protection Board’s Opinion 28/2024 of 17 December 2024 proceeds on a comparable footing: it treats measures such as filtering, exclusion of irrelevant sources and removal of unnecessary categories as relevant mitigations, rather than treating dataset size as itself the violation.

What none of that material says is that a corpus is minimised because it is useful. The consistent framing is conditional — large is permissible where selection, filtering and exclusion have been applied by reference to the purpose. The condition is the substance of the guidance and it is the part that gets dropped in the retelling.

Where the argument holds

It holds squarely against the naive objection, which deserves to lose. A rule that a smaller dataset is always more compliant would produce worse models with worse failure modes, and worse failure modes fall on data subjects: a model trained on a narrow corpus is measurably more likely to memorise individual examples and more likely to perform unevenly across populations. Insisting on volume reduction as such can increase the harm the principle exists to prevent, which is a genuine argument and not merely a convenient one.

It holds too against the instinct to treat the number of data subjects as the metric. Minimisation is about the data processed in relation to a purpose. Ten million people’s public writing, used to learn language structure, is a different proposition from ten thousand people’s medical histories used for the same objective, and a volume-based rule cannot tell them apart while a purpose-based one can. The same reasoning is why the compatibility analysis on reusing customer data for training turns on what the data is for rather than on how much of it there is.

Where it stops working

It is only as good as the purpose. Purpose-fit reasoning requires a purpose specific enough to test against, and Article 5(1)(b) independently requires purposes to be specified and explicit. “Training a general-purpose model” is not a purpose that can fail a minimisation test, because nothing is irrelevant to it. An argument that is unfalsifiable is not an argument that has been won, and a controller relying on purpose-fit against an unbounded purpose has smuggled the conclusion into the premise.

It says nothing about categories. Even granting that scale is fine, minimisation still operates within the corpus. Direct identifiers, contact details, precise location and Article 9 material are rarely necessary for learning language or reasoning structure, and the purpose-fit argument, properly applied, argues for stripping them. It is most often cited by controllers who have not stripped anything.

It does not survive the alternatives question. If the same capability can be reached with synthetic data, licensed data, or data with identifiers removed, the personal data is not necessary — and that reasoning belongs to Article 6(1)(f) necessity as much as to 5(1)(c), which is why the two are worked through together on writing a legitimate interest assessment. Purpose-fit answers “is this too much data for this purpose”. It does not answer “is any of it needed”.

It is a principle, not a basis. Satisfying 5(1)(c) does not make processing lawful. A corpus can be perfectly minimised and unlawfully collected. The argument is sometimes used as though clearing minimisation cleared the field, and it clears one principle out of six.

What follows for a controller

The practical consequence is that the purpose-fit reading is a strong position if you have done the work it presupposes, and an obviously weak one if you have not. What the work looks like is unglamorous: a purpose specific enough that some data can be shown to be outside it; a documented selection policy saying which sources were included and which excluded and why; identifier stripping with the fields listed; filtering for categories that carry disproportionate risk, with an honest note of what the filter misses; deduplication, which serves minimisation and reduces memorisation at the same time; and a retention period for the corpus, because 5(1)(c) and 5(1)(e) are different principles and passing one does not address the other.

None of that is burdensome relative to the cost of assembling a training corpus in the first place. The reason it is so often absent is that the purpose-fit argument is available as a substitute for it, which is the real problem with a correct argument used badly: it gives a controller something true to say instead of something to do. The claim that minimisation is about fit rather than volume should be the first line of an assessment, not the whole of one.