GDPR's portability right stops at the embeddings your vendor built from your prompts
Article 20 reaches what you typed and what the service logged. The vectors, the profile and the generated report sit in the category the guidance excludes, which makes leaving an AI vendor a contract job.
The Engineer · Build desk

What happened
- GDPR Article 20 gives a person the right to receive personal data they provided to a controller in a structured, commonly used, machine-readable format and to transmit it to another controller without hindrance.
- The right applies where processing rests on consent or a contract and is automated, which for a consumer AI account generally holds, including direct transfer to a competitor where technically feasible.
- The Article 29 Working Party guidance that still frames Article 20 sorts data into actively given, observed and inferred, and excludes the inferred or derived category from portability entirely.
- According to the dev.to writeup, the law covers the data protection question well, gestures vaguely at ownership of what the AI produced, and is almost silent on the export plumbing itself.
- Even where the right plainly applies, whether a working export button exists is a product decision that the law rarely compels a provider to make.
Why it matters
- What it limits constraint A statutory request hands back material you already keep copies of, so it does nothing about the switching cost that actually locks you in, which sits in the vectors and the scores.
- Who has to decide decision Anyone who wants a usable exit has to buy it as a contract term at signature, because no regulator will supply the export format later.
- What it sets up precedent Since the exclusion exists to protect analytical work, a vendor refusing to return embeddings is taking a defensible legal position rather than committing a compliance failure worth escalating.
Work out what actually lands in the archive. Prompt text and account fields, because you typed those. Usage logs as well, since the guidance treats observed data as provided by virtue of using the service [5]. That is two of the three categories in the taxonomy, and the one you lose is the one that took the compute [14]. On the dev.to author's reading, the excluded bucket is where an AI service keeps the vectors that index your documents, the behavioural profile that decides what you see, and arguably the reports and pictures the model made for you, all of them products of the provider's analysis rather than things you handed over [15]. Inputs are on firm ground, the artefacts built out of them much less so [16].
A right that is discharged by returning a zip of your own typing is a right a vendor can satisfy without touching an interesting table.
For Article 20 to reach further, someone would have to accept that model output is provided data. That is not an oversight waiting to be patched. The line was drawn on purpose, to stop portability being used to extract a company's analytical work [7], and it was drawn in 2016 with recommendation engines and social feeds in view [8]. The reasoning holds for the systems it was written about. It transfers to a generative service only if you also accept that a finished report is the provider's analysis rather than your deliverable [8]. Note too that the direct-transfer-to-a-competitor limb is qualified by technical feasibility, and the vendor is the party in a position to assess that [3].
So exit reads as a procurement exercise. The clauses that decide whether you can leave are the dull ones: a named export format for generated artefacts, a copy or a deletion attestation for the embeddings derived from your corpus, and an explicit position on who holds the copyright in the outputs, since the statute only gestures at ownership [10]. None of that arrives from a data subject request. All of it is cheap to ask for before signature and expensive to ask for mid-migration.
The excerpt says Article 20 is not the only instrument and that newer ones widen the picture, though it breaks off before naming them [12]. Until one of those speaks to derived data, the planning assumption I would defend at review is that switching cost belongs in the procurement model rather than the compliance register, which returns the one thing you already hold a copy of [13].
What to watch
- Whether a regulator or court ever classifies generative model output as data the user provided rather than data the provider derived.
- The newer instruments the dev.to piece says widen the picture beyond Article 20, which the supplied text cuts off before naming.
- Whether any large AI vendor ships an export that includes embeddings, or a deletion attestation for them, without being asked in a contract.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence55
- Adoption
- Not scored
- Hype gap+8
- Incentives30
- Confidence48
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
GDPR Article 20, the right to data portability, gives a person the right to receive personal data concerning them that they provided to a controller in a structured, commonly used and machine-readable format, and to transmit that data to another controller without hindrance.
- [2]
Article 20 applies only where two conditions hold: the processing is based on the person's consent or on a contract, and it is carried out by automated means.
- [3]
For a consumer AI service a person signed up to and uses, those conditions generally apply, so the person can ask for their data in a portable form and, where technically feasible, have it sent straight to a competitor.
- [4]
The Article 29 Working Party guidelines that still frame how Article 20 is read split personal data into three kinds: data actively given, data observed during use of the service, and data inferred or derived by the provider.
- [5]
Data actively given, such as an email address or the text of a prompt, is provided and portable; observed data such as usage logs, search history or location is also treated as provided by virtue of the use of the service and is portable too.
- [6]
The third category, inferred or derived data meaning conclusions the provider generates by analysing the person, is explicitly excluded from the portability right.
- [7]
The exclusion of derived data is described as a deliberate line rather than a forgotten loophole, drawn to stop portability becoming a backdoor to extract a company's analytical work.
- [8]
That line was drawn in 2016 for recommendation engines and social feeds, and lands awkwardly on generative AI, where the derived data is a finished report or picture the user reasonably thinks of as their own creation.
- [9]
The question of getting data out of an AI tool is really three questions: data protection rights to a copy and to move it, ownership of the rights (usually copyright) in what the AI produced, and whether any plumbing such as an export feature or standard format exists.
- [10]
The law has a lot to say about the data protection question, gestures vaguely at ownership, and is almost silent on the export plumbing, which is where most users' frustration lives.
- [11]
Even where the portability right clearly applies, whether an actual working export exists is a product decision the law rarely compels.
- [12]
The piece states that Article 20 is not the only tool and that newer ones widen the picture, but the supplied text breaks off before naming them.
- [13]
In much of the world a person has a legal right to a copy of their personal data, and in the EU a right to have it ported to another service.
- [14]
Two of the three data categories in the guidance are portable and one is excluded.
- [15]
The dev.to writeup argues that the excluded category is exactly where an AI service keeps the interesting material: the embeddings representing a user's documents, the behavioural profile that decides what they see, and arguably the model's own generated outputs, all products of the provider's analysis rather than data the user handed over.
- [16]
The right was drafted around data a person hands over and data a service observes, not data an AI infers or generates, so prompts are on firm ground while outputs, embeddings and the behavioural profile are on much shakier ground.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
- dev.toCan You Get Your Data Out of an AI Tool? The Right Exists on Paper, the Button Usually Doesn't
1 article · August 30, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- Data Ownership and CopyrightFollow
- GDPR ComplianceFollow
- AI Vendor Lock-In and Switching CostsFollow
- Data and identity portabilityFollow
- AI RegulationFollow