Start Free Trial
← All posts
Practical orientation, not legal advice

GDPR and live captioning: what to check

The moment captions appear on screen, personal data is being processed. Six questions to put to any vendor, and why the answer "nothing is stored" changes the whole analysis.

6 vendor questions 8 min read Published August 2026

An EU organisation adding live captioning is doing something unambiguous under the GDPR: processing personal data. A voice belongs to an identifiable person. A transcript records what identifiable people said — names, opinions, sometimes health details that surface mid-meeting without warning. None of this makes captioning risky by nature; captioning is often deployed precisely to serve accessibility and inclusion. It means the vendor behind the captions is processing personal data on your behalf, and you are accountable for knowing exactly what they do with it. Six questions cover nearly all of it. This is practical orientation, not legal advice — bring your DPO in early.

Why captions are personal data processing

Two data flows exist in every live captioning session. The audio — people's voices, captured and sent to a recognition engine. And the transcript — a written record of what identifiable individuals said, generated in real time. Both relate to identifiable people, so both are personal data; the GDPR applies regardless of whether anything is saved afterwards, because processing includes the transient handling itself.

The content compounds it. Meetings wander: a colleague mentions a medical appointment, a customer names another person, an HR discussion touches the exact categories the regulation treats as special. A captioning pipeline cannot know in advance that a sensitive sentence is coming — so the sensible posture is to choose a pipeline whose handling you would accept for the most sensitive sentence likely to cross it.

The reframe that makes this tractable: you are not assessing "a captioning feature." You are assessing a data processor that receives live audio of your staff, customers, or patients. Every question that follows falls out of that framing naturally.

The six questions to put to any captioning vendor

What lawful basis will this rest on — and does the tool support it?

Your organisation chooses the basis (legitimate interests for internal meetings, consent in some contexts, and so on) — but the tool must not undermine it. A tool that quietly retains or reuses data can invalidate an analysis built on minimal processing. Ask the vendor what processing actually occurs, in writing.

What is retained, and for how long?

The data-minimisation question, and the single most important one. Is audio stored at all — even transiently on disk? Do transcripts persist after the session? Are there backups, logs, or caches containing content? "We delete on request" and "we never keep it" are very different answers.

Where is the processing performed?

Which countries do audio and transcripts transit and rest in? If processing leaves the EU/EEA, what transfer mechanism covers it? Your DPO will need the storage and processing locations named, not gestured at.

Is the vendor a processor — and will they sign a DPA?

For organisational use the vendor should act as your processor, under a data processing agreement covering instructions, sub-processors, security measures, and breach notification. A vendor that treats meeting content as its own data to use is a different — and worse — arrangement.

Is session content used to train models?

Training on customer content is further processing for the vendor's own purposes, and it stretches or breaks the processor relationship. The clean answer is a flat no, stated in the terms — not an opt-out buried in settings.

How does deletion work — and can you prove it?

If content is retained at all, you need deletion mechanics that support data-subject rights: scope, timelines, backups included or not. An architecture that deletes everything at session end answers this question before it is asked.

Why "nothing stored" changes the whole analysis

Most GDPR effort around a new tool goes into governing data at rest: retention schedules, access controls, subject-access handling, deletion verification, breach exposure. An architecture that stores nothing removes the "at rest" category almost entirely — audio streams through recognition and is gone; the transcript exists for the session and is deleted when it ends.

  • Data minimisation is built in, not policied on: the system physically keeps the minimum, which is the strongest possible answer to the retention question.
  • Data-subject requests get short answers. Nothing is held about the meeting, so there is nothing to search, export, or forget.
  • Breach exposure shrinks structurally — an archive that does not exist cannot leak.
  • Your records of processing get simpler: transient processing for a defined purpose, no retention, one processor.

The trade-off is real and worth stating: nothing stored means no meeting archive, no searchable history, no minutes generated from last month's calls. For live comprehension — accessibility, multilingual meetings, interpreted encounters — that trade is usually exactly right; if you separately need an archive, treat that as a second, deliberately-governed processing activity rather than a captioning default.

The same shape as the HIPAA question

If this checklist feels familiar, it should. US healthcare organisations run the same analysis under HIPAA — who is the vendor to us legally, what do they retain, where does audio go, is content reused — with different statutes and the business-associate agreement in place of the DPA. We wrote the healthcare version in our HIPAA-compliant live captioning guide, and the two make a useful pair: a vendor with good answers to both has privacy in its architecture rather than its marketing.

The parallel holds down to the punchline: in both regimes, the architecture that stores nothing is dramatically easier to defend than the one that stores everything and promises to be careful. For EU health settings specifically — where both instincts apply — see telehealth live captioning and translation.

Bring the DPO in early — and what to hand them

None of the above replaces your organisation's own assessment. What it does is make the DPO conversation fast: arrive with the vendor's written answers to the six questions, the categories of meetings you intend to caption, and your proposed lawful basis, and the review becomes an afternoon instead of a quarter. If captioning will touch high-risk contexts — patient conversations, employee disciplinary meetings — expect your DPO to consider whether a data protection impact assessment is warranted, and let them make that call.

The encouraging conclusion: live captioning is one of the easier tools to bring through GDPR review well — the processing purpose is clear and sympathetic, the data flow is simple, and nothing-stored architectures exist. Choose one, document the answers, and captions stop being a compliance question at all.

Frequently asked

Is AI transcription GDPR compliant?

It can be — GDPR compliance depends on the specific tool's architecture and the way you deploy it, not on the category. The checks that matter: what the vendor retains (audio and transcripts), where processing happens, whether the vendor acts as your processor under a DPA, whether content is used for model training, and how deletion works. Unicaption is GDPR compliant with a nothing-stored design: audio is never stored, transcripts are deleted at session end, and content is never used for training.

Are live captions and transcripts personal data under GDPR?

Yes. A voice belongs to an identifiable person, and a transcript records what identifiable people said — names, opinions, and sometimes special-category details that surface mid-conversation. The GDPR applies even when nothing is saved afterwards, because transient handling is still processing. That is why the vendor behind a captioning tool should be assessed as a data processor receiving live audio, not as a cosmetic meeting feature.

What should I ask a captioning vendor before using it in the EU?

Six things, in writing: what processing occurs and on what lawful basis it can rest; what is retained and for how long; where audio and transcripts are processed and stored; whether the vendor signs a data processing agreement as your processor; whether session content trains models; and how deletion works. A vendor that stores nothing — Unicaption deletes transcripts at session end and never stores audio — gives the shortest defensible answers to all six.

Do we need consent from everyone in a meeting to use live captions?

Not necessarily — consent is one lawful basis among several, and many organisations rest internal captioning on legitimate interests, particularly for accessibility. Which basis fits depends on the context, the participants, and what the tool actually does with the data, which is why a nothing-stored tool makes the analysis easier. This is practical orientation, not legal advice: your DPO should choose the basis and the participant-notice approach.

Captions your DPO can approve

GDPR compliant, end-to-end encrypted, audio never stored, transcripts deleted at session end. 30 free minutes every week — no credit card.

Start Free Trial →