Is AI dictation safe for privileged material? An Irish answer

Every conversation about AI dictation at the Irish bar arrives at the same question within a minute: is it safe for privileged material? It's the right question, asked in the right order — before the demo, not after the breach. The honest answer is: it can be, and in some respects it can be safer than what most practices do today — but only if the vendor is built for it, and only if you ask the questions that actually matter. Here they are.

What privilege actually protects

Legal professional privilege in Ireland comes in two forms. Legal advice privilege attaches to confidential communications between lawyer and client made for the purpose of giving or seeking legal advice — the principle the Supreme Court set out in Smurfit Paribas Bank Ltd v AAB Export Finance Ltd. Litigation privilege covers confidential communications made for the dominant purpose of actual or contemplated proceedings, and extends to third parties brought in for that purpose.

Both forms rest on the same foundation: confidentiality. Privilege belongs to the client, and what destroys it is not the involvement of a machine or a service — it's the loss of confidentiality. That's the lens for everything that follows.

Dictation has always involved third parties

A barrister's dictation has passed through other hands for as long as there have been Dictaphones: typists, secretaries, transcription bureaus, couriers carrying tapes. None of that ever stripped privilege from the underlying material, because those parties handled it in confidence, as the lawyer's agents, to help produce the legal work. The law has never required that privileged material be touched only by counsel's own fingers — it requires that confidentiality be maintained.

An AI processor operating under contract, on your documented instructions, bound to confidentiality, sits in the same structural position as the typing bureau always did. In one respect it sits in a better one: with a properly built AI pipeline, no human outside your practice ever hears the recording at all. The typing bureau could never say that.

Where the real risk lives

The dividing line isn't AI versus human. It's controlled disclosure versus uncontrolled disclosure. A consumer AI tool whose terms of service let the vendor retain your content, use it to train models, or route it through whichever data centre is cheapest that day is not a confidential agent — it's a disclosure you can't call back. That's where the privilege analysis, the GDPR analysis and plain professional prudence all converge on the same answer: don't.

The professional bodies and the courts have been urging exactly this caution about putting client material into general-purpose AI tools. The caution is warranted. It just shouldn't be mistaken for a verdict on the technology — it's a verdict on the terms under which most of it is offered.

Privilege was never about the tool. It's about confidentiality — who can see the material, and under what obligations.

The six questions to put to any AI dictation vendor

1. Does your content train their models? If the answer is anything but an unqualified, contractual no — stop here.
2. Where is audio processed and stored? EU regions, or a third country? If any leg runs through the US, you inherit a transfer analysis and the residual CLOUD Act question that comes with US providers.
3. Will they sign an Article 28 Data Processing Agreement? A processor acts on your documented instructions. A tick-box terms of service is not that.
4. Who — human or machine — can access the material? Ask specifically whether vendor staff review recordings or outputs, and under what controls. Encryption at rest and in transit should be table stakes.
5. What are the retention and deletion terms? You should be able to delete a matter and mean it.
6. Is there an audit trail? For a practice, knowing who touched what and when is the difference between an incident and an unanswerable letter.

GDPR is the enforcement arm of the same idea

Privilege is the professional duty; GDPR is the statutory overlay that puts teeth in it. An AI dictation vendor processing client material is your processor under Article 28, which means a signed DPA, appropriate technical and organisational measures under Article 32, and — if data leaves the EU — Standard Contractual Clauses plus a transfer impact assessment. If there's a notifiable breach, Article 33 gives you 72 hours to tell the Data Protection Commission. Every one of those obligations is easier to meet when the entire stack, model included, stays in the EU — because the hardest questions never arise.

How clerk& answers the six questions

• Training: your dictations, briefs and documents are never used to train any model — not by us, not by our sub-processors. Contractually guaranteed.
• Location: matter content is stored in Dublin and Belgium, and all AI processing runs on Google Vertex AI in Belgium. Nothing transfers to the United States.
• DPA: a signed Article 28 agreement comes back within 48 hours of asking.
• Access: processing is automated end-to-end — no human outside your practice hears the audio. TLS 1.3 in transit, AES-256 at rest, and Postgres row-level security scoping every read and write to your workspace.
• Retention: delete a matter or your account and the content goes with it; the specifics are in the privacy policy, in plain language.
• Audit: every document carries an activity log — who dictated, who edited, who sent.

The honest answer

Is AI dictation safe for privileged material? Built and contracted properly — EU-only processing, no training on your content, an Article 28 DPA, no third-party humans in the loop — it preserves confidentiality at least as well as the dictation workflows the bar has trusted for decades, and better than audio files sitting in an inbox waiting for a typist. The tool was never the risk. The terms are. Ask the six questions, read the answers like the lawyer you are, and hold every vendor — us included — to them.