Deposition Transcription App Private: Keeping Witness Interviews Off the Cloud

Deposition preparation is one of the highest-stakes conversations a litigator ever has. You are testing a witness's memory, exposing weak spots, and often previewing the exact narrative opposing counsel will try to dismantle under oath. The last place that conversation belongs is on someone else's server.

Yet the default workflow for most litigation teams in 2026 involves uploading witness prep audio to a cloud transcription service, letting a third-party AI summarize it, and hoping the vendor's retention policy and security posture hold up. This article walks through why that default is increasingly risky, what the current authorities actually say, and what a genuinely private deposition transcription app looks like from an architectural standpoint.

Why Deposition Prep Is a Uniquely Sensitive Recording

A witness prep session is not a routine client meeting. It typically contains, in the same recording:

• Attorney mental impressions and case theory (classic Hickman v. Taylor opinion work product).
• Draft narrative choices — what to emphasize, what to concede, what to say "I don't recall" to.
• The witness's unrehearsed first answers, which may differ from later testimony and become impeachment gold if disclosed.
• Candid discussion of documents, weaknesses, and prior inconsistent statements.

Under ABA Model Rule 1.6, an attorney must make reasonable efforts to prevent the unauthorized disclosure of, or unauthorized access to, information relating to the representation. When the recording of a prep session lives on a vendor's servers, "reasonable efforts" has to include an honest assessment of who else can reach it: the vendor's engineers, its subprocessors, its subpoena-response team, and any attacker who breaches it.

The Third-Party Problem After Heppner

The most important recent decision for lawyers thinking about AI tools is US v. Heppner (S.D.N.Y. Feb 2026), where Judge Rakoff held that a litigant's chats with a public AI platform were not privileged. The reasoning tracked the classic third-party doctrine: once you voluntarily hand content to an outside service, the expectation of confidentiality that supports privilege is weakened.

Heppner was about a pro se litigant using a consumer chatbot, not an attorney using a vendor transcription tool with a signed agreement. But the underlying analytical move — treating disclosure to an AI service as disclosure to a third party — is exactly the argument opposing counsel will make about your cloud-hosted prep recordings. A subpoena or a motion to compel does not need to win to hurt you; it only needs to force you to litigate the question, produce logs, and explain to your client why the recording of their preparation is now the subject of a discovery fight.

The safest architectural answer is that there is nothing to fight over: the audio and transcript never left the device.

What ABA Formal Opinion 512 Actually Requires

In July 2024, the ABA issued Formal Opinion 512 on generative AI. It is worth reading in full, but three obligations are directly relevant to deposition transcription:

1. Competence (Rule 1.1): understand, at a reasonable level, how the tool handles your inputs.
2. Confidentiality (Rule 1.6): evaluate whether inputs will be used to train models, retained, or accessible to the vendor.
3. Informed client consent when the tool's use of client information goes beyond what a reasonable client would expect.

The NYC Bar Formal Opinion 2025-6 (December 2025) built on this, reinforcing that lawyers must understand a tool's data flow before feeding it privileged material. "The vendor says it's secure" is not diligence; reading the DPA, understanding retention, and knowing whether human reviewers ever touch the data is.

A cloud transcription vendor can be perfectly reputable and still fail this test if its architecture allows the recording to be accessed by anyone other than you. The test is not the vendor's good intentions. It is the structure of who can reach the file.

The Otter Problem: When the Notetaker Is the Defendant

The risks are not hypothetical. In Brewer v. Otter.ai (2025), the plaintiff alleged that Otter's transcription service captured and processed conversations in ways users did not adequately understand. Whatever the ultimate merits, the case illustrates the structural problem: when a third-party notetaker sits in the middle of your conversations, its practices become your problem. Its litigation becomes discoverable context around your files.

Compare this to West Technology Group v. Sundstrom (D. Conn. 2024), where the court examined whether communications routed through certain third-party channels retained their privileged character. The recurring theme in this line of cases is that architecture — who holds the data, who can access it, what the terms of service permit — matters more than the label the parties put on the relationship.

What "Private" Should Actually Mean in a Deposition Transcription App

The word "private" has been diluted to near-meaninglessness in vendor marketing. For deposition prep, a useful working definition is that a tool is genuinely private only if:

• Audio is captured, transcribed, and summarized on your own hardware.
• Nothing is transmitted to the vendor's servers as a condition of normal use.
• The vendor has no ability to access the content, even under compulsion, because they do not possess it.
• There are no subprocessors in the transcription pipeline, because there is no pipeline outside your machine.

Anything less is a matter of policy rather than architecture. Policies change. Companies get acquired. Terms of service get updated. Architecture is harder to reverse.

Comparing Common Deposition Prep Transcription Workflows

The rightmost row is what "privilege-safe by architecture" means in practice. It is not a guarantee that no dispute will ever arise about a prep recording — a court can still order you to produce your own files if it finds waiver on independent grounds — but it eliminates the entire category of risk that involves a third-party vendor.

Consent, Recording Laws, and the Witness in the Room

A private transcription architecture does not solve consent. It is still your responsibility to know whether your jurisdiction is one-party or all-party consent for recording, to disclose recording to the witness where required, and to document that disclosure. Many prep sessions occur in all-party states like California, Florida, and Illinois, where every participant must be informed.

A well-designed prep workflow therefore includes:

• A spoken notice at the start of the session that the meeting is being transcribed for attorney work-product purposes.
• A written consent log tied to the matter file.
• Clear labeling of the resulting transcript and summary as Privileged & Confidential — Attorney Work Product.
• A retention decision made deliberately at the matter level, not defaulted to "forever in the cloud."

None of this is unique to AI transcription. It is what careful litigators have always done with prep recordings. AI tools should make it easier, not harder, to maintain those habits.

Practical Checklist Before You Record a Prep Session

1. Confirm the jurisdiction's consent rule and give the required notice on the record.
2. Confirm your tool's data flow: is anything transmitted off your device during or after capture?
3. Confirm retention: where does the file live, for how long, and who else can reach it?
4. Label outputs as work product from the moment they are generated.
5. Store the transcript with the matter, not in a general "AI notes" folder that mixes matters.
6. Decide, per matter, whether the recording itself or only the transcript should be preserved.
7. If you use any cloud AI tool, make sure the DPA, retention policy, and training-opt-out are documented in your file.

For related reading, see our notes on Basil for Law and the broader question of how on-device AI changes the privilege analysis for meeting notes.

How Basil approaches this

Basil is a fully on-device meeting notetaker built by a practicing lawyer. Audio capture, transcription, and summarization all run locally on the Apple Neural Engine. Nothing is uploaded. There is no Basil server that holds your recording, no subprocessor in the transcription pipeline, and no ability for us to access your files — because we never receive them. For in-person prep, you record on your Mac. For virtual prep sessions on Zoom, Teams, or Meet, Computer mode captures the call on-device without a bot joining the meeting.

The general Basil app is available today with a 60-minute-per-month free tier. The Basil for Law edition — with privilege attestation, consent log, matter organization, and automatic Privileged & Confidential labeling — launches in August 2026 at $19.99/month or $199.99/year for solo attorneys, with a 3-day monthly trial or 7-day annual trial. Basil signs DPAs and NDAs on request.

The design goal is simple: reduce the risk that your deposition prep ever becomes a discovery fight about your notetaking tool.

This article is for information only and is not legal advice.

Frequently asked questions

Is an AI transcript of a witness prep session privileged?

The transcript itself is not what creates privilege; the underlying communication is. A prep session between attorney and client is generally covered by the attorney-client privilege and the resulting notes by the work-product doctrine. The risk with cloud AI tools is that routing the recording through a third party can complicate that analysis, as the reasoning in US v. Heppner (S.D.N.Y. Feb 2026) illustrates. Keeping the transcription on-device avoids introducing a third party in the first place.

Do I have to tell the witness the session is being recorded and transcribed?

In all-party consent jurisdictions, yes. In one-party consent jurisdictions, you as a participant can generally consent for yourself, but many careful practitioners disclose recording anyway and document the disclosure. Basil for Law includes a consent log tied to the matter to make that documentation routine rather than ad hoc.

What does ABA Formal Opinion 512 require me to do before using an AI transcription tool?

Formal Opinion 512 (July 2024) frames the analysis around competence, confidentiality, and informed client consent. In practice, this means understanding how the tool handles inputs, whether data is retained or used for training, and whether the client would reasonably expect that use. On-device tools simplify this analysis because there is no vendor data flow to disclose.

Is on-device transcription accurate enough for deposition prep?

Modern on-device speech recognition on Apple Silicon is good enough for the working purposes of prep — reviewing what the witness said, spotting inconsistencies, drafting outlines. It is not a substitute for a certified court reporter's transcript when one is required. Treat the AI transcript as attorney work product for internal use, not as a record for filing.

What happens if opposing counsel subpoenas my transcription vendor?

If your vendor holds the file, they can be served and will have to respond. Whether they must produce depends on the privilege analysis and the terms of your agreement, but you will be litigating the question. If the file never left your device, there is no vendor to serve. This is what we mean by 'no server to subpoena.'

Can I use Basil for deposition prep today?

The general Basil app is available now with a 60-minute-per-month free tier and captures both in-person and virtual meetings on-device. The Basil for Law edition, with privilege attestation, matter organization, consent log, and Privileged & Confidential labeling, launches in August 2026 at $19.99/month or $199.99/year for solo attorneys.

Keep client conversations on your device

Basil transcribes and summarizes entirely on-device — no cloud, no bot, no server to subpoena. See Basil for Law → · Legal-tool reviews →

This article is for information only and is not legal advice.