Best AI Transcription App for Lawyers, Therapists, and Journalists in 2026

The three professions that most need accurate transcription are also the three that can least afford a data breach. Attorneys owe clients a duty of confidentiality under ABA Model Rule 1.6. Therapists are bound by HIPAA and state licensure rules. Journalists have an ethical, and in many jurisdictions legal, obligation to protect sources.

Yet the dominant AI transcription products of 2024 and 2025 were built for sales teams. Their default architecture — upload audio to a vendor cloud, run a large server-side model, store the transcript indefinitely, and (in many cases) use the content to train the next model — is a reasonable trade for a sales call. It is a professional-responsibility problem for a deposition prep session, a therapy intake, or a background interview with a whistleblower.

This guide lays out a framework for evaluating AI transcription apps when the conversation itself is confidential, and compares the architectures available in 2026.

What changed in 2024–2026

Three developments reshaped the confidentiality analysis for AI transcription:

• ABA Formal Opinion 512 (July 2024) made clear that lawyers using generative AI must understand how the tool handles client information, obtain informed consent where appropriate, and consider confidentiality, competence, and supervision duties. Read it directly at the ABA.
• NYC Bar Formal Opinion 2025-6 (December 2025) extended that analysis, addressing vendor due diligence and the specific risks of tools that retain or train on client data.
• US v. Heppner (S.D.N.Y. Feb 2026): Judge Rakoff held that a litigant's chats with a public AI platform were not privileged, drawing an explicit analogy to the third-party doctrine. The decision does not govern every vendor arrangement, but it signals how courts are beginning to treat voluntary disclosures to AI providers.

Add West Technology Group v. Sundstrom (D. Conn. 2024) on preservation obligations for AI-generated work product, and the 2025 Brewer v. Otter.ai class action over recording and transcript handling, and the landscape for cloud-based transcription in confidential practice has narrowed considerably.

The buyer's framework: four questions that matter

For lawyers, therapists, and journalists, the right AI transcription app is the one that can answer four questions well:

1. Where is the audio processed? On the user's device, or on a vendor server?
2. Who holds the data at rest? The user (in their own storage account), the vendor, or a chain of subprocessors?
3. Is content used for model training? Ever. By default. Opt-out or opt-in.
4. What is the subpoena surface? If a third party serves the vendor tomorrow, what can the vendor produce?

Marketing pages rarely answer these directly. The answers live in the Data Processing Addendum, the subprocessor list, and the retention policy. If a vendor will not sign a DPA, that is itself an answer.

Cloud transcription: the default architecture, and its costs

Most well-known AI transcription tools — Otter, Fireflies, Fathom, Read.ai, Grain, tl;dv, and the various Zoom/Teams/Meet built-ins — share a broadly similar shape. A bot joins the call, or audio is uploaded after the fact. The audio is transcribed on a vendor-operated GPU cluster. The transcript is stored in the vendor's database. A chain of subprocessors (hosting, LLM inference, analytics, email) each receive some slice of the data. Enterprise plans typically disable training-on-content; free and lower-tier plans often do not.

For a sales organization, this is fine. For a lawyer debriefing a client, a therapist conducting an intake, or a reporter meeting a confidential source, the architecture creates concrete problems:

• Subpoena surface. The vendor and every subprocessor is a potential discovery target. The content exists on their infrastructure, so it is producible.
• Breach surface. Each hop adds a party whose security posture the professional has not personally vetted.
• Consent complexity. Informed consent under Model Rule 1.6 or HIPAA requires the professional to actually understand and disclose where the data goes.
• Bots in the room. A visible bot named "Otter" or "Fireflies" on a Zoom call with opposing counsel or a patient changes the dynamic, and in some jurisdictions triggers two-party consent analysis.

On-device transcription: the alternative

Modern Apple Silicon ships with a Neural Engine capable of running Whisper-class speech models locally. The user's Mac does the transcription. No audio is uploaded to a vendor, because the vendor never has a server in the loop. Storage lives in the user's own iCloud account, under the user's Apple ID — not in a vendor database.

This architecture collapses most of the questions above:

• Audio is processed on the device.
• Data at rest sits in storage the user controls.
• There is no vendor content to train on, because the vendor never receives it.
• The subpoena surface for the vendor is essentially empty with respect to recordings and transcripts, because the vendor does not hold them.

It is not magic. On-device models have real constraints: they run on the user's hardware, so battery and thermals matter; model size is bounded by device RAM; and if the user enables iCloud Sync, the library (audio included) lives in the user's iCloud account with Apple. That is a trade-off the user can see and control, not a hidden subprocessor chain.

Comparison: architectures at a glance

What each profession should weigh

Lawyers

The analysis flows from Model Rule 1.6 and ABA Formal Opinion 512. A lawyer using any AI tool must understand how the tool handles client information and, where appropriate, obtain informed consent. A product that cannot answer "who holds the recording tonight" in one sentence makes that duty harder to satisfy.

The Heppner decision is a useful stress test. If a vendor holds the transcript, and a third party serves the vendor, the lawyer is in an argument about whether the vendor relationship was sufficient to preserve privilege. If the vendor never received the transcript, that argument is much shorter. This is architectural risk reduction, not a guarantee — courts decide privilege on facts, not on marketing.

For a deeper look at how courts have treated AI vendors in discovery, our Basil for Law pillar page collects the relevant opinions and ethics guidance in one place.

Therapists and healthcare professionals

HIPAA's Privacy and Security Rules treat protected health information strictly, and the Business Associate Agreement framework assumes that if a vendor processes PHI, a BAA is required. The simpler architecture is one where the vendor does not process PHI at all because the audio is processed locally.

Therapists should also weigh the clinical effect of a visible meeting bot on rapport. On-device capture avoids that entirely for in-person sessions, and for telehealth sessions it means no unfamiliar participant appears in the patient's view.

Journalists

Source protection is both an ethics obligation under the SPJ Code of Ethics and, in many jurisdictions, a legal one under shield laws. The threat model is specific: if a reporter is subpoenaed over a story, any third party who holds the raw interview audio is a potential alternate target.

A cloud transcription vendor that holds the recording is an alternate target. An on-device processor that never receives the recording is not. The Reporters Committee for Freedom of the Press has long advised journalists to minimize the number of parties holding sensitive source material.

Features that matter after the privacy question is settled

Once the architectural question is answered, the usual feature comparison applies:

• Accuracy on domain vocabulary. Legal, clinical, and technical terminology varies widely. Test on your own recordings.
• Virtual meeting support. Many on-device tools, including Basil, capture Zoom, Teams, and Meet audio on macOS without joining the call as a bot.
• In-person support. For client meetings, intakes, and interviews, local microphone capture matters as much as virtual-meeting support.
• Organization. Matter-based or case-based grouping for lawyers; client-based grouping for therapists; story-based grouping for journalists.
• Labeling and consent logs. Privileged & Confidential headers, consent timestamps, and attestations are common requirements in regulated practice.
• Export. Plain text, DOCX, PDF, and the ability to purge.

For a longer treatment of the recording-consent question specifically, see our guide on one-party vs. two-party consent for AI meeting recording.

How to evaluate a vendor in 20 minutes

A short, concrete checklist to run before a trial:

1. Ask for the DPA. If one is not available for solo or small-firm tiers, note it.
2. Ask for the current subprocessor list. Count the entries.
3. Ask, in writing: "Is customer content used to train your models under my plan, by default?"
4. Ask: "Where is audio processed, and where are transcripts stored?"
5. Ask: "What is your response to a third-party subpoena for my customers' recordings?"
6. Check the retention policy. Default retention of "indefinite" is a flag.
7. Check whether a bot joins meetings, and whether that bot's presence is visible to other participants.

The answers go in the file for the matter, client, or story. Under Opinion 512, documenting the diligence is part of the diligence.

How Basil approaches this

Basil was built by a practicing lawyer, specifically for work that cannot tolerate a vendor cloud in the middle. Transcription and summarization run on the Apple Neural Engine, on the user's Mac. Basil has no server and no subprocessors, and Basil never receives recordings, transcripts, or summaries. iCloud Sync is on by default, which means the user's library — audio included — is stored in the user's own iCloud account with Apple, not with Basil; users can turn it off in Settings › iCloud.

For virtual meetings, Basil captures Zoom, Teams, and Meet audio on macOS in Computer mode, without joining the call as a bot. For in-person work, Basil captures from the device microphone the same way. The Basil for Law edition adds privilege attestations, consent logs, matter organization, and Privileged & Confidential labeling — $19.99/month or $199.99/year, with a 3-day free trial on monthly and a 7-day free trial on annual, one trial per Apple Account. The general Basil app is available with a free tier of 60 minutes per month.

Basil signs DPAs and NDAs on request. The architecture is the point: the vendor cannot produce what the vendor never received.

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

Keep client conversations on your device

Basil makes transcripts and summaries on your device. No bot joins the call, and Basil never receives your client conversations. See Basil for Law → · Legal-tool reviews →

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