Dictation for Lawyers: Privilege and Confidentiality

Lawyers can use dictation and cloud transcription. The ethics rules ask for something more specific: reasonable efforts to keep client information confidential, which in practice means knowing where the audio goes, who can hear it, how long it is kept and whether the vendor uses it to improve or train its products. The most important single check is whether the tool learns from what you put into it. If it does, and its outputs could disclose client information, the ABA says the client's informed consent is required before you enter that information ¹. If the audio is processed on your own computer and no vendor stores it, the vendor-side risk goes away.

This guide reads the ABA opinions and the state bar opinions directly, explains when a third-party tool can affect privilege and when it cannot, and ends with five questions to put to any vendor before client audio touches it. For the broader picture of where your voice goes, see our overview of voice privacy and security and whether voice to text is safe.

Lawyers may use dictation tools if they take reasonable steps to protect client information

The duty comes from Model Rule 1.6(c), which requires a lawyer to make "reasonable efforts to prevent the inadvertent or unauthorized disclosure of, or unauthorized access to, information relating to the representation of the client" ². Individual states adopt their own version. Arkansas, for example, requires the same reasonable efforts in its Rule 1.6(c) ³.

The standard is a judgment, not a checklist. The ABA says that what counts as reasonable efforts "is not susceptible to a hard and fast rule, but rather is contingent upon a set of factors" ⁴. Arkansas's comment to its rule names the factors, including how sensitive the information is and how likely disclosure is without extra safeguards, and the cost and difficulty of those safeguards ⁵. For dictation, the sensitivity of the matter and the vendor's terms are the two factors that matter most.

Six ethics opinions now address AI tools and client confidentiality, and their requirements overlap.

The Texas opinion asks for a specific judgment. The lawyer "should be reasonably satisfied that the program will not reveal confidential information to others or permit the use of such information to the disadvantage of the client" ⁶. The test rests on the lawyer's own judgment, so that judgment has to be made against the tool's actual terms.

The stakes are rising because more lawyers use AI tools. In the ABA's Legal Technology Surveys, 11 percent of respondents were using AI-based technology in 2023 ⁷, and 30 percent were in 2024 ⁸. That is about 2.7 times the share in one year ⁹. The survey counts AI-based technology tools broadly, not generative AI alone, so read the figures as a trend.

Sending recordings outside the firm is not new. Legal Futures reports that digital dictation recordings can be encrypted in real time and then uploaded to a cloud platform, or routed over a local network, an FTP server or email for transcription ¹⁰. A Pennsylvania firm had its lawyers record notes on smartphones and upload the audio files to the firm's data center for transcription in 2014 ¹¹. Using an outside transcriber is an established practice. What the rules ask is that the firm can show it chose the vendor and its settings on purpose.

ABA Formal Opinion 512 applies the confidentiality duty to AI tools that learn from inputs

ABA Formal Opinion 512 is the ABA's opinion on generative AI. Its confidentiality section starts from Rule 1.6 and then deals with a specific risk: tools that learn from what users enter. The opinion says that self-learning tools "by their very nature, raise the risk that information relating to one client’s representation may be disclosed improperly" ¹².

That is why consent is the central requirement. For self-learning tools whose outputs could disclose client information, the opinion concludes that "informed consent is required" before the lawyer enters client information ¹.

A general clause in the engagement letter is not enough. The opinion says that "merely adding general, boiler-plate provisions to engagement letters purporting to authorize the lawyer to use GAI is not sufficient" ¹³. A firm that wants to use a self-learning tool needs more than a boilerplate clause.

For a shorter answer to what the opinion covers, see what ABA Formal Opinion 512 says about AI tools. Keep the opinion's scope in mind. Its consent requirement is about tools that learn and may disclose. A transcription engine that processes audio, returns text and keeps nothing is a different case, but you only know which case you are in by reading the vendor's terms. The ABA opinion is persuasive guidance. Your state's rules and state bar opinions decide what applies to your practice.

Sending audio to a vendor can keep privilege intact, but only when the vendor acts for the lawyer and keeps the information confidential

Privilege protects confidential communications made to obtain legal advice. Courts have not limited it to lawyers' own offices. In United States v. Kovel, the Second Circuit held that an accountant whom a firm used to help it advise a client could be covered, and the court wrote that "the presence of the accountant is necessary, or at least highly useful, for the effective consultation between the client and the lawyer which the privilege is designed to permit" ¹⁴. Kovel concerned a person, not a software product, so applying it to a dictation vendor is an analogy. It does show that courts ask what role the third party plays.

The opposite rule matters too. Cornell's legal encyclopedia notes that if a third party is present during a privileged communication, confidentiality may be compromised unless the third party is essential to the attorney-client relationship ¹⁵. A cloud transcription service is not in the room, but the same test applies in spirit: is the vendor there to serve the lawyer's advice, and is it bound to keep the content private?

The Texas opinion puts that test in terms of vendor trust. For third-party service providers, the lawyer must be "reasonably confident that the confidential character of the information will be respected and protected by the service provider" ¹⁶. The test is about the vendor's conduct and terms, not whether the audio left the device.

Human access is one place where vendor conduct has gone wrong. In 2019 The Guardian reported that Apple contractors regularly heard confidential medical information, drug deals and recordings of couples having sex while doing quality control for Siri ¹⁷. Google said that one of its language reviewers had leaked confidential Dutch audio in violation of its data security policies ¹⁸. Neither case was a court ruling on privilege. Both show why a firm should ask who can listen to audio, not only where it is stored.

We did not find a court ruling that a dictation vendor waives privilege by processing audio. The practical question is narrower: does the vendor's contract and settings limit access and use to providing the service?

Courts have started to rule on whether AI conversations are privileged

A New York federal court has now ruled on the question. On February 13, 2026, Judge Jed S. Rakoff of the Southern District of New York held that a non-attorney's communications with a generative AI platform are not protected, and the court found the work product claim failed too ¹⁹. Commentators differ on the exact date, so check the docket in United States v. Heppner before citing it.

The defendant used Claude, a publicly available AI assistant, to analyze his legal situation. He did so "on his own initiative, without direction from his attorneys" ²⁰. That detail limits the ruling. The court did not decide what happens when a lawyer directs an AI tool as part of representing a client, and a dictation tool that only converts a lawyer's speech to text is further from the facts than a chatbot conversation about case strategy.

Two lessons carry over to dictation. First, the AI platform's terms decide how much confidentiality a user can expect, so read them before using any consumer tool for legal work. Second, anything a lawyer types or dictates into a chat-style assistant to get advice about a matter is a different kind of record from a transcript. If your firm uses an assistant for case work, the question is whether it is inside the firm's controlled environment and under the lawyer's direction. For that question, see whether conversations with AI chatbots are privileged and whether cloud dictation waives privilege.

State bars add their own guidance on cloud and AI tools

State bars add detail on process. The table below compares the three opinions we read in full or in the detailed text we could open.

Texas puts vetting first. Its opinion says that with any generative AI tool, the lawyer "should be reasonably satisfied" about confidentiality ⁶, and it adds that the lawyer "may need to secure client consent" when confidential information is used with these tools ²¹. Florida reaches a similar place, recommending that the affected client's informed consent be obtained before using a third-party program that would expose confidential information ²².

The three opinions differ in how firm the consent duty is. The ABA requires consent for self-learning tools that may disclose information, Texas says a lawyer "may need" consent, and Florida recommends it. Do not treat them as one uniform rule. Read your own state's opinion. New York also treats recording a call as a separate question, covered below.

AI notetakers on client calls need consent before they record

Most of the sharpest guidance on AI notetakers comes from New York. New York City Bar Formal Opinion 2025-6 says clients must be told, and must consent, whenever an AI-driven system records a call with them ²³. Its later opinion on AI recording of calls with non-clients says the default should be not to record unless there is a good reason ²⁴.

Recording law is separate from ethics. California's Penal Code 632 punishes a person who intentionally records a confidential communication without the consent of all parties ²⁵. In California, a notetaker that records a confidential client call without every party's consent can create liability even if the ethics rules were satisfied. For the state-by-state picture, see recording consent laws and the all-party consent term. For a direct answer on notetakers, see whether lawyers can use AI notetakers on client calls.

On-device dictation takes the third party out of the analysis

The simplest way to limit confidentiality risk is to keep recognition on the device. Apple's Mac Help says general text Dictation can be "processed on your device and not sent to Siri servers", depending on the language, so check the setting for yours ²⁶. On iPhone, Dictation is processed on the device in many languages and needs no internet connection ²⁷. Recognition that stays on the device means no vendor receives the audio at all, which is the strongest answer to the third-party question. For how this works in general, see on-device speech recognition and Apple dictation privacy on Mac.

On-device does not mean every setting is safe. Apple's dictation settings include an opt-in to Improve Siri and Dictation, which stores and reviews a sample of audio ²⁸, and Apple says only its own employees, under confidentiality obligations, can access that audio ²⁹. Dictation sent to Apple's servers is not stored unless you opt in ³⁰. A firm should turn off the opt-in and confirm the setting on each device.

Cloud recognition is not automatically unsafe. Microsoft documents that for real-time speech to text in Azure, audio "is processed only on the Azure's server memory, and no data is stored at rest" ³¹. That is a vendor storing nothing, which is a different risk profile from a vendor keeping recordings. It is still a third party's server, so the contract question remains.

Yap, which we make, is an on-device dictation app for Mac and Windows, so recognition runs on the computer you are using. Check the specific claim for your platform and version in the app's own documentation before relying on it for client work. Current versions are on the download page.

Ask every vendor five questions before client audio touches it

The table below compares what four widely used tools say in their own documentation. Confirm each answer against the current policy before you rely on it, because these terms change.

Google's default is the one to note. Its documentation says "By default, Cloud Speech-to-Text does not log customer audio data or transcripts" ³², and logging is something the customer opts into. If logging is on, Google says only a limited set of authorized Google employees and contractors can access the data ³³. ChatGPT dictation keeps the audio for as long as the chat exists in your history ³⁴. Apple's history of Siri and Dictation requests can be kept for up to two years, linked to a random identifier rather than your Apple Account ³⁵.

Ask these five questions before a vendor's tool touches client audio:

1. Where is recognition done? Ask whether audio is processed on the device, in the vendor's memory, or stored on its servers. Microsoft's real-time path and Apple's on-device path are different answers to this question.
2. How long is audio or a transcript kept, and can you delete it? Ask for the default retention period and the deletion route. The retention periods above range from nothing stored at rest in Microsoft's real-time path to up to two years for Apple's request history ³⁵.
3. Does the vendor use your audio to train or improve its products? Ask for the setting and its default. Google's data logging is off unless the customer opts in. Anthropic says its mobile apps delete the recording after converting speech to text and do not use it for training ³⁶.
4. Who can listen to the audio? Ask whether human reviewers ever hear recordings, who they are and what confidentiality duties bind them. Human review has produced leaks and exposure in the past, as the Apple and Google cases show.
5. Will the vendor commit to these answers in writing? The Texas test is whether the lawyer is reasonably confident the provider will protect the information. Get the answers in the contract or the settings you can document, and keep the record. Florida's opinion adds the client consent question if the tool exposes confidential information.

A firm can answer these questions in one page and keep the page with its matter files. For how to turn those answers into a written rule for everyone who dictates, see dictation at work: enterprise and government policies. For how to find out whether a vendor trains on your voice, see do dictation apps train AI on your voice. For a lawyer's specific tools, see legal dictation software for lawyers and law firms. For the transcription side of legal work, see legal transcription: the job, formats and pay.

Lawyers who dictate for clients can use cloud tools, but the defensible choice is one the firm can explain: a tool whose processing and storage are documented, whose settings are set on purpose, and whose consent duties are met where the opinions require them. Where the answer is unclear, use on-device recognition for client material until the vendor's terms are clear. Ask your state bar and your firm's ethics counsel before you rely on any reading in this guide for a specific matter.