I recently had occasion to answer an ethics question about what AI transcription tools are safe for attorneys. If you have researched this and could not reach a solid conclusion, I am not surprised. The question is genuinely hard, and I have wrestled with it myself.
Here is the fundamental problem. AI transcription tools that upload the conversation so that they can use AI to transcribe the content introduce a third party into privileged conversations. While many companies promise privacy in their marketing, these are just vendor statements. Vendor statements are not enforceable confidentiality commitments.
Attorneys Cannot Just Trust Vendor Promises
When it comes to attorney/client confidentiality, we cannot simply trust the vendors’ promises. We need to do a close analysis of the specific tool’s privacy policy and terms of service. This is straightforward enough, though I find that like many cloud-based businesses, the terms of service tend to leave a lot of wiggle room. Beyond the wiggle room, the problem is that most transcription tools are built on top of other AIs. Those AIs may or may not protect confidentiality. Rarely do the transcription tools tell us what the other AIs are doing with the data, nor where that data is traveling. They may or may not identify the other AIs, and if they do, they don’t tend to tell what level of AI is being used. For example, if the transcription tool goes through Claude or ChatGPT, is it using the enterprise version of those tools? And if it is, where is the data going? Is it traveling to the UK, the EU, or even China, where privacy requirements are quite different from the US?
In short, there is a gap between the marketing language and even the terms of service/privacy policies of the transcription tools and the full processing chain. That gap is why it is hard to figure out which tools are acceptable and which are not. We attorneys are obligated to make reasonable efforts to make certain that we protect our clients’ data under ethics rule 1.6. To me, reasonable efforts include determining not only what the tools we purchase do with the data, but also, what the tools they are built on top of do with the data. As a result, to clear any AI tool, including transcription, we need to analyze the terms of service/privacy policies not only for the tools themselves, but for the underlying tools, including where those tools send the audio and the transcription.
My Transcription Tool Recommendations
New transcription tools are being introduced every day, but the analysis I described takes time to perform properly. In the meantime, based on my own review, there are two setups I am comfortable recommending. The first I have recommended previously, Microsoft Teams with a microphone.
Recommendation 1: Teams for client meetings
For client meetings, I recommend using Microsoft Teams and Copilot with transcription enabled, using a Teams-certified conference microphone. For example, you could look into microphones by Jabra, Poly, and Anker. These microphones cost roughly $100 to $200. When you are not in the office or in a room without a computer, a Bluetooth speakerphone and the Teams mobile app work well.
Why do I find this particular setup acceptable? The transcript is stored in your own Microsoft 365 tenant, and the processing stays inside what Microsoft calls the Microsoft 365 service boundary. It is important to note that Microsoft does route model processing to whatever region has capacity, so unless you have paid for data residency, it does not promise the processing stays in one country.
If you use Teams, you do not need to be concerned about a new vendor or unknown sub-processor chain. Along with Microsoft comes a data protection agreement, retention policies, compliance controls, and audit rights. Further, under Microsoft’s enterprise terms, tenant data is not used to train foundation models. Please make certain to confirm this in your own license terms, because it depends on your plan. As long as you have the right plan, you are simply extending a tool your practice can already trust with client data.
Conditions for This Setup
You must have Microsoft 365 with Teams licensing that includes transcription and Copilot, so verify your plan before relying on this, and make certain transcription is turned on. Some states, like Pennsylvania, are all-party consent states under the state wiretap act. This means all people involved in the conversation must consent to the recording. Other states are one-party consent. This means only one party in the conversation needs to consent. That said, even in one-party consent states, I recommend getting everyone’s consent for meetings. I specifically recommend documented consent from every participant, covering both recording and AI transcription, at the start of every meeting.
Teams Caveats
Two things to remember. You must review every transcript and summary before relying on it, because AI transcripts contain errors and occasional invented dialogue. And you will need a workflow that includes starting the meeting and enabling transcription. Test the process before you use it with a client.
Recommendation 2: Whisper, run locally, for your own dictation
If you want to dictate for yourself, Whisper is a good choice. Whisper is OpenAI’s open-source speech-to-text model, free to download and run entirely on your own computer. If you run it locally, the audio never leaves your machine. No third party ever receives the communication.
Normally this is a two-step process. You record yourself, then run the recording through Whisper, most commonly as an mp3 or a .wav file. There are any number of ways to record yourself. There is one exception to the two-step process. It is called Const-me WhisperDesktop, which captures and transcribes live from a microphone with no separate recording needed. It is available via GitHub.
Remember, this is for solo dictation and for transcribing recordings you already lawfully have. It is not for live multi-party client meetings. While Whisper can transcribe all of the voices in a room, it does not identify who is speaking separately. This would create a very confusing transcription.
Whisper Caveat
The caveat is that local Whisper gives you architectural security with no vendor behind it. The burden of verifying the setup is yours. So a few rules:
- Download only from the official source, listed below. Do not search for “Whisper download,” because there are lookalike sites.
- Use local processing only. Never the OpenAI API option, which sends audio to OpenAI’s cloud and puts you right back in the third-party problem.
- Test before trusting. Disconnect from the internet and confirm it still transcribes. If it needs a connection, it is not local.
- If you use a consumer app built on Whisper rather than the model itself, check for cloud sync features and re-check after every update. Many “Whisper” apps quietly add cloud components.
One more warning about the name. Dozens of products are called Whisper-something. The name means nothing. Many of them send audio to the cloud. The disconnect test is the only way to be sure.
These download links were all verified via GitHub as of July 2026:
- The official model: https://github.com/openai/whisper. OpenAI’s repository. It requires Python and the command line, so it is the most technical option, but it is the canonical one.
- whisper.cpp: https://github.com/ggml-org/whisper.cpp. The lightweight version most local apps are built on. Also command line.
- Windows, point-and-click: https://github.com/Const-me/Whisper. Download WhisperDesktop.zip from the Releases section. It can transcribe files and live microphone audio. Be aware that the website whisperdesktop.com impersonates this project and should not be trusted. Download only from the GitHub Releases page.
- Mac: apps like MacWhisper exist, but I have not verified their current data handling, so the disconnect test still applies.
Final Thoughts
While neither of the options that I suggest is as simple as downloading an app or purchasing a tool with built-in transcription, my recommendations are a reasonable step to using transcription while protecting client data. That, of course, is the whole point.