Voice privacy and security
Speech to Text Privacy: Where Your Voice Goes
Dictation is only as private as the place your voice is processed and the rules of whoever processes it. On-device recognition keeps the audio on your hardware, while cloud recognition sends it to a server that may store it, train models on it or let people review it. Voice is personal data everywhere and biometric data when it identifies you, so privacy laws, HIPAA, legal ethics rules and recording consent statutes all apply, and the reliable way to dictate privately is on-device recognition with history, training and sharing settings switched off.
Updated October 16, 202620 min read60 cited facts, each traced to its source
80MB
In 2019 Google moved Gboard voice typing onto the phone with an 80 MB speech recognizer that works even when the phone is offline, first on Pixel phones in American English.
Johan Schalkwyk, Google Research Blog, 2019
2years
Apple links Siri and Dictation request history, including transcripts, to a random rotating identifier rather than your Apple Account, and may keep it for up to two years.
Apple, Apple, 2026
0.2percent of audio snippets
Google said in July 2019 that language experts review only around 0.2 percent of all audio snippets.
Google, Google Blog, 2019
3seconds
A 2023 Microsoft research paper reported that its VALL-E model can synthesize personalized speech from only a 3-second enrolled recording of an unseen speaker.
Chengyi Wang et al., arXiv (Microsoft Research), 2023
Guides in this topic
Each guide takes one part of the subject in depth.
- Is Voice to Text Safe? How to Dictate PrivatelyVoice to text is safe when recognition runs on your device and the app keeps neither the audio nor the transcript. Cloud recognition may store audio, use it for training or let people review it, so check where recognition runs, how long text is kept, whether it trains AI and who can read it.
- Voice Data Retention: What Apple, Google and Microsoft KeepApple and Google do not keep your voice audio by default, and Microsoft makes voice clips optional, but each keeps something else: request history or chats. Apple keeps Siri and Dictation request history for up to two years, Google keeps Gemini app activity for 18 months by default, and Amazon lets you choose between saving recordings and automatic deletion.
- Is Whisper AI Safe? Privacy and Security ExplainedWhisper is safe for private audio when you run the open-source model on your own computer, because the audio never leaves it. The hosted OpenAI API sends audio to OpenAI, which keeps abuse monitoring logs for up to 30 days by default and does not train on API data unless you opt in.
- Is Dictation HIPAA Compliant? What Makes It SafeNo dictation app is HIPAA compliant on its own. Dictation is compliant when a covered entity uses it inside a compliant program: if a cloud recognizer receives patient audio, that means a signed business associate agreement plus encryption, access controls and a risk analysis. On-device dictation keeps the audio off the vendor's servers, but the device itself must still meet the Security Rule.
- Is Your Voice Biometric Data? GDPR, BIPA and CCPAA recording of your voice is personal data under almost every privacy law. It becomes biometric data in the legal sense when it is processed to identify who is speaking, because that process creates a voiceprint. That step triggers the strictest rules: explicit consent under GDPR Article 9, written consent and damages under Illinois BIPA, and sensitive-information limits under the CCPA.
- Do Dictation Apps Train AI on Your Voice?Some do and some do not. The answer sits in one clause of the privacy policy and one setting: Apple uses Dictation audio for improvement only if you opt in, Microsoft says it uses Copilot voice data for AI training unless an exception applies, and Otter's policy lists training on de-identified audio.
- Do Companies Listen to Your Voice Recordings?Yes. In 2019 reporters showed that human reviewers at Apple, Google, Amazon, Microsoft and Facebook listened to some voice clips, and each company then paused, limited or disclosed the practice. Today human review is mostly opt-in, and the only sure way to keep people from hearing your audio is to keep it off the company's servers.
- Recording Consent Laws: One-Party vs All-Party StatesFederal law requires the consent of only one party to a conversation, but several states, including California, Florida, Maryland and Washington, require every participant to agree. A call across state lines can fall under the stricter state's law, and the UK, EU and Canada regulate recording through other rules.
- Dictation App Permissions: Mic, Full Access, ClipboardA dictation app needs the microphone to hear you. The second permission, which lets it put text into other apps, can reach much further than your voice: Full Access on an iPhone keyboard, Accessibility on a Mac, or the clipboard on any system. Grant each one only when the app needs it, and review them in your system's privacy settings.
- Voice Cloning and Deepfakes: Risks and ProtectionA few seconds of recorded speech is now enough to make a voice clone that can fool people and some bank systems, so a familiar voice is no longer proof of identity. Scammers use clones to imitate relatives, executives and officials. The defenses that work happen outside the call: call back on a number you already have, and agree a safe phrase with the people you would be called by.
- Dictation for Lawyers: Privilege and ConfidentialityYes, with conditions. Lawyers may use cloud dictation when they make reasonable efforts to protect client information, which means knowing where the audio is processed, who can access it, how long it is kept and whether the vendor uses it to train models. ABA Formal Opinion 512 requires a client's informed consent before client information goes into self-learning AI tools.
- Dictation at Work: Enterprise and Government PoliciesMost organizations allow dictation when the tool has a data processing contract, a security certification and retention controls. They restrict consumer tools that may keep or train on what staff say, and governments use formal authorization programs and device bans.
Where your voice goes depends on where it is recognized. If the speech recognizer runs on your device, the audio stays there. If it runs in the cloud, your audio travels to a company's servers, and from that moment its storage, training and review rules decide who else can hear it. Everything else on this page (the law, the health and legal rules, the settings) follows from that one fact.
This overview covers the whole subject: on-device versus cloud recognition, what vendors keep, training on your voice, human review, voice as biometric data, HIPAA, attorney-client privilege, workplace policy, recording consent, app permissions, voice cloning and a short routine for dictating privately. Each section gives the answer and the evidence, and links to the guide that goes deeper.
Dictation is private when recognition runs on your device and the vendor keeps nothing
A spoken sentence passes through up to five places before it becomes text: the recognizer, the network, the vendor's storage, its model-training pipeline and its human reviewers. On-device recognition removes the network and everything after it. Cloud recognition adds all five, and each is a separate decision a vendor makes in its policy. Our guide Is Voice to Text Safe? walks through the five places and a quick test you can run yourself.
Vendors say plainly which mode they use. On a Mac, Apple says voice inputs and transcripts for general text Dictation are processed on your device and not sent to Siri servers 1, and the Keyboard settings show whether your dictation is on-device or sent to Apple 2. On iPhone, Apple says dictation is processed on the device in many languages and needs no internet connection 3. Google moved Gboard voice typing onto the phone in 2019, first on Pixel phones in American English, with an 80 MB recognizer that works offline 4.
Cloud dictation is still common, and it is not hidden. Speech recognition on a web page sends your audio to a server-based service by default 5. Microsoft says Microsoft 365 Dictate sends your speech to Microsoft to produce the text, though it does not store the audio or the text 6. Claude Code's voice dictation streams recorded audio to Anthropic's servers for transcription 7.
| Setup | Where the audio is recognized | What to check |
|---|---|---|
| Apple Dictation on Mac or iPhone (supported languages) | On the device | Keyboard settings show on-device or server mode |
| Gboard voice typing with offline language pack | On the phone | The offline pack must be downloaded |
| Browser speech recognition | A server by default | The browser vendor's service terms |
| Microsoft 365 Dictate | Microsoft's cloud | Not stored, but sent to Microsoft |
| Business speech APIs (Azure, Google Cloud, AWS) | The vendor's cloud | Retention and logging terms per service |
| Open models such as Whisper run locally | On your computer | Confirm the app does not call a hosted API |
| Alexa and similar smart speakers | The vendor's cloud | Recording and review settings |
You can check any tool in a moment. Switch on airplane mode and dictate. If text still appears, recognition is local. Our guides on open-source dictation software that runs offline and on offline speech to text on Windows cover the test and the best options, and the glossary entry for on-device speech recognition defines the term. For how the models behind both modes work, see our overview of speech recognition.
What vendors keep is different for each platform, and "not by default" is the common answer
Many large platforms have stopped keeping your audio by default, but they often keep something: a transcript, a request history under a random identifier, or logs for abuse monitoring. The chart below puts three published maximums on one axis.
The three numbers describe different things, so read them carefully:
- Apple. Dictation sent to Apple's servers is not stored unless you opt in to Improve Siri and Dictation 8. Request history, including transcripts, is linked to a random rotating identifier rather than your Apple Account and may be kept for up to two years 9.
- Microsoft Azure. For real-time speech to text in the Azure business service, audio is processed only in server memory and no data is stored at rest 10. That is a business service, not Windows dictation, and batch transcription lets you choose where audio and transcripts are stored 11. Our guide has the split for Microsoft.
- OpenAI API. Abuse monitoring logs are generated for all API feature usage and retained for up to 30 days by default 12. The glossary entry on zero data retention explains what a stricter arrangement involves.
- Amazon. Amazon told Echo owners, in an email that The Register reported in March 2025, that from March 28 voice recordings would be sent to and processed in the cloud and deleted after Alexa processes the request 13.
Deleting your history is a different control from opting out of improvement programs, and a business API is governed by different terms than the consumer assistant from the same company. The guide to voice data retention puts every vendor's rules and deletion controls in one dated table, and the questions on Google, Alexa and cloud speech APIs answer each directly. To check what your own device saves, see does voice typing save your audio and how to delete voice recordings.
Training on your voice is a policy choice, and Apple makes sharing opt-in
Whether a vendor trains models on what you say is stated in one clause of its policy and controlled by one setting. After 2019, Apple moved from sharing to asking. It said it would no longer retain audio recordings of Siri interactions by default and would let users opt in to improvement from audio samples 14. On a Mac, choosing Share Audio Recordings when you enable Dictation lets Apple store audio of your interactions and review a sample 15. You can reverse the choice under System Settings, Privacy & Security, Analytics & Improvements by turning off Improve Siri & Dictation 16.
Smaller AI dictation apps vary most, so read the clause. Aqua Voice says that with Privacy Mode off it may store transcript data on its servers to improve the product 17. Superwhisper says that on its Free plan dictation uses local Whisper models and nothing dictated leaves the device 18, while its Pro plan adds cloud models and AI processing 19. The HEX app's optional post-processing sends the transcript, the foreground app and the website hostname to the model provider you configure 20.
The pattern is the same in all three examples: a tool that is local by default can still send text to a cloud model when you turn on a rewriting or cleanup feature. Our question on whether AI dictation apps send your words to a language model lists what to look for. The guide Do Dictation Apps Train AI on Your Voice? covers the policies, why training matters (models can memorize rare phrases), what de-identification protects, and how on-device learning, called federated learning, improves a model without uploading audio. If you want to stop it, see how to stop apps training on your voice and the question on Otter.
People did listen to voice recordings, and Apple suspended human grading in 2019
In July 2019 the Guardian reported that Apple contractors regularly heard confidential medical information, drug deals and recordings of couples having sex while grading Siri 21. Apple told the paper that less than 1% of daily Siri activations were used for grading 22. Google had also said in July 2019 that one of its language reviewers violated data security policies by leaking confidential Dutch audio 23, and that its language experts review only around 0.2 percent of all audio snippets 24.
Small percentages are not small numbers when a service handles a very large volume of requests, which is why the disclosures mattered. Apple responded by saying it had immediately suspended human grading and begun a review of its practices 25. Years later, Apple moved to settle the litigation: Reuters reported in January 2025 that Apple agreed to pay $95 million to settle a proposed class action claiming Siri violated users' privacy 26. The settlement needed court approval. Today, Apple says that for people who opt in to Improve Siri and Dictation, only its own employees, under confidentiality obligations, can access the audio 27.
The timeline, the settlements and the current programs are in Do Companies Listen to Your Voice Recordings?. Related questions: did Apple contractors listen to Siri recordings, what the Google Assistant lawsuit was about, how often assistants activate by accident (see false accept) and whether your phone is listening for ads. A related finding for wake words: Microsoft says a Windows PC detects "Hey Copilot" locally and keeps only a 10-second audio buffer in memory that is never recorded or stored 28.
The only guarantee that no person hears you is that the audio never reaches a server.
Voice is personal data everywhere, and biometric data only when it identifies you
Law treats three things differently. A recording of speech is personal data. A transcript of it is personal data if it identifies someone. A voiceprint, a model of how a particular person sounds, is biometric data, and that triggers the strictest rules. Our guide Is Your Voice Biometric Data? reads the statutes directly. The short version follows.
GDPR. The regulation defines biometric data as personal data from specific technical processing of a person's characteristics that allows or confirms unique identification 29. Article 9 then prohibits processing biometric data for the purpose of uniquely identifying a person, along with other special categories, unless one of the listed exceptions applies 30. Recital 51 makes the key point for photographs: they count as biometric data only when processed through specific technical means that allow unique identification 31. We read the same logic as applying to audio, though the recital does not say so. On that reading, a dictation app that turns speech into text processes personal data, and a tool that builds a speaker profile to recognize who is talking processes biometric data. See the entry on special category data and the question is a voice recording personal data under GDPR.
Illinois BIPA. The Biometric Information Privacy Act defines biometric identifiers to include voiceprints, alongside fingerprints and iris scans 32. Its penalties are what made it famous: $1,000 for each negligent violation and $5,000 for each intentional or reckless violation 33. The guide covers the current version of the law, including later amendments, and the lawsuits over meeting assistants and voice analytics. See does BIPA cover voiceprints.
California and children's rules. California's privacy law lists as sensitive personal information the processing of biometric information for the purpose of uniquely identifying a consumer 34 (see is voice data sensitive under the CCPA). At the federal level, the FTC's amended children's privacy rule adds biometric identifiers that can be used to recognize a person to the definition of personal information 35.
So where does plain dictation fall? Usually in ordinary personal data, not biometric data, because it converts speech to words without identifying the speaker. The line is crossed by speaker identification, voice ID and tools that keep a speaker profile. Ask any voice tool whether it builds one and how long it keeps it. The question does speech to text collect biometric data answers it directly, and the glossary defines biometric data, voiceprint and speaker verification. Microsoft's documentation makes the same point about audio: speech recordings and their transcripts may be considered personal or sensitive data under various privacy laws 36.
HIPAA makes dictation a contract question: a vendor that hears patient audio needs a business associate agreement
No dictation app is HIPAA compliant by itself. What matters is how a covered entity (a clinic, a hospital, a health plan) uses it. The Privacy Rule allows a covered entity to let a business associate create, receive, maintain or transmit protected health information on its behalf only if the entity obtains satisfactory assurance that it will be safeguarded 37. In practice that assurance is a written business associate agreement, and it is why a cloud service that hears protected health information must sign one before a clinician dictates a patient note into it.
The Security Rule adds duties that apply even when audio stays on the device. It requires an accurate and thorough risk analysis of risks to the confidentiality, integrity and availability of electronic protected health information 38, and it lists encryption of that information as an addressable implementation specification 39. On-device dictation therefore removes the vendor from the picture but not the device security duties: a lost, unencrypted laptop with dictated notes is still a problem.
Consumer voice typing comes without a business associate agreement, so check the vendor's terms before you use it for patient information. One vendor, Superwhisper, recommends bringing your own cloud agreement: its documentation advises using your own API keys for organizations that already have HIPAA business associate agreements with an AI provider 40. Ambient AI scribes add a second question, whether patients must consent to being recorded, which we treat with the consent laws below.
The guide to HIPAA-compliant dictation quotes the rule text and sorts common tools. The questions on Microsoft dictation, Google voice typing, whether you need a BAA, on-device dictation and AI scribe consent each give a short answer. This is a description of the rules, not individual legal advice. For how clinicians use dictation day to day, see our overview of dictation at work.
Lawyers may use dictation if they protect client information, and the vendor question decides it
Attorney-client privilege protects confidential communications between a lawyer and client about legal advice. It can be compromised: if a third party is present during the privileged communication, confidentiality may be compromised unless that third party is essential to the attorney-client relationship 41. Whether a transcription vendor counts as such a third party depends on the contract, the vendor's access and the jurisdiction, so careful firms treat the question as one of vendor selection rather than a ban.
The professional-conduct rules point the same way. ABA Formal Opinion 512 applies the confidentiality duty of Model Rule 1.6 to generative AI 42. Commentary on the opinion says it concludes that informed consent is required before client information is entered into self-learning tools whose outputs could disclose it 43. That makes the vendor's training policy, its retention and its access controls the facts that matter. On-device dictation takes the third party out of the analysis, because nothing is sent.
Our guide Dictation for Lawyers: Privilege and Confidentiality reads the ABA opinions and state bar guidance and ends with five questions to ask every vendor. Related: does cloud dictation waive privilege, what Opinion 512 says, whether AI chatbot conversations are privileged and whether lawyers can use AI notetakers on client calls. The term is defined in attorney-client privilege. This is general information, not legal advice.
Employers restrict dictation for the same reason they restrict documents: speech carries the same secrets
Organizations treat dictation as a data flow. Approved tools with contracts and security reviews are allowed, and consumer cloud tools that may store or train on what staff say are restricted. The best-known case came in 2023, when CNBC reported that Samsung confirmed it was temporarily restricting generative AI on company computers after misuse of the technology 44, following a report that some staff had uploaded sensitive code to ChatGPT 45.
Governments approve cloud services through frameworks such as FedRAMP, and procurement teams check certifications and data processing terms. On-device tools fit strict environments, including air-gapped ones, because they need no connection. A workable policy names approved tools, banned content and retention.
The guide Dictation at Work: Enterprise and Government Policies documents real policies and gives an outline. The questions cover what your employer can see, why companies ban AI notetakers, government use and what a policy should include.
Recording a conversation is a separate legal question from dictating your own words
Dictating your own words is not recording a conversation. Recording a meeting or a call is, and the law of consent applies. Federal law says it is not unlawful to intercept a wire, oral or electronic communication when you are a party to it or one party has given prior consent 46. That is the one-party consent rule. Delaware follows it 47, and many other states do too. Others require everyone's agreement. California, for example, punishes anyone who intentionally records a confidential communication without the consent of all parties 48, an all-party consent rule. A call between two states can fall under the stricter law.
The same rule matters for tools that join meetings. An AI notetaker records the conversation, so its host needs to know the law of every participant's location, and a transcript is a record that can be requested later. The guide Recording Consent Laws: One-Party vs All-Party States maps the statutes with citations and covers the UK, the EU and Canada, where recording is also regulated through privacy law. The questions are which states require all-party consent, AI notetakers without consent, calls across state lines and recording calls in the UK and EU. The safe practice is simple: announce the recording and keep evidence of consent.
Permissions matter more than the microphone: text insertion is the powerful one
The microphone permission is the obvious one. The permission that puts text where you type is the powerful one, and its form depends on the platform: full access for a phone keyboard, Accessibility on a Mac, input simulation or the clipboard on Windows.
On iPhone, Apple's developer documentation says a custom keyboard runs in a sandbox that by default disallows network access, and that users must switch on Allow Full Access to change this 49. Apple's warning, quoted in a 2014 article, states that with Full Access enabled, developers are permitted to access, collect and transmit the data you type 50. A dictation keyboard that needs the network for cloud recognition therefore needs Full Access, while a fully on-device one may not.
Clipboard-based insertion leaves your words where other software can read them, and clipboard sync can copy them elsewhere. Apple's Universal Clipboard requires both devices to be signed in to the same Apple Account 51, and content stays on the clipboard for a short time 52. You also decide which apps can use the microphone: on a Mac, Apple says you can choose which apps are allowed 53, and in Windows 11 apps can use the microphone only when Microphone access is turned on under Privacy & security 54.
Our guide Dictation App Permissions explains what each permission allows, which ones a dictation app genuinely needs and how to audit and revoke them. The questions are can a third-party keyboard see what you type, why a dictation app needs Accessibility access on a Mac, can apps read your clipboard and how to tell if an app is using your microphone. Step-by-step setup is in the Mac, Windows and phones and tablets guides.
A few seconds of audio can make a voice clone, so stored recordings carry risk
Voice cloning is the privacy risk that grows after recording. A 2023 Microsoft research paper reported that its VALL-E model can synthesize personalized speech from only a 3-second enrolled recording of an unseen speaker 55. That is a lab claim, and quality varies, but it shows how little audio a clone can need. Any public or stored recording of you is raw material.
Regulators are starting to treat synthetic voices as a fraud and consent problem. In February 2024 the FCC ruled unanimously that AI-generated voices are artificial voices under the Telephone Consumer Protection Act 56, so robocalls that use them need the same consent as other prerecorded calls. Voice ID at banks and call centers may be fooled by a good clone, which is why voice alone is no longer a safe password.
What works is verification outside the call: a call-back to a known number, a family verification phrase, and not relying on voice authentication. Detection of fakes by ear or software is unreliable. Dictation on-device does not hand a vendor a recording of you in the first place, which cuts one source of exposure but not the public ones.
The guide Voice Cloning and Deepfakes sets out what researchers and regulators say. The questions cover how much audio it takes, voice authentication security, whether cloning is legal, how to protect yourself and whether stored recordings can clone you. The term is defined in voice cloning.
Whisper is safe when it runs on your machine, and an app named after it may not be
Whisper is two different things for privacy: an open model you can run entirely on your own computer, and a hosted OpenAI API that receives your audio. The code and model weights are released under the MIT License 57, a license that permits embedding in other apps. An app that bundles the model keeps audio local. An app that calls the hosted API sends audio to OpenAI under its API data terms, with the retention described above. MacWhisper, for one, says it processes sensitive content locally, without data leaving the Mac 58, and Handy says transcription happens on your own computer without sending information to the cloud 59. Treat any such claim as the vendor's statement and confirm it with the airplane-mode test.
Whisper also has a risk for records. In a 2024 study of one dataset, roughly 1 percent of Whisper transcriptions contained entire hallucinated phrases or sentences that were not in the audio 60. When a transcript is a medical, legal or financial record, review it against the audio.
The guide Is Whisper AI Safe? separates the model from the API and shows how to tell which one an app uses. The questions are can Whisper run offline, does OpenAI store Whisper audio, is Whisper HIPAA compliant and do Whisper apps send audio to OpenAI. Developers who need recognition without a cloud API can use the engines compared in offline speech recognition libraries, with questions on Python, Raspberry Pi, ESP32, Vosk vs Whisper and mobile apps. For a broader map of dictation tools see dictation software and apps and browsers.
A private dictation setup takes six steps
Dictating privately means closing the five places your words can go. This routine works on a recent Mac or iPhone, and the same ideas apply on other platforms.
- Pick on-device recognition. On a Mac, open System Settings > Keyboard and check that Dictation is set to on-device processing for your language. The Keyboard settings show whether your audio is processed on the device 2.
- Run the airplane-mode test. Turn on airplane mode, dictate a sentence and confirm text appears. If it does not, the tool needs the network.
- Turn off sharing. Go to System Settings > Privacy & Security > Analytics & Improvements and turn off Improve Siri & Dictation 16.
- Read the policy for four things. Storage, training, human review and sharing. The question what a dictation app's privacy policy should tell you lists them.
- Audit permissions. Keep microphone access to the apps that need it, grant full access or Accessibility only to tools you trust, and turn off clipboard sync for sensitive work.
- Match the tool to the content. For patient, client or classified material, use an on-device tool or a cloud tool under a written agreement.
Yap, which we make, is an on-device dictation app for Mac and Windows, and it is one option among those described in our guide to choosing a dictation app. If the built-in dictation in your system is enough, use it: Apple Dictation on Mac and Windows voice typing are free. See /offline-dictation-mac for offline setup, and /online-dictation for the free browser tool, which is cloud-based, so keep sensitive content out of it.
All the guides in this topic
| Guide | What it answers |
|---|---|
| Is Voice to Text Safe? | The five places your words can go and a six-step private setup |
| Open-source dictation software | Offline recognition you can run yourself |
| Voice Data Retention | What Apple, Google, Microsoft and Amazon keep |
| Is Whisper AI Safe? | Local model versus hosted API |
| HIPAA-Compliant Dictation | BAAs, the Security Rule and on-device use |
| Is Your Voice Biometric Data? | GDPR, BIPA and CCPA |
| Offline Speech Recognition Libraries | Engines for developers |
| Do Dictation Apps Train AI on Your Voice? | Training policies and opt-outs |
| Do Companies Listen to Your Recordings? | The 2019 revelations and what changed |
| Recording Consent Laws | One-party and all-party states |
| Dictation App Permissions | Mic, full access, Accessibility, clipboard |
| Voice Cloning and Deepfakes | Risks and protection |
| Dictation for Lawyers | Privilege and confidentiality |
| Dictation at Work | Enterprise and government policy |
For the other sides of the subject, see our overviews of dictation, speech recognition, transcription, accessibility and dictation at work.
Evidence
Every number on this page, with where it comes from. Open a fact to see the exact words in its source.
- 1.
- 2.
- 3.
- 4.
- 5.
- 6.
- 7.
- 8.
- 9.
- 10.
- 11.
- 12.
- 13.
- 14.
- 15.
- 16.
- 17.
- 18.
- 19.
- 20.
- 21.
- 22.
- 23.
- 24.
- 25.
- 26.
- 27.
- 28.
- 29.
- 30.
- 31.
- 32.
- 33.
- 34.
- 35.
- 36.
- 37.
- 38.
- 39.
- 40.
- 41.
- 42.ABA Formal Opinion 512 says lawyers must make reasonable efforts to prevent inadvertent or unauthorized disclosure of, or unauthorized access to, information relating to a client's representation.
ABA Standing Committee on Ethics and Professional Responsibility, American Bar Association, 2024 · Original
- 43.
- 44.
- 45.
- 46.
- 47.
- 48.
- 49.
- 50.
- 51.
- 52.
- 53.
- 54.
- 55.
- 56.
- 57.
- 58.
- 59.
- 60.
Questions people ask
- Are conversations with an AI chatbot privileged?Probably not, at least for a consumer chatbot a non-lawyer uses on their own. In February 2026 a federal court in New York held that a defendant's conversations with Claude were not protected by attorney-client privilege or work product. The court did not decide how a tool used at a lawyer's direction would be treated.
- Are open-source offline dictation apps like Handy and Murmur safe to use?Generally yes. Handy and Murmure, the app people often search for as Murmur, are designed to transcribe on your own computer, and their code is public. Their safety in practice depends on three more things: downloading only from the official release page, checking that your operating system can verify the file, and granting only the permissions the app needs.
- Can a third-party keyboard see what you type?A third-party keyboard sees the characters you type into ordinary fields, because it produces them. On an iPhone it cannot send that text over the network unless you turn on Allow Full Access for it, and the system keyboard handles password fields. With Full Access on, Apple warns that the developer may collect and transmit what you type, so read that keyboard's privacy policy before you turn the switch on.
- Can AI models leak what you dictated?Yes, in research. Language models can repeat rare pieces of their training text word for word, and speech models can too. The studies used research models and attacks, so they do not show how often a dictation app leaks what it heard. Training on your audio is the condition that creates the exposure.
- Can apps read your clipboard?Yes. Whether an app can read what you copied depends on the system: iPhone asks for your intent before one app reads what another app copied (from iOS 16), and Android 12 and later show a notice each time an app pastes clipboard data. Clipboard sync and Universal Clipboard can also copy dictated text to your other devices, so turn them off where you do not need them.
- Can government employees use AI dictation?Yes, if the agency has approved the tool through its own process. For US federal agencies, cloud services that fall within FedRAMP's scope must hold a FedRAMP authorization, and agencies write their own rules for generative AI. Some governments ban named AI tools on government devices, and state and local rules differ.
- Can lawyers use AI notetakers on client calls?Only with the client's consent, and in some places the law requires it before anyone records. New York's City Bar says clients must be told and must consent before an AI system records a call with them. California and Florida require the consent of all parties before a confidential conversation is recorded, so a notetaker can create liability even when the ethics rules are met.
- Can speech to text run on an air-gapped computer?Yes, if the speech engine and its language files are installed on the computer before it is cut off from networks. Cloud services cannot run there at all, and some built-in tools need an internet connection for their first-time setup.
- Can Whisper AI run offline?Yes. The open-source Whisper model runs offline once its model file is on your computer, so the audio never has to leave the machine. You need a connection to install it and to download the model the first time, while OpenAI's hosted API cannot run offline at all.
- Can you use speech to text for confidential documents?Yes, for many documents, but only with a tool that processes speech on the device or a cloud service your organization has approved for that content. Dictation that stays on the computer avoids the main risk, because the audio never reaches a server. Cloud tools also need a contract, retention controls and a check of their default settings, and privileged or regulated records need more than that.
- Can your employer see what you dictate?Sometimes. The answer turns on who owns the device, which dictation tool you use and whose account it runs on. Apple says device management cannot access personal information on a managed device, but work-account AI tools keep audit records of who used them, when and where, and sometimes the text of the prompt.
- Can your stored voice recordings be used to clone your voice?Yes, in principle. A stored recording is the same raw material as any other clip of your voice, and a clone can be made from a few seconds of clear speech. The real risk depends on who can reach the copy: company staff or vendors, a breach, or someone who gets into your account. Check whether the service keeps your audio, for how long, and whether people review it.
- Did Apple contractors listen to Siri recordings?Yes. In July 2019 the Guardian reported that Apple contractors regularly heard confidential details in Siri recordings, and Apple said it suspended human grading of Siri requests. Apple later agreed to pay $95 million to settle a privacy class action, and human review now requires you to opt in.
- Do AI dictation apps send your words to a language model?Some AI dictation apps do and some do not. A speech recognizer turns audio into words, and a separate language model may then edit or rewrite them, which sends your text to that model's provider. Check whether the app runs a model step, which provider it names, and whether a plain speech-to-text mode is available.
- Do cloud speech-to-text APIs store your audio?It depends on the service and the mode you use, not on the brand. Azure's real-time speech to text keeps no audio at rest when endpoint logging is off, and Google Cloud Speech-to-Text does not log audio or transcripts unless you opt in to data logging. Azure batch jobs keep audio and transcripts in storage you choose, for a retention time you set, and Amazon Transcribe has job data and artifacts that you can delete with its delete APIs.
- Do patients have to consent to AI scribes?It depends on the state, not on HIPAA. HIPAA lets a clinic use patient information for treatment, including documentation, but recording the conversation is a separate act, and about 11 states primarily require every party to agree. Ask every patient before an AI scribe records the visit, whatever your state, and write down the answer.
- Do Whisper-based apps send your audio to OpenAI?Not always. Some apps named after Whisper send your audio to OpenAI through an API key you enter, others run the model on the device, and a few send it to a server you host. An app's name does not tell you which path it takes, so check its settings and documentation, then run an airplane-mode test on a real recording.
- Do you need a BAA for dictation software?Yes, if the dictation software creates, receives, maintains or transmits patient information for you, which is the case when audio or transcripts go to a vendor's servers. No vendor BAA is needed for a tool that never receives patient information, such as software that recognizes speech entirely on your device, but the device still has to meet the HIPAA Security Rule.
- Do you need an OpenAI API key to use Whisper offline?No. The open-source Whisper model runs offline on your own computer without an OpenAI API key, and you need one only for OpenAI's hosted transcription API, which sends audio to OpenAI and cannot work offline. An app that asks for your key may use it for cloud transcription, so check its settings before you assume it works offline.
- Does Alexa store your voice recordings?Yes, unless you turn saving off. Since March 2025, Amazon has told some Echo owners that their voice recordings are sent to the cloud for processing, and they are kept unless you choose not to save them. If you choose not to save them, they are deleted after Alexa processes each request. Amazon's announcement does not say which option is the default, so check your setting.
- Does Azure speech to text store your data?It depends on the feature. Real-time speech to text stores no data at rest when endpoint logging is off and no customizations are used, and Microsoft says it does not use customer data to improve its Speech models. Batch transcription stores audio and output in storage you choose, and custom speech stores uploaded data in Microsoft's storage by default, so check which feature you use.
- Does Google keep your voice recordings?Not by default. Google keeps audio from voice interactions in your Google Account only if the voice and audio activity setting is on, and that setting is off unless you turn it on. Google still keeps a text history of Gemini conversations, by default for 18 months, and chats that people reviewed for up to three years.
- Does Illinois BIPA cover voiceprints?Yes. Illinois's Biometric Information Privacy Act names voiceprints as biometric identifiers, so a business that collects one must first get written consent. The 2024 amendment, SB 2979, limits damages for repeated collection of the same identifier from the same person by the same method to one violation.
- Does Microsoft keep your voice recordings?It depends on the feature. Windows voice typing sends your speech to Microsoft to make the transcript, and Microsoft says it does not store, sample or listen to recordings without your permission. Voice clips are optional, and Microsoft says it keeps the ones you contribute for up to two years. Copilot keeps conversations for 18 months by default and Copilot Voice audio for 48 hours on work or school accounts.
- Does OpenAI store audio sent to the Whisper API?By default, OpenAI does not use data sent to its API to train its models, unless you opt in. By default, abuse monitoring logs are kept for up to 30 days, and the page does not say whether they include audio. OpenAI's endpoint table lists no abuse monitoring retention for its transcription endpoint, but its general statement says otherwise, and the page does not say which applies to your account.
- Does Otter.ai use your recordings to train its AI?Yes, according to its privacy policy, effective June 16, 2026. Otter lists training its proprietary AI on de-identified audio recordings and transcripts as a purpose. We found no opt-out from that training in the policy, and the opt-outs it names cover the sale of personal information and targeted advertising.
- Does speech to text collect biometric data?Usually not. Speech to text turns your words into text, and the audio and transcript are personal data. It collects biometric data only when the same system also builds a model of your voice that identifies or verifies you, which is a voiceprint.
- Does speech to text share your data with third parties?Sometimes. Dictation sends your audio to a third party only when the recognition runs on a server or the app passes it to another company, such as a cloud provider or an AI model maker. Recognition that runs on your device sends the audio nowhere, and the app's own policy still decides what else it shares.
- Does using cloud dictation waive attorney-client privilege?We found no court ruling that sending dictation audio to a cloud vendor waives privilege on its own. Confidentiality is what keeps privilege intact, so the risk grows when the vendor keeps recordings, trains on them or lets people listen. Check the vendor's terms before client audio goes there.
- Does voice typing save your audio recordings?Voice typing saves your audio only when the company that recognizes your speech keeps it, and that depends on the product and your settings. Apple does not store dictation sent to its servers unless you opt in, while ChatGPT keeps dictated audio for as long as the chat stays in your history, so check the setting or policy for the exact tool you use.
- How can you tell if an app is using your microphone?Look for the indicator your system shows, then check the app list in its privacy settings. On an iPhone, an orange indicator in the status bar means an app is using the microphone. On a Mac, the Recording Indicator in Control Center shows when the microphone is in use or has been used recently. On Windows 11, apps can use the microphone only when Microphone access is on, so check those switches.
- How do I delete my voice recordings from Google, Apple, Amazon and Microsoft?Delete each company's voice history in the place it keeps it: Siri and Dictation history on a Mac, voice data on Microsoft's privacy dashboard, and your Alexa privacy settings for Amazon. Deleting history does not turn off future saving, so check the setting that controls that too, and check Google's own help page because we could not verify its steps.
- How do you protect yourself from voice cloning scams?Do not decide based on the voice. Hang up, call the person back on a number you already have, check with someone else who knows them, and agree a safe phrase in advance with the people who might ask you for money. These steps work because they check the route to the person, which a scammer controls only on the call, rather than the voice, which a clone can copy.
- How do you stop apps from training AI on your voice?Find the setting that names training and switch it off, then check what it does not cover. On a Mac, turn off Improve Siri & Dictation. In Copilot, turn off Training on voice conversations. In Wispr Flow, turn off Improve the model for everyone. If an app has no such setting, keep your audio on your device, where it is not sent to a server.
- How does a voice recognition security system work?A voice recognition security system records a sample of your voice to build a voiceprint, a stored model of how your voice sounds. On later calls it compares the new voice with the voiceprint for the identity the caller claims, then accepts or rejects that claim. Because the check only tests whether the voice matches, a good voice clone can pass it, so treat it as one check among several and never as the only approval for a payment.
- How much audio does it take to clone a voice?As little as 3 seconds of speech can be enough to start a voice clone, according to research and a bank's warning, and a vendor recommends 1 to 2 minutes of clear audio for its instant voice cloning. The vendor says audio quality matters most, so a short, clean clip from any public source can be enough to start a clone.
- How often do voice assistants activate by accident?In a 2020 study that played TV shows to smart speakers, a Google Home Mini averaged 0.95 activations per hour without the wake word while playing one show, and the Apple HomePod averaged 0.38 across all the shows. Real homes will differ, so there is no single daily count, and the rate depends on the device and on what it hears.
- Is a voice recording personal data under GDPR?Yes. A voice recording relates to a person who can be identified, so it is personal data under GDPR Article 4(1). It becomes special category data under Article 9 only when it is processed to uniquely identify that person, for example to build a voiceprint.
- Is Google voice typing HIPAA compliant?Not on the evidence we found. Google Docs is in the Google Workspace list covered by Google's BAA, but Docs voice typing sends speech through the browser's speech-to-text service, and we found no Google statement that this step is covered. Until Google confirms it in writing, do not dictate patient information with it.
- Is it legal to clone someone's voice?It depends on what the clone is used for, and the laws we checked target misuse rather than cloning itself. Using a cloned voice to deceive someone for money can be a federal crime under wire fraud law, AI-voice robocalls fall under US call rules, and Tennessee protects a voice against unauthorized use. Get consent before you clone someone else's voice for anything beyond a private test, and check the rules where you live.
- Is it legal to record a phone call in the UK or EU?Recording your own personal call is generally treated differently from a business recording, but there is no single yes or no. In the UK and EU, the criminal law on interception and the data protection law both matter, and a recording that identifies people is personal data. Announce the recording, have a clear reason for it, and take local legal advice before you record a call that matters.
- Is it legal to use an AI notetaker without telling everyone?Often not. An AI notetaker records every participant, so in an all-party state, or with a participant who is entitled to be told, you need every person's agreement before it starts. Even where one-party consent applies, tell everyone, because a notetaker can keep recordings and use audio in ways the host did not expect.
- Is it safe to dictate passwords and sensitive information?Not passwords or PINs. Anyone nearby can hear them, the transcript can be stored, and some dictation features are switched off in password fields for that reason. Other sensitive information can be dictated only with on-device recognition, no kept transcript, no bystanders and an agreement where a regulation requires one.
- Is Microsoft dictation HIPAA compliant?No HHS-approved HIPAA certification exists for a business associate, so the answer depends on the feature. We found no Microsoft statement that Windows voice typing is covered by a Business Associate Agreement, so treat it as not covered. Dictate in Microsoft 365 may fall under Microsoft's BAA, but get that confirmed in writing before you dictate patient information.
- Is my phone listening to me for ads?No reliable evidence shows that a phone or smart speaker records everyday conversation and sells what it hears to advertisers. The documented cases are a company that falsely claimed it could do this, Meta's disclosed use of AI chats to personalize ads, and accidental wake-ups that record speech by mistake.
- Is offline speech recognition more private than cloud speech recognition?Yes, for the audio. Offline speech recognition turns your voice into text on your device, so the recording never crosses the network to a server, which removes the largest privacy risk. It does not protect the text after recognition, and some systems need a download first or fall back to a server.
- Is on-device dictation HIPAA compliant?Yes for the recognition step, with conditions. When speech recognition runs on the device, no vendor receives the audio, so the BAA question largely falls away for that step. On-device dictation does not make a practice compliant by itself: the device must still meet the HIPAA Security Rule, and you must confirm that no transcript or cloud step leaves the computer.
- Is the Whisper AI model safe to download and run on your own computer?Yes, if you get it from the official project and run it on a computer you control. OpenAI released Whisper's code and model weights under the MIT License, and a local copy does not send your recordings to OpenAI. The risks sit around the model: the file you load, the software that loads it, and any app built on it that uses the cloud.
- Is voice authentication still secure?No, not on its own. Voice authentication checks whether a voice matches a stored voiceprint, and researchers have bypassed such systems with up to a 99% success rate after six tries, so a voice should never be the only check that approves an account change or a payment. It can still be one factor among several, provided a second check uses a channel a caller cannot control.
- Is voice data sensitive personal information under the CCPA?Only when it is used to identify a consumer. The CCPA lists voice recordings as biometric information, but biometric information counts as sensitive personal information only when it is processed for the purpose of uniquely identifying a consumer. Plain dictation usually falls outside that category.
- Is voice to text GDPR compliant?No voice to text tool is compliant or non-compliant in general, because the GDPR regulates what the organization that processes your voice does with it. Your use can be compliant only if that organization has a lawful basis, a binding contract with any cloud provider, a set retention period, and, when it identifies speakers at scale, an impact assessment. Check those points before you dictate personal or client information.
- Is Whisper AI HIPAA compliant?No, Whisper is not HIPAA compliant by itself, because compliance describes how a setup handles protected health information, not the name of the model. Local use can fit HIPAA with a documented risk analysis and no outside party handling the audio. The hosted OpenAI API can process protected health information only under a signed Business Associate and Healthcare Addendum and a provisioned organization.
- What does ABA Formal Opinion 512 say about AI tools?ABA Formal Opinion 512 asks lawyers to understand the AI tools they use, to make reasonable efforts to protect client information, and to get a client's informed consent before entering client information into self-learning tools that could disclose it.
- What does Google's speech to text privacy policy say about your voice?Google's general privacy policy says that, depending on your settings, it can save audio recordings of your voice interactions with Search, Assistant, Maps and Gboard. What is kept in practice depends on the product: Gboard keeps donated speech snippets for no more than 18 months, and Cloud Speech-to-Text does not log audio or transcripts unless you opt in to data logging.
- What is voice biometrics and how does a voice biometric system work?Voice biometrics is the use of how a voice sounds to recognize or verify who is speaking. The system first enrolls a voice to build a stored model called a voiceprint, then compares new speech with it, either to accept or reject a claimed identity or to search enrolled voices for the closest match.
- What should a company dictation policy include?A workable policy names approved tools and the account each must use, bans personal accounts for work speech, and sets required settings for contracts, training and logging. It also states how long audio and transcripts are kept, and names an owner with a review date.
- What should a dictation app's privacy policy tell you?A dictation app's privacy policy should answer four questions in writing: how long audio and text are kept, whether they train AI models, who can review them, and which vendors or AI models receive them. Start with the storage and retention section, then check each answer against the product's settings before you dictate anything sensitive.
- What was the Google Assistant recording lawsuit about?It was a class action that said Google Assistant recorded private conversations after waking by mistake, and that Google used those recordings for targeted advertising. Google denied wrongdoing and agreed in January 2026 to pay $68 million to settle.
- Which dictation apps and software are HIPAA compliant?No dictation app is HIPAA compliant on its own. Compliance depends on whether the vendor signs a business associate agreement (BAA) that covers the feature you use, and on the safeguards your organization keeps around it. Dragon Medical One's terms name a BAA, Wispr Flow and Superwhisper offer one on some plans or setups, and Superwhisper's local-only mode needs no vendor BAA.
- Which is more private: Apple dictation or Google voice typing?For dictation on a Mac, Apple's general dictation is the clearer case: Apple says it is processed on your device when Keyboard settings show that, and dictation sent to its servers is not stored unless you opt in. For voice typing on a phone, the answer depends on the app: Gboard's microphone sends speech to Google on an iPhone, while Gboard's offline recognizer works without a connection on Android.
- Which law applies when you record a call across state lines?No single rule settles it. When the two ends of a call are in different states, the stricter state's consent rule can apply, and a California court applied California law to calls with California clients. Meet the stricter standard whenever any participant is in an all-party state.
- Which states require all-party consent to record?California, Florida, Maryland and Washington clearly require every participant to consent before a private conversation or call is recorded. Other states on published lists could not be checked against their statute text here, so confirm each one before you rely on it.
- Why do companies ban AI notetakers and dictation apps?Companies ban a tool when the vendor, not the company, controls what happens to what people say, or when recording breaks consent rules. Personal accounts add risk because they sit outside company logging and retention controls. The public bans we found target chatbots and a meeting notetaker, not dictation apps by themselves.
- Why does a dictation app need Accessibility access on a Mac?Accessibility is the permission macOS uses to let one app access and control other apps through accessibility features. A dictation app that types into the window you are using needs a route like this, so macOS blocks it until you grant it in Privacy & Security settings. The grant reaches more than your voice, so allow it only for a dictation app you know and trust.
Terms in this topic
- All-party consentAll-party consent is a recording rule under which every participant in a conversation must agree before it can be recorded or intercepted. California, Florida and Maryland use it in their statutes, and it is stricter than one-party consent.
- Attorney-client privilegeAttorney-client privilege protects confidential communications between a lawyer and their client about seeking legal advice or services. A third party who is present, or a tool that handles the conversation, can put that protection at risk.
- Biometric dataBiometric data is personal data produced by technical processing of a person's physical, physiological or behavioral characteristics that allows or confirms who that person is. A voice recording becomes biometric data when it is processed to identify the speaker.
- Business associate agreement (BAA)A business associate agreement (BAA) is the written contract that HIPAA requires before a covered entity lets a vendor create, receive, maintain or transmit protected health information on its behalf. Dictation tools used in clinics often need one, and a BAA alone does not make a product compliant.
- False accept (accidental activation)A false accept, also called a false wake or accidental activation, is when a voice assistant or dictation feature starts listening or recording although the wake word was not said. Smart speakers have been measured to wake by accident, and that can capture speech the user did not mean to share.
- Federated learningFederated learning is a way to train one machine learning model on data spread across many devices. Each device computes an update from its own data, and only those updates are sent to a server, so the raw data does not have to be collected in one place.
- On-device speech recognitionOn-device speech recognition converts speech to text with a model that runs on your own computer or phone, so the audio does not need to be sent to a server. It is also called local, offline or on-device dictation.
- One-party consentOne-party consent is a recording rule under which a conversation can be recorded lawfully when one of its participants agrees, which is usually the person doing the recording. The US federal wiretap law uses this rule, and individual states may set stricter ones.
- Protected health information (PHI)Protected health information (PHI) is individually identifiable health information that HIPAA protects when it is held or sent in electronic or other form by a covered entity or its business associate. Patient names with diagnoses, in dictated notes or voice recordings, are PHI. It is also called ePHI when held electronically.
- Speaker verificationSpeaker verification is a voice biometric check in which a system accepts or rejects a claimed identity by comparing a voice with stored models of the person claimed. It is also called voice authentication, voice ID or voice biometrics.
- Special category data (GDPR)Special category data is the group of personal data that Article 9 of the GDPR bans organizations from processing unless a listed exception applies. It includes health data and biometric data used to identify a person, and a voice recording can fall into it when it is processed that way.
- Voice cloningVoice cloning is the use of a model trained on recordings of one person to generate new speech in that person's voice. Some products need only minutes of clear audio, and cloned voices have been used in fraud, so recordings of your voice carry a real risk.
- VoiceprintA voiceprint is a stored model of a person's voice, built from measurable features of their speech and used to recognize or verify that person. Illinois law lists voiceprints among biometric identifiers, and other vendors call the same idea a voice template or speaker model.
- Zero data retention (ZDR)Zero data retention (ZDR) is a setting or contract term under which a service keeps none of your data once it has handled your request. Each vendor sets its own scope, so check what the label excludes before you rely on it.
Referenced by
- Dictation and Voice Typing: How to Talk Instead of TypeTopic
- Dictation Software: How to Choose the Right AppTopic
- Mac Dictation: How to Dictate and Control a Mac by VoiceTopic
- Dictation on Windows: Voice Typing, Voice Access, AppsTopic
- Voice Typing in Google Docs, Word and Every Other AppTopic
- Dictation and Voice to Text on iPhone and AndroidTopic
- Speech Recognition: What It Is and How Well It WorksTopic
- Transcription: How to Turn Audio and Video Into TextTopic
- Speech to Text AI: Dictation, Assistants and Voice AgentsTopic
- Is Voice to Text Safe? How to Dictate PrivatelyGuide
- Best Dictation Software and Apps, Compared by NeedGuide
- Writing a Book by Dictation: Software and WorkflowGuide
- Free Dictation Software: The Best Free OptionsGuide
- Voice Data Retention: What Apple, Google and Microsoft KeepGuide
- Is Whisper AI Safe? Privacy and Security ExplainedGuide
- Is Dictation HIPAA Compliant? What Makes It SafeGuide
- Is Your Voice Biometric Data? GDPR, BIPA and CCPAGuide
- Do Dictation Apps Train AI on Your Voice?Guide
- Do Companies Listen to Your Voice Recordings?Guide
- Recording Consent Laws: One-Party vs All-Party StatesGuide
- Dictation App Permissions: Mic, Full Access, ClipboardGuide
- Dictation for Academics, Researchers and ThesesGuide
- Voice Cloning and Deepfakes: Risks and ProtectionGuide
- Dictation for Lawyers: Privilege and ConfidentialityGuide