Facts

Voice privacy and security: facts and numbers

286 facts, each traced to the exact words of its source.

A 2018 study shows that a language model can be trained on client devices without exporting sensitive user data to servers, and that the approach matched or improved prediction recall.Andrew Hard et al., arXiv, 2018
3 secondsA 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
A 2024 study found that when a speech model has seen a recording during training, a noisy version of that recording can make the model transcribe its memorized sensitive transcript.Matthew Jagielski et al., arXiv, 2024
25 million US dollarsA deepfake scam led an employee of the engineering firm Arup in Hong Kong to pay out $25 million to fraudsters.Kathleen Magramo, CNN (reprinted by ABC17 News), 2024
A Windows policy named AllowInputPersonalization controls whether users can enable online speech recognition, and when the policy is disabled, users cannot turn speech services on in Settings.Microsoft Learn, undated
ABA Formal Opinion 477R says what counts as reasonable efforts to protect client information is not a fixed rule, but depends on a set of factors.ABA Standing Committee on Ethics and Professional Responsibility, American Bar Association, 2017
ABA Formal Opinion 512 concludes that informed consent is required before client information is entered into self-learning generative AI tools whose outputs could disclose it.Harvard Journal of Law and Technology (JOLT Digest), 2026
ABA Formal Opinion 512 says a general clause in an engagement letter is not enough to obtain the informed consent needed before using generative AI with client information.ABA Standing Committee on Ethics and Professional Responsibility, American Bar Association, 2024
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
ABA Formal Opinion 512 says lawyers need not become generative AI experts, but must have a reasonable understanding of the capabilities and limitations of the specific generative AI technology they might use.Robert Ambrogi, LawNext, 2024
ABA Formal Opinion 512 says that self-learning AI tools, by their nature, raise the risk that one client's information is disclosed to others.ABA Standing Committee on Ethics and Professional Responsibility, American Bar Association, 2024
0.38 activations per hourAcross all the TV shows tested in the 2020 study, the Apple HomePod averaged 0.38 activations per hour without the wake word. That was just behind the Invoke with Cortana and the second-generation Echo Dot, which averaged 0.40.Daniel J. Dubois et al., Proceedings on Privacy Enhancing Technologies (PETS 2020), 2020
After reports of people listening to Siri recordings, Apple said it immediately suspended human grading of Siri requests and began a review of its practices.Apple, Apple Newsroom, 2019
2663 US dollars per violationAfter the California Privacy Protection Agency's inflation adjustment effective January 1, 2025, the CCPA administrative fine under Civil Code 1798.155(a) is up to $2,663 for each violation and up to $7,988 for each intentional violation or violation involving consumers under 16.California Privacy Protection Agency, 2024
Amazon Connect Voice ID does not ask an enrolling caller to repeat a particular word or phrase.Amazon Web Services, Amazon Connect Customer Administrator Guide, 2026
Amazon Connect Voice ID generates a voiceprint from the caller's audio, compares it with the enrolled voiceprint for the identity the caller claims, and returns an authentication result.Amazon Web Services, Amazon Connect Customer Administrator Guide, 2026
30 seconds of speechAmazon Connect Voice ID needs 30 seconds of customer speech, excluding silence, before it can create a voiceprint and enroll the customer.Amazon Web Services, Amazon Connect Customer Administrator Guide, 2026
Amazon says an organization can opt out of having its Transcribe content used to improve Amazon Transcribe and its other machine-learning services, by using an AWS Organizations opt-out policy.Amazon Web Services, n.d.
Amazon says customers can use its Delete APIs to delete the data and other artifacts associated with transcription jobs.Amazon Web Services, n.d.
Amazon says that if you choose not to save your Alexa voice recordings, they are deleted automatically after Alexa processes each request.Amazon, n.d.
30 daysAmazon says transcripts of your Alexa requests remain available for 30 days before it begins to delete them automatically.Amazon, n.d.
3 or 18 monthsAmazon says you can keep saving Alexa voice recordings and have them automatically deleted on an ongoing three-month or 18-month basis.Amazon, n.d.
Amazon told Echo owners that from March 28 their voice recordings would be sent to and processed in the cloud and deleted after Alexa processes the requests.The Register, The Register, 2025
Apple advises users to be cautious and grant access to a Mac only to apps they know and trust.Apple Support, Mac User Guide, 2026
730 daysApple may keep Siri and Dictation request history for up to about 730 days.Calculated from other facts
Apple said in 2022 that from iOS 16 the system confirms the user's intent before an app reads pasteboard items written by another app.Apple Developer, 2022
Apple said that, by default, it would no longer retain audio recordings of Siri interactions, and that users could opt in to let it learn from audio samples.Apple, Apple Newsroom, 2019
Apple says an app that has been notarized has been checked for malicious software, and none was detected.Apple, Apple Support, 2026
Apple says data is listed as linked to you when it is collected in a way tied to your identity, such as your account, your device or your phone number.Apple Inc., n.d.
Apple says Siri transcripts and related request data are not used to build a marketing profile, and are never sold to anyone.Apple, Apple, 2026
Apple says supervision lets IT use controls that are unavailable for other deployment models, such as personally owned devices.Apple, 2022
Apple says that if you allow apps to use Speech Recognition for transcription, the audio to be transcribed may be sent to Apple.Apple, Apple, 2026
Apple says that on-device machine learning benefits privacy because the data for inference workloads stays on the device rather than on a server.Apple Machine Learning Research, Apple Machine Learning Research, 2022
Apple says that when a user taps a secure text input object, such as a password field, the system temporarily replaces a custom keyboard with the system keyboard.Apple, Apple Developer Documentation (archived), n.d.
Apple says that when you give an app access to your Mac, you also give it access to your contacts, calendar and other information, and that the app is then subject to its own terms and privacy policy rather than Apple's.Apple Support, Mac User Guide, 2026
Apple says the MDM framework can never access personal information on a managed device, including email, messages and browser history.Apple, 2022
less than 1 percent of daily Siri activationsApple told the Guardian in 2019 that less than 1% of daily Siri activations were used for grading, and that the clips were typically only a few seconds long.Alex Hern, The Guardian, 2019
27 million US dollarsApple's $95 million Siri settlement was $27 million larger than Google's $68 million Assistant settlement. Both are settlements of class actions over voice assistant recordings, and neither company admitted wrongdoing.Calculated from other facts
Apple's Siri, Dictation and Privacy page says that by using Siri or Dictation you consent to Apple processing and storing the information with trusted third-party service providers.Apple, Apple, 2026
Article 35(3)(b) of the GDPR lists processing on a large scale of special categories of data, which include biometric data used to identify a person, as a case that requires a data protection impact assessment.gdpr-info.eu (text of Regulation (EU) 2016/679), 2016
2 August 2026Article 50 of the EU AI Act comes into force on 2 August 2026.European Union, Regulation (EU) 2024/1689 (EU Artificial Intelligence Act), text mirror, 2024
Article 9 of the GDPR prohibits processing special categories of personal data, including biometric data used for the purpose of uniquely identifying a natural person, unless one of the listed exceptions applies.gdpr-info.eu (Intersoft Consulting), 2016
Arup confirmed that fake voices and images were used in the deepfake scam.Kathleen Magramo, CNN (reprinted by ABC17 News), 2024
AWS ended support for Amazon Connect Customer Voice ID, effective May 20, 2026, after it had stopped accepting new customers in May 2025.Amazon Web Services, Amazon Connect Customer Administrator Guide, 2026
27 percentBecause listeners spotted 73% of the speech deepfakes in the study, about 27% went unnoticed.Calculated from other facts
BIPA gives a private right of action to any person aggrieved by a violation of the Act.Hogan Lovells, Hogan Lovells, 2019
20 million euros, or 4 percent of worldwide turnoverBreaches of the GDPR's core processing rules can be fined up to 20 million euros, or up to 4 percent of total worldwide annual turnover for an undertaking.gdpr-info.eu (Intersoft Consulting), 2016
10 seconds of speechBy default, Amazon Connect Voice ID needs 10 seconds of a caller's speech to authenticate them against an enrolled voiceprint.Amazon Web Services, Amazon Connect Customer Administrator Guide, 2026
California Penal Code 632 punishes a person who intentionally records a confidential communication without the consent of all parties.California Public Law, 2026
Canada's Criminal Code exempts a person who has the express or implied consent of the originator of a private communication, or of the intended recipient.Parliament of Canada, Justice Laws Website, Government of Canada, n.d.
8 vendors with critical flawsCitizen Lab found critical vulnerabilities in keyboard apps from eight of the nine vendors it tested in 2024.Citizen Lab, University of Toronto, 2024
Citizen Lab recommends that Android sandbox keyboard apps by default, as iOS does, so they cannot send network traffic until the user grants access.Citizen Lab, University of Toronto, 2024
Citizen Lab recommends that at-risk users switch from a cloud-based keyboard app to one that works entirely on-device.Citizen Lab, University of Toronto, 2024
Citizen Lab says most of the vulnerable keyboard apps it found in 2024 can be exploited by a passive network eavesdropper, who needs no action from the user.Citizen Lab, University of Toronto, 2024
Citizen Lab says third-party keyboards on iOS cannot transmit network data without the Full Access permission.Citizen Lab, University of Toronto, 2024
9 vendors testedCitizen Lab tested keyboard apps from nine vendors in its 2024 report, including Baidu, Huawei and Tencent.Citizen Lab, University of Toronto, 2024
CNBC relayed Bloomberg's report that some Samsung staff had uploaded sensitive code to ChatGPT, which led to the restriction.Arjun Kharpal, CNBC, 2023
CNBC reported in May 2023 that Samsung confirmed it was temporarily restricting generative AI use on company computers after misuse of the technology.Arjun Kharpal, CNBC, 2023
Connecticut allows a phone call to be recorded without all-party consent if the recording party gives a verbal notification at the start that is recorded and is part of the call.Connecticut General Assembly, Connecticut General Assembly, n.d.
Connecticut places its telephone-recording rule in a statute titled as a civil action for illegal recording of private telephone communications.Connecticut General Assembly, Connecticut General Assembly, n.d.
Consumer Reports found that a majority of the six voice cloning products it assessed lacked meaningful safeguards against fraud or misuse.Consumer Reports, Consumer Reports, 2025
4 of six productsConsumer Reports researchers easily created a voice clone from publicly available audio in four of the six voice cloning products they tested.Consumer Reports, Consumer Reports, 2025
Cornell Law School's legal information site says attorney-client privilege protects confidential communications between a lawyer and their client about seeking legal advice or services.Legal Information Institute, Cornell Law School (Wex), 2026
Cornell's legal encyclopedia says 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.Legal Information Institute, Cornell Law School (Wex), 2026
$880,000 US dollarsCox Media Group agreed to pay $880,000 to settle FTC charges that it falsely claimed to offer active listening ad targeting to small businesses.Wendy Davis, MediaPost, 2026
24.9 percentCyberhaven reports that 24.9 percent of Gemini usage happens through personal accounts.Cyberhaven, undated
32.3 percentCyberhaven reports that 32.3 percent of ChatGPT usage happens through personal accounts rather than corporate ones.Cyberhaven, undated
58.2 percentCyberhaven reports that 58.2 percent of Claude usage happens through personal accounts.Cyberhaven, undated
60.9 percentCyberhaven reports that 60.9 percent of Perplexity usage happens through personal accounts.Cyberhaven, undated
Cyberhaven says personal AI accounts bypass SSO enforcement, centralized logging, enterprise retention policies and controls on model training.Cyberhaven, undated
Delaware's interception statute allows a person to intercept a communication when the person is a party to it, or when one of the parties has given prior consent.Delaware General Assembly, Delaware Code Online, n.d.
89 percentEight of nine vendors tested in the 2024 Citizen Lab report had critical flaws, which is 89 percent.Calculated from other facts
1-2 minutesElevenLabs recommends about 1 to 2 minutes of clear audio, without reverb or background noise, for its instant voice cloning.ElevenLabs, ElevenLabs, n.d.
ElevenLabs says audio quality is the most important factor in instant voice cloning.ElevenLabs, ElevenLabs, n.d.
Florida Ethics Opinion 24-1 recommends obtaining the affected client's informed consent before using a third-party generative AI program that would expose confidential information.Professional Ethics Committee of The Florida Bar, The Florida Bar, 2024
Florida's wiretap statute makes it lawful to intercept an oral, wire or electronic communication when all of the parties have given prior consent to the interception.Florida Legislature, Florida Senate, 2025
For custom speech and custom voice, Microsoft says your uploaded data is stored in Microsoft's storage by default, unless you prepare your own storage account.Microsoft, Microsoft Learn, 2026
From Android 12, Google shows a toast each time an app reads and pastes clipboard data, which makes attacks harder to miss.Android Developers, n.d.
GDPR Recital 51 says photographs should not systematically be treated as special category data, because they count as biometric data only when processed through specific technical means that allow unique identification or authentication.gdpr-info.eu (Intersoft Consulting), 2016
1095 daysGoogle keeps Gemini chats that human reviewers read for up to about 1,095 days.Calculated from other facts
0.2 percent of audio snippetsGoogle said in July 2019 that language experts review only around 0.2 percent of all audio snippets.Google, Google Blog, 2019
Google said in July 2019 that one of its language reviewers violated its data security policies by leaking confidential Dutch audio data.Google, Google Blog, 2019
3 yearsGoogle says chats that human reviewers read are not deleted when you delete your Gemini activity, and are retained for up to three years.Google, n.d.
Google says Gemini apps wait in standby until activated, and that audio recordings may begin a few seconds before activation, which means a recording can start by accident.Google, n.d.
Google says Google Cloud and Google Workspace support HIPAA compliance within the scope of a business associate agreement, and that the customer remains responsible for evaluating its own HIPAA compliance.Google Cloud documentation, n.d.
Google says it may delete audio recordings from its servers sooner than the auto-delete period a user chooses, once the audio is no longer needed for improving its audio recognition.Google, n.d.
Google says its voice and audio activity setting does not control audio saved and managed by other Google services, such as Google Voice and YouTube.Google, n.d.
Google says its voice and audio activity setting is off unless you choose to turn it on, so audio from voice interactions is not saved to your Google Account by default.Google, n.d.
Google says that Android versions older than Android 10 let background apps read clipboard information from the foreground app.Android Developers, n.d.
Google says that by default your Gemini Live audio, video and screenshares are not used to improve Google services.Google, n.d.
Google says that donating audio snippets to Gboard is voluntary, and that you may be asked to share them when you use dictation.Google, Google Gboard Help, 2026
72 hoursGoogle says that even when Keep Activity is off, conversations are saved with your account for up to 72 hours, so Gemini can provide the service and process feedback.Google, n.d.
Google says that if you turn off its voice and audio activity setting, audio already saved to your account is not deleted.Google, n.d.
Google says that only a limited set of authorized Google employees and contractors can access customer data collected under Speech-to-Text data logging, and that Google uses it solely to train and improve its products.Google Cloud, 2026
18 monthsGoogle says the default auto-delete period for Gemini Apps activity is 18 months, and that users can change it to 3 or 36 months or turn auto-delete off.Google, n.d.
15 secondsGoogle says the speech snippets you donate to Gboard's conventional learning include a maximum of 15 seconds of your speech.Google, Google Gboard Help, 2026
July 10, 2019 dateGoogle told the Hamburg data protection authority that it had suspended human reviews of Google Assistant audio across all of Europe on July 10, 2019, after learning of a data leak.Natasha Lomas, TechCrunch, 2019
548 daysGoogle's default 18-month auto-delete period for Gemini Apps activity is about 548 days.Calculated from other facts
Google's HIPAA page lists Speech-to-Text among its Google Cloud services and tells customers who have a BAA covering PHI not to opt in to data logging for it.Google Cloud documentation, n.d.
Google's HIPAA page says the customer must determine whether it is a covered entity or a business associate, and whether it requires a business associate agreement with Google.Google Cloud documentation, n.d.
Google's privacy policy says that, depending on your settings, Google can save audio recordings of voice interactions with Search, Assistant, Maps and Gboard, to develop and improve its audio technologies.Google, Google, 2026
Handy's README says that on a restricted network, where Handy cannot download models automatically, you can download the model files yourself and install them.cjpais, Handy project, GitHub repository, 2026
HHS says most vendors and contractors that provide services to covered entities, and that involve access to protected health information, are business associates.U.S. Department of Health and Human Services, Office for Civil Rights, 2016
HHS says that lacking an encryption key does not exempt a cloud service provider from business associate status under HIPAA.U.S. Department of Health and Human Services, Office for Civil Rights, n.d.
HHS says the covered entity and a cloud service provider that holds its health information must sign a HIPAA-compliant business associate agreement.U.S. Department of Health and Human Services, Office for Civil Rights, n.d.
HIPAA permits a covered entity to use or disclose protected health information for its own treatment, which is the permission clinicians rely on when they document care.Legal Information Institute, Cornell Law School (reproducing the Code of Federal Regulations), n.d.
HIPAA's Privacy Rule lets a covered entity allow a business associate to create, receive, maintain or transmit protected health information on its behalf only if the entity obtains satisfactory assurance that it will be safeguarded.Legal Information Institute, Cornell Law School, 2026
$1,000 and $5,000 US dollars per violationIllinois BIPA carries statutory penalties of $1,000 for each negligent violation and $5,000 for each intentional or reckless violation.Mintz, Mintz, 2019
Illinois's Biometric Information Privacy Act defines biometric identifiers to include voiceprints, alongside fingerprints and iris scans.Hogan Lovells, Hogan Lovells, 2019
In 2006, the California Supreme Court found a true conflict between California and Georgia law over recordings of calls with California clients.Supreme Court of California, Supreme Court of California, 2006
In 2021 a federal judge found that Google's privacy policy did not sufficiently tell Assistant users it would use recordings made without manual activation or a hot word.NBC News, NBC News, 2021
In 2024 the US House of Representatives barred staff from using Microsoft Copilot on their government-issued devices.Priya Singh, Business Today, 2024
84 keyboard appsIn a 2015 peer-reviewed study, 84 of 139 keyboard apps on Google Play could be misused as keyloggers.Junsung Cho et al., 2015 13th Annual Conference on Privacy, Security and Trust (PST), IEEE, 2015
In a 2019 Interspeech paper on speaker verification, the decision threshold is the setting that controls the ratio between the false rejection rate and the false acceptance rate.Victoria Mingote et al., Proceedings of Interspeech 2019, ISCA Archive, 2019
In a 2020 study of smart speakers exposed to TV audio, the authors found no evidence that the devices constantly record conversations. They did find that the devices wake up frequently, usually for short intervals.Daniel J. Dubois et al., Proceedings on Privacy Enhancing Technologies (PETS 2020), 2020
In a 2021 study, an attack on the GPT-2 language model recovered hundreds of verbatim text sequences from the model's training data.Nicholas Carlini et al., USENIX Security Symposium 2021, 2021
73 percentIn a 2023 PLOS ONE study, listeners correctly spotted speech deepfakes only 73% of the time.Kimberly T. Mai and Lewis Griffin, PLOS ONE, 2023
10000 examplesIn a 2023 study, researchers recovered over ten thousand examples from ChatGPT's training dataset by prompting the model to diverge from normal output.Milad Nasr et al., arXiv, 2023
6 developersIn a 2025 Stanford study of the privacy policies of six US AI developers, all six use users' chat data to train their models by default.Jennifer King and Stanford Institute for Human-Centered Artificial Intelligence, Stanford Institute for Human-Centered Artificial Intelligence (HAI), 2025
28 percentIn a Starling Bank survey published in September 2024, 28% of UK adults said they think they have been targeted by an AI voice cloning scam in the past year.Starling Bank, Starling Bank, 2024
In Amazon Connect Voice ID, voice spoofing scores are not returned when only speaker authentication is enabled.Amazon Web Services, Amazon Connect Customer Administrator Guide, 2026
65 percentIn an internal survey conducted in April 2023, Samsung found that 65 percent of respondents believed generative AI services pose a security risk.The National (Bloomberg report), 2023
In August 2019, after questions about Skype Translator and Cortana audio, Microsoft said it could do better at specifying that humans sometimes review this content, and it updated its privacy statement and product FAQs.TechCrunch staff, TechCrunch, 2019
In August 2019, Amazon changed Alexa's settings so that users could opt out of having their voice recordings manually reviewed by its human workers. The change took effect on Friday, August 2, 2019, after Bloomberg reported the practice.Natasha Lomas, TechCrunch, 2019
In August 2019, Facebook said it had paused human review of audio from Messenger more than a week before its statement, and that it would follow Apple and Google in curtailing the practice.CBS News, CBS News, 2019
$68 million US dollarsIn January 2026, Google agreed to pay $68 million to settle a class action that accused it of illegally recording private conversations through Google Assistant. Google denied wrongdoing.Rohit Singh, MediaNama, 2026
In July 2025, OpenAI's chief executive Sam Altman told a Federal Reserve conference that some financial institutions still accept a voiceprint as authentication, and that this terrifies him.Associated Press, WSLS (Associated Press story), 2025
In Kearney, the California Supreme Court concluded that the comparative impairment analysis supported applying California law to the calls.Supreme Court of California, Supreme Court of California, 2006
In Kearney, the California Supreme Court said that if a business tells a caller at the start of a telephone call that it is being recorded, the recording would not violate California's statute.Supreme Court of California, Supreme Court of California, 2006
60 percentIn the 2015 study, 84 of 139 keyboard apps, or 60 percent, could be misused as keyloggers.Calculated from other facts
0.95 activations per hourIn the 2020 study, a Google Home Mini playing the TV show The West Wing averaged 0.95 activations per hour without the wake word, which the authors singled out as a notable case.Daniel J. Dubois et al., Proceedings on Privacy Enhancing Technologies (PETS 2020), 2020
0.40 activations per hourIn the 2020 study, the second-generation Amazon Echo Dot averaged 0.40 activations per hour across all the TV shows tested, tied with the Harman Kardon Invoke for the most activations of the devices tested.Daniel J. Dubois et al., Proceedings on Privacy Enhancing Technologies (PETS 2020), 2020
97 percentIn the 2023 study, 97 percent of the apps with a Data Not Collected privacy label had privacy policies indicating otherwise.Mir Masood Ali et al., arXiv, 2023
11 percentIn the ABA's 2023 survey, 11 percent of respondents were using AI-based technology.Bob Ambrogi, LawNext, 2025
30 percentIn the ABA's 2024 Legal Technology Survey Report, 30 percent of respondents said they were using AI-based technology.Bob Ambrogi, LawNext, 2025
In the same Stanford study, only some of the six AI developers say they remove personal information before using data for training.Jennifer King and Stanford Institute for Human-Centered Artificial Intelligence, Stanford Institute for Human-Centered Artificial Intelligence (HAI), 2025
In United States v. Heppner, the defendant used Claude, a publicly available AI assistant, to analyze his legal situation on his own, without direction from his attorneys.Orrick, Herrington & Sutcliffe LLP, Orrick, 2026
In United States v. Kovel, the Second Circuit said an accountant's presence does not destroy the privilege where it is necessary, or at least highly useful, to the consultation between client and lawyer.United States Court of Appeals for the Second Circuit, Federal Reporter, Second Series, 1961
Keyboard settings on a Mac show whether your general text dictation is processed on the device and not sent to Siri servers.Apple Support, n.d.
Maryland's exception that lets a party to a communication record it applies only where all of the parties have given prior consent.Maryland General Assembly, Maryland General Assembly, n.d.
McMahan and colleagues proposed federated learning, an approach that leaves the training data distributed on mobile devices and learns a shared model by aggregating locally computed updates.H. Brendan McMahan et al., Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS), JMLR Workshop and Conference Proceedings volume 54, 2017
Meta said in October 2025 that it would use people's interactions with its AI features to personalize the content and ads they see.Meta, Meta Newsroom, 2025
Meta says it does not use your microphone unless you have given permission and are actively using a feature that requires the microphone.Meta, Meta Newsroom, 2025
Microsoft documents that for real-time speech to text in Azure, audio is processed only in server memory and no data is stored at rest.Microsoft, Microsoft Learn, 2026
Microsoft says a customer can set a retention time for generated transcription text files with a parameter called timeToLive.Microsoft, Microsoft Learn, 2026
Microsoft says an updated Copilot app has been available since August 18, 2026, and that its privacy controls page describes only the older version of the app.Microsoft Support, 2026
Microsoft says audio of people speaking and its transcripts may be considered personal data or sensitive data under various privacy regulations and laws.Microsoft, Microsoft Learn, 2026
Microsoft says device-based speech recognition in Windows 11 processes the voice on the device and sends no voice data to Microsoft.Microsoft, Microsoft Support, n.d.
Microsoft says it does not use customer data to improve its Speech models.Microsoft, Microsoft Learn, 2026
Microsoft says Microsoft 365 Dictate does not store your audio or transcribed text, but your speech is sent to Microsoft to produce the text.Microsoft, 2026
Microsoft says that clearing your voice activity on its privacy dashboard removes your audio recordings but may not remove all information associated with your voice activity.Microsoft Support, 2026
Microsoft says that if endpoint logging is disabled and no customizations are used, no customer data is stored by the Speech service.Microsoft, Microsoft Learn, 2026
Microsoft says that in batch transcription the customer chooses where the audio and transcripts are stored, and controls their retention.Microsoft, Microsoft Learn, 2026
Microsoft says that Windows voice typing, which it calls Dictation in other products, converts speech to text using online speech recognition.Microsoft, Microsoft Support, n.d.
Microsoft says the system automatically creates an audit log when a user interacts with Copilot or another AI application, and that each record shows the user, the time and the place of the interaction.Microsoft Learn (Microsoft Purview documentation), 2026
Microsoft says voice data it collected and linked to your account before October 30, 2020 still shows on your dashboard for as long as Microsoft keeps it.Microsoft Support, 2026
Microsoft says you can pin clipboard items and sync your clipboard history to the cloud.Microsoft Support, n.d.
Microsoft's audit record for a Copilot interaction has a Messages field that contains details of the prompt and the response.Microsoft Learn (Microsoft Purview documentation), 2026
548 daysMicrosoft's default storage period for Copilot conversation activity, 18 months, is about 548 days.Calculated from other facts
730 daysMicrosoft's stated maximum retention for contributed voice clips, two years, is 730 days.Calculated from other facts
Murmure's privacy policy says that on startup the app checks for updates by contacting its updater endpoint, which is hosted on GitHub, and that other network features are opt-in.Kieirra, Murmure project, GitHub repository, 2026
Murmure's privacy policy says the app does not include analytics or telemetry.Kieirra, Murmure project, GitHub repository, 2026
Murmure's README roadmap lists signing the inner murmure.exe on Windows, not only the installers, as an open item marked under consideration.Kieirra, Murmure project, GitHub repository, 2026
New York City Bar Formal Opinion 2025-6 says clients must be told, and must consent, when an AI-driven system records a call with them.New York City Bar Association Committee on Professional and Judicial Ethics, New York City Bar Association, 2025
New York City Bar Formal Opinion 2026-2 says the default for AI recording of conversations with non-clients should be not to record, unless there is a good reason.New York City Bar Association Committee on Professional and Judicial Ethics, New York City Bar Association, 2026
NIST's glossary defines an air gap as an interface between two systems that are not physically connected and whose logical connection is not automated.NIST Computer Security Resource Center, 2022
Oklahoma's governor banned the DeepSeek app on all state-owned devices, citing security risks.Office of the Governor of Oklahoma, 2025
OMB memorandum M-24-15 says federal agencies must obtain and maintain a FedRAMP authorization when a cloud product or service falls within the scope of FedRAMP.Office of Management and Budget, Executive Office of the President, 2024
270 daysOMB memorandum M-25-21 says federal agencies should develop a generative AI acceptable-use policy within 270 days.Office of Management and Budget, Executive Office of the President, 2025
OMB says a FedRAMP authorization is not an endorsement of a product or service. It records that the security assessment is complete and presumed adequate.Office of Management and Budget, Executive Office of the President, 2024
On a Mac, the Recording Indicator light in Control Center shows when the microphone is in use or has been used recently.Apple Inc., Apple Support, macOS User Guide, n.d.
On an iPhone running iOS 14 or later, an orange indicator in the status bar means an app is using the microphone.Apple Support, 2026
2026-02-13 dateOn February 13, 2026, Judge Jed S. Rakoff of the Southern District of New York ruled that a non-attorney's communications with a generative AI platform were not protected by attorney-client privilege or work product.Orrick, Herrington & Sutcliffe LLP, Orrick, 2026
On iPhone, the microphone indicator appears as an orange square when the Differentiate Without Color setting is turned on.Apple Support, 2026
180 daysOnce an organization turns on pay-as-you-go audit logging, Microsoft keeps the logs of user interactions with non-Microsoft AI applications for 180 days.Microsoft Learn (Microsoft Purview documentation), 2026
48 percentOpen-source audio deepfake detectors lost 48% of their AUC, a measure of accuracy, on the 2024 in-the-wild benchmark compared with earlier benchmarks.Nuria Alina Chandra et al., arXiv, 2025
45 percentOpen-source image deepfake detectors lost 45% of their AUC on the 2024 in-the-wild benchmark compared with earlier benchmarks.Nuria Alina Chandra et al., arXiv, 2025
50 percentOpen-source video deepfake detectors lost 50% of their AUC on the 2024 in-the-wild benchmark compared with earlier benchmarks.Nuria Alina Chandra et al., arXiv, 2025
30 daysOpenAI says abuse monitoring logs are generated for all API feature usage and retained for up to 30 days by default, unless longer retention is required by law or to protect its services.OpenAI, OpenAI Platform documentation, 2026
OpenAI says its Zero Data Retention and Modified Abuse Monitoring controls currently require prior approval by OpenAI and acceptance of additional requirements.OpenAI, OpenAI Platform documentation, 2026
OpenAI says its Zero Data Retention control excludes customer content from abuse monitoring logs.OpenAI, OpenAI Platform documentation, 2026
OpenAI says that customers who have signed its Business Associate and Healthcare Addendum can use BAA-eligible endpoints to process protected health information once their organization is provisioned for Private Retention, even if data is retained.OpenAI, OpenAI Platform documentation, 2026
OpenAI's per-endpoint table lists /v1/audio/transcriptions with no abuse monitoring retention, no application state retention, and as eligible for Zero Data Retention.OpenAI, OpenAI Platform documentation, 2026
Oregon's ban on obtaining a telecommunication that you are not a participant in does not apply when at least one participant consents.Oregon Legislature, Oregon Revised Statutes (oregon.public.law), n.d.
Oregon's statute makes it an offence to obtain a conversation by device when not all participants have been specifically informed that the conversation is being obtained.Oregon Legislature, Oregon Revised Statutes (oregon.public.law), n.d.
Orrick says Anthropic's consumer privacy policy, to which every user consents, lets the company use user inputs for training and reserves the right to disclose them to third parties.Orrick, Herrington & Sutcliffe LLP, Orrick, 2026
Otter's privacy policy lists a right to opt out of the sale of your personal information and of its use for targeted advertising. The rights section does not list an opt-out from AI training.Otter.ai, 2026
Otter's privacy policy says it discloses personal information to vendors and service providers, including data labeling providers that use the data it shares to create training and evaluation data for its product features.Otter.ai, 2026
Otter's privacy policy says it keeps de-identified data in de-identified form and does not try to re-identify it, except to test whether its de-identification works.Otter.ai, 2026
Otter's privacy policy says that if it combines aggregated data with your personal information so that it can identify you, it treats the combined data as personal information.Otter.ai, 2026
Otter's privacy policy, effective June 16, 2026, lists training its proprietary AI technology on de-identified audio recordings and on transcriptions among the purposes for which it uses information.Otter.ai, 2026
$95 million US dollarsReuters reported in January 2025 that Apple agreed to pay $95 million in cash to settle a proposed class action claiming Siri violated users' privacy.Jonathan Stempel, Reuters (via Yahoo Finance), 2025
Samsung's memo said the company was concerned that data sent to AI platforms such as Google Bard and Bing is stored on external servers, where it is hard to retrieve and delete.The National (Bloomberg report), 2023
3 years of imprisonment, maximumSection 201(1) of Germany's Criminal Code makes it an offence, punishable by up to three years in prison or a fine, to record without authorisation another person's spoken word that is not spoken publicly.Federal Ministry of Justice of Germany, gesetze-im-internet.de, n.d.
Stanford HAI reports that Anthropic changed its consumer terms so that conversations with its Claude chatbot are used for training by default, unless the user opts out.Jennifer King and Stanford Institute for Human-Centered Artificial Intelligence, Stanford Institute for Human-Centered Artificial Intelligence (HAI), 2025
Starling Bank says a phrase that no one else knows lets close friends and family verify that they are really speaking to each other.Starling Bank, Starling Bank, 2024
3 secondsStarling Bank says criminals can now clone a person's voice from as little as three seconds of audio.Starling Bank, Starling Bank, 2024
Superwhisper says that for its cloud models, the customer must sign Superwhisper's own business associate agreement.Superwhisper, n.d.
Superwhisper says that in its local-only setup no audio or transcripts leave the device, and that no BAA with Superwhisper is required for those models.Superwhisper, n.d.
Tennessee's ELVIS Act defines voice as a sound in a medium that is readily identifiable and attributable to a particular individual, whether or not the sound is the actual voice or a simulation of it.State of Tennessee, ETO AGORA (reproduction of the enacted bill text), 2024
Tennessee's ELVIS Act took effect on July 1, 2024.State of Tennessee, ETO AGORA (reproduction of the enacted bill text), 2024
Texas Opinion 705 applies the same test to generative AI tools as to other third-party service providers: the lawyer must be reasonably confident the provider will protect the information.Professional Ethics Committee for the State Bar of Texas, State Bar of Texas, 2025
Texas Opinion 705 says a lawyer should be reasonably satisfied that a generative AI tool will not reveal confidential information to others before the lawyer uses it with confidential material.Professional Ethics Committee for the State Bar of Texas, State Bar of Texas, 2025
Texas Opinion 705 says a lawyer who puts confidential information into a generative AI tool should consider telling clients about the risks, and may need their consent.Professional Ethics Committee for the State Bar of Texas, State Bar of Texas, 2025
139 keyboard appsThe 2015 peer-reviewed study analyzed the security behaviors of 139 keyboard apps available on Google Play.Junsung Cho et al., 2015 13th Annual Conference on Privacy, Security and Trust (PST), IEEE, 2015
200 US dollarsThe 2023 attack that recovered over ten thousand ChatGPT training examples used queries costing $200 USD.Milad Nasr et al., arXiv, 2023
The 2023 Google study states that neural networks can unintentionally memorize their training examples, which causes privacy concerns, and it sets out to measure this in large speech recognition models.Lun Wang et al., arXiv, 2023
The 2023 study says that Apple does not validate App Store privacy labels, but requires developers to provide a privacy policy that can be compared with them.Mir Masood Ali et al., arXiv, 2023
The ASVspoof 5 challenge summary reports that attacks significantly compromise its baseline systems, while submitted systems improve substantially.Xin Wang et al., arXiv (ASVspoof 5 Workshop, Interspeech 2024 satellite), 2024
The authors of a 2024 study say that speech can potentially identify speakers, so sharing voice recordings raises privacy concerns.Daniela Wiepert et al., JMIR AI, 2024
The authors of a 2024 study state that speech models are often trained on sensitive data to improve their performance, which can lead to privacy leakage.Matthew Jagielski et al., arXiv, 2024
1 documentThe Carlini et al. extraction attack worked on sequences that appeared in only one document in the training data.Nicholas Carlini et al., 30th USENIX Security Symposium, 2021
hundreds verbatim text sequencesThe Carlini et al. study recovered hundreds of verbatim text sequences from the training data of GPT-2.Nicholas Carlini et al., 30th USENIX Security Symposium, 2021
The CCPA defines biometric information to include voice recordings, from which a template such as a voiceprint can be extracted.California Legislature, via california.public.law, 2018
The CCPA lists as sensitive personal information the processing of biometric information for the purpose of uniquely identifying a consumer.California Legislature, via california.public.law, 2018
$2,500 and $7,500 US dollars per violationThe CCPA sets an administrative fine of up to $2,500 for each violation and up to $7,500 for each intentional violation.California Legislature, via california.public.law, 2018
The FBI advises you to verify the identity of anyone who calls or sends voice messages before you respond.FBI Internet Crime Complaint Center, Federal Bureau of Investigation (IC3), 2025
The FBI warned in May 2025 that malicious actors sent text messages and AI-generated voice messages claiming to come from senior US officials.FBI Internet Crime Complaint Center, Federal Bureau of Investigation (IC3), 2025
The FBI's December 2025 update says the impersonation activity dates back to 2023 and has widened to officials' family members and personal acquaintances.FBI Internet Crime Complaint Center, Federal Bureau of Investigation (IC3), 2025
The federal Wiretap Act says it is not unlawful for a person to intercept a wire, oral or electronic communication when they are a party to it or one party has given prior consent to the interception.Legal Information Institute, Cornell Law School, 2026
The FTC advises that, to check a call about a family emergency, you use a phone number you know is right to call or message the family member or friend who supposedly contacted you.Federal Trade Commission, FTC Consumer Advice, n.d.
The FTC advises that, when a caller claiming to be a family member demands urgent money, you should resist the pressure to react and send money immediately.Federal Trade Commission, FTC Consumer Advice, n.d.
The FTC said Cox Media Group's active listening service did not listen to consumers' conversations or use voice data at all.Wendy Davis, MediaPost, 2026
The GDPR defines biometric data as personal data from specific technical processing of a person's physical, physiological or behavioural characteristics that allows or confirms the unique identification of that person.gdpr-info.eu (Intersoft Consulting), 2016
The GDPR defines personal data as any information relating to an identified or identifiable natural person.gdpr-info.eu (Intersoft Consulting), 2016
The Guardian reported in July 2019 that Apple contractors regularly hear confidential medical information, drug deals and recordings of couples having sex as part of quality control for Siri.Alex Hern, The Guardian, 2019
The HIPAA regulation's definitions section says protected health information means individually identifiable health information, with some exclusions, such as education records covered by the federal education privacy law.Legal Information Institute, Cornell Law School (text of the Code of Federal Regulations), n.d.
The HIPAA Security Rule lists automatic logoff, which ends an electronic session after a predetermined time of inactivity, as an addressable implementation specification.U.S. Department of Health and Human Services (text reproduced by Cornell Legal Information Institute), 2013
The HIPAA Security Rule makes an accurate and thorough risk analysis of the risks to the confidentiality, integrity and availability of electronic protected health information a required implementation specification.Legal Information Institute, Cornell Law School, 2026
The HIPAA Security Rule requires a unique name or number for identifying and tracking each user of an electronic system that holds electronic protected health information.U.S. Department of Health and Human Services (text reproduced by Cornell Legal Information Institute), 2013
The HIPAA Security Rule's technical safeguards include an audit controls standard: mechanisms that record and examine activity in systems that contain or use electronic protected health information.U.S. Department of Health and Human Services (text reproduced by Cornell Legal Information Institute), 2013
The House memo that restricted Copilot gave as its reason the risk of data leaking to unauthorized cloud services.Priya Singh, Business Today, 2024
The Nuance Gatekeeper documentation says a voice biometrics system compares spoken utterances with previously enrolled voiceprints, and determines whether they are spoken by the same speaker.Nuance Communications, Gatekeeper documentation, n.d.
The Nuance Gatekeeper documentation says text-dependent verification uses a passphrase to identify the speaker.Nuance Communications, Gatekeeper documentation, n.d.
The Nuance Gatekeeper documentation says text-independent verification uses conversational audio of the speaker, rather than a fixed passphrase.Nuance Communications, Gatekeeper documentation, n.d.
The openai-whisper loader calls torch.load with weights_only set to True on PyTorch 1.13 and later, which restricts what the unpickler will load.OpenAI, GitHub (openai/whisper, main branch), 2026
The openai-whisper loader takes the expected SHA256 checksum from the download address of the model, then compares it with the hash of the file saved on disk.OpenAI, GitHub (openai/whisper, main branch), 2026
The openai-whisper package checks each model file against a SHA256 checksum and downloads it again when the checksum does not match.OpenAI, GitHub (openai/whisper, main branch), 2026
The openai-whisper README introduces command-line use by saying that the following command will transcribe speech in audio files.OpenAI, GitHub (openai/whisper), 2026
The plaintiffs in the Google Assistant lawsuit said Google had no right to use conversations for targeted advertising when Assistant misheard speech as a hot word.NBC News, NBC News, 2021
The PLOS ONE study found that giving listeners examples of speech deepfakes improves their results only slightly.Kimberly T. Mai and Lewis Griffin, PLOS ONE, 2023
The PyTorch documentation for torch.load warns that the function uses an unpickler and that users should never load data from an untrusted source.PyTorch contributors, PyTorch, 2026
11 statesThe Reporters Committee for Freedom of the Press says about 11 US states primarily require the consent of all parties before a conversation is recorded.Reporters Committee for Freedom of the Press, Reporters Committee for Freedom of the Press, n.d.
The Seventh Circuit held that the 2024 amendment limiting BIPA damages is a procedural change and must be applied to cases that were pending when the amendment was enacted.Paul Hastings, 2026
2.7 timesThe share of ABA survey respondents using AI-based technology rose about 2.7 times between 2023 and 2024.Calculated from other facts
The terms of service for Dragon Medical One name a business associate agreement among the vendor, the customer and the customer's healthcare group as one of the agreements that govern use of the product.Dragon Medical USA, 2025
The UK data protection regulator says all biometric data is personal data, because it relates to an identified or identifiable person.Information Commissioner's Office (UK), 2025
The UK regulator lists voice recognition among the physical or physiological biometric identification techniques, alongside facial recognition and iris scanning.Information Commissioner's Office (UK), 2025
The UK regulator says biometric data is special category data whenever it is processed for the purpose of uniquely identifying a natural person.Information Commissioner's Office (UK), 2025
The UK regulator's guidance says an organisation must include its lawful basis and its purposes for using personal information in its privacy information.Information Commissioner's Office, Information Commissioner's Office, n.d.
The Waterloo researchers say their method can fool most voice authentication systems within six attempts.University of Waterloo, University of Waterloo, 2023
U.S. law defines a FedRAMP authorization as a certification about a cloud computing product or service.Legal Information Institute, Cornell Law School, undated
Under Article 13(2)(a) of the GDPR, a controller must tell the person, when the personal data are obtained, how long the data will be stored, or the criteria used to set that period if a fixed period is not possible.gdpr-info.eu (text of Regulation (EU) 2016/679), 2016
Under Article 2(2)(c) of the EU GDPR, the regulation does not apply to processing of personal data by a natural person in the course of a purely personal or household activity.gdpr-info.eu (text of Regulation (EU) 2016/679), 2016
Under Article 28(1) of the GDPR, processing by a processor must be governed by a contract or other legal act that binds the processor to the controller.gdpr-info.eu (text of Regulation (EU) 2016/679), 2016
Under Article 42 of the GDPR, certification is voluntary, and a certification does not reduce the responsibility of the controller or the processor for compliance with the regulation.gdpr-info.eu (text of Regulation (EU) 2016/679), 2016
Under Article 50 of the EU AI Act, a deployer of a deepfake must disclose that the content has been artificially generated or manipulated.European Union, Regulation (EU) 2024/1689 (EU Artificial Intelligence Act), text mirror, 2024
Under BIPA, before a business collects a biometric identifier, it must obtain written consent from the data subject or their legal representative.Hogan Lovells, Hogan Lovells, 2019
1 year years of imprisonment, maximumUnder California Penal Code 632, a violation can carry imprisonment in a county jail for up to one year.California Public Law, 2026
5 years years of imprisonment, maximumUnder federal law, a person who violates the wiretap ban can be fined or imprisoned for up to five years, or both.Legal Information Institute, Cornell Law School, 2026
Under Federal Rule of Evidence 502, an inadvertent disclosure of privileged material does not waive the privilege when the holder took reasonable steps to prevent it and promptly took reasonable steps to fix the error.Legal Information Institute, Cornell Law School (Wex), 2026
third-degree felony felony classUnder Florida law, a person who violates the ban on intercepting communications is guilty of a felony of the third degree.Florida Legislature, Florida Senate, 2025
18 monthsUnder Gboard's opt-in conventional learning, Google says it will not keep the speech snippets you donate beyond 18 months.Google, Google Gboard Help, 2026
Under GDPR Article 28(2), a processor cannot engage another processor without the controller's prior written authorization, either specific or general.gdpr-info.eu (text of Regulation (EU) 2016/679), 2016
60 calendar daysUnder HIPAA's Breach Notification Rule, a covered entity must notify affected individuals without unreasonable delay and in no case later than 60 calendar days after discovering a breach.Legal Information Institute, Cornell Law School (reproducing the Code of Federal Regulations), n.d.
5 years years of imprisonment, maximumUnder Maryland law, a violation of the wiretap statute is a felony punishable by imprisonment for up to five years.Maryland General Assembly, Maryland General Assembly, n.d.
Under Tennessee's ELVIS Act, a person who violates the act is guilty of a Class A misdemeanor, and can also be sued in a civil action for money or an injunction.Kim Peretti, Alston & Bird, Cyber & Data Strategy Blog, 2024
Under the 2024 amendment to Illinois BIPA (SB 2979), a private entity that collects the same biometric identifier or biometric information from the same person by the same method in more than one instance commits only one violation, and the aggrieved person can recover at most once.Paul Hastings, 2026
5 years years of imprisonment, maximumUnder the Criminal Code of Canada, wilful interception of a private communication is an indictable offence with a maximum of five years in prison.Parliament of Canada, Justice Laws Website, Government of Canada, n.d.
Under the EU GDPR, personal data that has been pseudonymised still counts as information about an identifiable person when it can be attributed to that person with additional information.Official Journal of the European Union (EUR-Lex), 2016
Under the EU GDPR, processing personal data is lawful only if and to the extent that at least one of the listed conditions applies, and consent is one of them.Official Journal of the European Union (EUR-Lex), 2016
20 years in prison (maximum)Under the federal wire fraud statute, a scheme to obtain money by false pretenses, carried out by wire, radio or television communication, can be punished by a fine or up to 20 years in prison, or both.Cornell Law School Legal Information Institute, Cornell LII, n.d.
Under the GDPR, the ban on processing special category data does not apply when the person has given explicit consent for one or more specified purposes.gdpr-info.eu (Intersoft Consulting), 2016
Under the HIPAA Security Rule, a covered entity may let a business associate create, receive, maintain or transmit electronic protected health information only if it obtains satisfactory assurances that the business associate will safeguard it.Legal Information Institute, Cornell Law School, 2026
Under the UK Investigatory Powers Act 2016, it is an offence to intentionally intercept a communication in the course of its transmission without lawful authority.UK Parliament, legislation.gov.uk, 2016
99 percentUniversity of Waterloo researchers reported an attack that bypasses voice authentication with up to a 99% success rate after six tries.University of Waterloo, University of Waterloo, 2023
Washington makes it unlawful to record a private communication by device without first obtaining the consent of all the participants.Washington State Legislature, Washington State Legislature, n.d.
2 ways to delete stored batch outputWhen Azure batch transcription output is stored in Microsoft storage, you can delete it with a deletion API or set the timeToLive parameter so it is deleted automatically after a set time.Microsoft, Microsoft Learn, 2026
When you load a named Whisper model, the openai-whisper loader fetches the model file from its download address with urllib, unless a file with the right checksum is already in the cache folder.OpenAI, GitHub (openai/whisper, main branch), 2026
Windows clipboard sync is tied to a Microsoft account or a work account, so each device must use the same sign-in.Microsoft Support, n.d.
Wispr Flow says BAA signing is available on Mac, Windows and iOS.Wispr Flow Help Center, n.d.
Wispr Flow says its Growth and Enterprise organization plans both include a standard HIPAA business associate agreement.Wispr Flow Help Center, n.d.
Wispr Flow says members of organizations with active billing cannot sign personal BAAs and should ask their admin.Wispr Flow Help Center, n.d.
Wispr Flow says that an applicable signed business associate agreement locks both of its data-sharing controls off.Wispr Flow Help Center, n.d.
Wispr Flow says that downloading a signed BAA requires a recorded signer name and an individual, non-enterprise account.Wispr Flow Help Center, n.d.
Wispr Flow tells users to confirm both data-sharing settings are off before dictating protected health information.Wispr Flow Help Center, n.d.