In 2019, a team of researchers at Northeastern University and Imperial College London ran one of the most thorough independent tests of smart speaker privacy to date. Over six months, they played 125 hours of scripted dialogue and ambient household sound past eight popular smart speaker models. They measured how often each device woke up and began recording without being deliberately addressed. Every device had false activations. The Amazon Echo activated unintentionally an average of 1.5 to 19 times per day, depending on the wake word and the content playing nearby. The researchers published their findings and shared them with manufacturers. Amazon disputed the methodology. The speakers kept recording.
That study, published in the Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, didn’t receive much coverage outside privacy circles. It should have.
The short answer: Smart speakers listen constantly for a wake word, and the local model gets it wrong often, up to 19 times a day for some Echo setups. Each false trigger uploads real audio to Amazon, Google, or Apple servers, where it’s stored by default and sometimes reviewed by human contractors. Police have obtained Echo recordings in court cases. You can cut the exposure without new hardware: use the hardware mute button, delete voice history, set auto-delete, and turn off human review and voice purchasing.
Jump to: Why false activations matter · What the companies store · Law enforcement access · Voice data and ads · What to change · Network-level separation
Why false activations are the core problem
Smart speakers operate in a state of constant, low-level listening. The device’s onboard chip runs a wake-word detection model. This is a small, local system built to recognise one specific trigger phrase. When it thinks it hears that phrase, it opens an audio stream to the company’s cloud servers. From that point, everything you say is sent, processed, and stored.
The local model is the problem. It is not very good. It has to be small enough to run on a modest chip, which means it makes mistakes. “Unboxing” and “Texas” have both been documented as false triggers for Alexa. A TV news broadcast, a podcast, a snatch of conversation from another room: any of these can cross the threshold. When that happens, the audio stream opens. Seconds or minutes of audio get uploaded, analysed, logged with a timestamp, and linked to your account. In most default configurations, that log is permanent.
Google Assistant’s wake phrase, “Hey Google,” has been triggered by the phrase “OK, Boomer.” Siri has been activated by the sound of zips. These are not hypothetical edge cases; they appear in user-reported complaints, contractor accounts, and in the data that privacy researchers have obtained through subject access requests.
The Northeastern and Imperial College study was significant precisely because it was systematic rather than anecdotal. The researchers created a controlled environment and measured activation rates methodically. The finding was up to 19 false activations per day for some Echo configurations. That means a device in a typical home could send dozens of accidental recordings to Amazon’s servers each week. It records ambient conversation that was never meant for any server anywhere.
What the companies actually store
Each major assistant has a different approach to retention, but they share a common default: they keep your recordings.
Amazon Echo stores voice recordings indefinitely unless you manually delete them. Each recording is accessible in the Alexa app under Settings → Alexa Privacy → Review Voice History. You can scroll back through years of entries. More significant than the storage itself is what Amazon does with it. The company employs teams of contractors in multiple countries, including Romania, India, and Costa Rica, who listen to voice recordings as part of a “quality assurance” programme. Bloomberg reported on this in April 2019. The contractors heard domestic arguments and medical conversations. They heard what appeared to be a sexual assault being described. They heard material captured by accident, from people who had no idea anyone was listening. Workers told Bloomberg that some recordings were obvious mistakes, but they were being reviewed regardless.
Google Nest stores voice and audio activity in your Google account, under a category called Voice & Audio Activity. For years, this was on by default. GDPR enforcement in the EU eventually pushed Google to change the default for European users. In 2019, a Belgian public broadcaster, VRT NWS, obtained and published a sample of more than 1,000 Google Assistant recordings. They came from a contractor with access to the review programme. The recordings included fragments of domestic conversations, apparent medical symptoms, and confidential exchanges. Google acknowledged the leak and suspended the human review programme in the EU.
Apple Siri attracted similar controversy in July 2019. The Guardian reported that Apple contractors were regularly hearing private medical discussions between doctors and patients, confidential business meetings, and details of what appeared to be criminal activity. All of it was captured via accidental Siri activations. Unlike Amazon and Google, Apple asserted that Siri recordings are not linked to Apple IDs. Apple uses random identifiers rather than account-linked ones, which is a meaningful distinction for targeted profiling. But the recordings were still being reviewed by humans who had not been disclosed to users. Apple paused the programme and later made it opt-in.
These were not isolated events. They describe a structural feature of how these systems work: audio goes to servers, humans occasionally review it, and the disclosure to users was, in each case, insufficient.
Law enforcement and your recordings
Stored recordings can be accessed by law enforcement. This is not a hypothetical — it has happened in court cases.
In 2017, prosecutors in Benton County, Arkansas, sought recordings from an Amazon Echo that had been present during an alleged murder. The device had belonged to James Bates, who was charged with killing a friend in his hot tub. Amazon initially challenged the warrant on First Amendment grounds, arguing that the assistant’s responses constituted speech with constitutional protections. Amazon ultimately provided the recordings — not because the warrant challenge failed, but because Bates’s own defence counsel consented to the disclosure. Charges were later dropped, but the legal exchange established a template.
Later in 2017, in Farmington, New Hampshire, prosecutors sought Echo recordings in a double murder case. A couple, Timothy and Meredith Sullivan, had been killed, and police believed an Echo in the home may have captured audio around the time of the killings. Amazon challenged that subpoena. The recordings were eventually provided.
The Electronic Privacy Information Center (EPIC) is a Washington-based privacy advocacy organisation. It tracked at least 11 cases through 2021 in which Amazon complied with law enforcement requests for Alexa voice recordings. The actual number is almost certainly higher, because many such disclosures happen with no public record.
The legal standard matters here. The Wiretap Act provides meaningful protection against real-time interception of communications. Authorities generally need a wiretap order, which requires showing probable cause to a judge. But stored recordings are a different category. Under the Stored Communications Act, law enforcement can access stored records with a subpoena, which requires a lower standard than a full warrant. Your recordings, once stored, are subject to this lower threshold. A subpoena doesn’t require a judge’s sign-off. A grand jury or even a government attorney can issue one in many circumstances.
Outside the United States, the protections vary considerably. In the UK, the Investigatory Powers Act 2016 gives GCHQ and the police broad powers to request data from communications service providers, which Amazon, Google, and Apple are. Amazon’s global law enforcement request numbers are disclosed in transparency reports. But these aggregate figures obscure the specific categories of data requested.
Voice data and the advertising economy
Amazon’s primary business is not selling smart speakers. It is selling advertising and driving purchases. The Echo is, among other things, a data collection node.
Voice interactions with Alexa inform the wider Amazon advertising system. “Alexa, add this to my cart” is a purchase signal. “Alexa, remind me to buy milk” is a signal about household routines. “Alexa, what time does the pharmacy close?” is a health-adjacent query. Amazon’s Privacy Notice acknowledges that it uses information collected from Alexa to improve its services — and “services” includes advertising.
The specific mechanics of how voice interaction data is used to personalise advertising are not publicly disclosed in granular form. Amazon has been cautious about making explicit claims in either direction. What is documented: Amazon’s advertising business generated $56.2 billion in revenue in 2024. The Echo lives in people’s kitchens and living rooms, listening to household life. That role is part of what makes the broader data ecosystem valuable.
Google is more explicit. Its Privacy Policy states that information collected across its services can be used to provide “more relevant ads.” That includes voice interactions via Google Nest. Voice data joins the wider Google account profile: search history, YouTube viewing, Chrome browsing, Maps location history, and Gmail content (in the case of Smart Compose and related features). Each signal reinforces the others.
Apple makes stronger claims about keeping advertising data separate. It does not run a behavioural advertising business the way Amazon and Google do. Siri data’s random identifier system also keeps some separation. That distinction is real. But it does not address the human review issue, and it doesn’t prevent law enforcement access.
What you should actually change
The good news: there are concrete steps that reduce exposure meaningfully. None of them require buying new hardware.
Use the mute button. Every major smart speaker — Echo, Nest, HomePod — has a physical mute that disables the microphone at the hardware level. This is more reliable than any software setting. When you are in a sensitive conversation, when you are on a work call in the same room, when guests are over — use the mute button. The speaker cannot send audio it cannot record.
Amazon Echo: Open the Alexa app. Go to More → Settings → Alexa Privacy → Review Voice History. Delete all recordings. Then go to Manage Your Alexa Data and set “Automatically delete recordings” to either 3 months or “Don’t save recordings.” Disable “Use of voice recordings” and “Help improve Amazon services and develop new features.” If you have voice purchasing enabled, disable it under Settings → Account → Voice Purchasing. This doesn’t stop recordings from being made temporarily during processing, but it limits long-term retention.
Google Nest: Log into your Google account. Go to Data & Privacy → My Activity. Find “Web & App Activity,” click Manage, then filter by “Voice and Audio.” Delete all existing voice activity. To stop future saving: back on the main Data & Privacy page, find “Voice & Audio Activity” and turn it off. EU-based users should check whether Google has re-enabled this without notification; it has happened following app updates.
Apple HomePod and Siri: On iPhone, go to Settings → Siri & Search → scroll to “Improve Siri & Dictation” and disable it. This opts you out of the human review programme. On HomePod itself, use the Mute button for sensitive conversations. Apple’s random identifier system already provides more separation than Amazon or Google, but the opt-out removes the human review component entirely.
Disable features you don’t use. Alexa’s “Drop In” function lets two Echo devices open an audio channel between them with minimal friction. It was designed for quick inter-room communication within a household. But if your account credentials were ever compromised, an attacker with access to your Amazon account could potentially use Drop In to listen into a room. If you don’t actively use Drop In, disable it: Alexa app → Device Settings → Communications → Drop In → Off.
Network-level separation
If you want stronger protection without removing the devices, network segmentation is worth understanding.
A home router that supports VLANs — virtual local area networks — can place your smart speakers on a separate network segment that cannot communicate with your computers, phones, or NAS drives. This means that even if a speaker’s firmware were compromised, the attacker cannot pivot to other devices on your network. Eero, Ubiquiti’s UniFi range, and many other consumer routers support this. The setup requires moderate technical comfort, but guides exist for each major platform.
DNS-level blocking via Pi-hole or NextDNS lets you block known telemetry domains at the network level. NextDNS maintains curated blocklists that include Amazon and Google telemetry endpoints. Blocking these does not prevent all data transmission — recordings go to core processing endpoints that cannot be blocked without breaking the device — but it can reduce ancillary data collection. Be aware that some blocks will break device functionality in ways that are difficult to trace.
These are not steps for everyone. But they exist. And for someone who has a home office where sensitive client conversations happen near an Echo that they haven’t reconfigured, they’re worth knowing about.
The 2019 Northeastern and Imperial College paper ended with a measured call for more transparency from manufacturers about false activation rates and what happens to accidentally captured audio. Six years on, Amazon, Google, and Apple have each made incremental changes to defaults and human review policies. All three still store recordings by default. All three still share data with law enforcement on request. The false activation problem hasn’t been solved — it’s been managed.
The mute button, the deletion settings, and the opt-outs exist. They require you to go looking for them. For the rest of the connected home, our smart home privacy guide covers TVs, cameras, and thermostats.
What you can do right now
Want to see what data brokers expose about you? See what a scan finds → shows an example report: the kinds of sites that list your name, phone, and address. Once KnytX Shield launches it will find and remove these automatically. Join the waitlist to get in early.