Amazon Alexa Is Listening: What They Don't Want You to Know
Amazon placed a microphone in more than 500 million homes and called it a convenience device. In 2019, Bloomberg revealed that thousands of Amazon employees were listening to those recordings every day. What followed was a masterclass in corporate damage control — partial admissions, reframed disclosures, and buried settings. Here is the full documented record of what Alexa actually collects, who hears it, and what Amazon has quietly filed patents to do with it.
The pitch was simple: a cylinder on your kitchen counter that plays music, answers questions, and controls your smart lights. You talk to it, it talks back. Amazon launched the first Echo device in November 2014 and by 2019 had sold an estimated 100 million units worldwide. By 2024, that number had grown past 500 million Alexa-enabled devices. Each one contains a microphone that is, by design, always active.
Amazon's official position is that Alexa only records and transmits audio after detecting its wake word — “Alexa,” “Echo,” or “Computer.” What the company did not prominently disclose in its marketing materials, its setup process, or its terms of service summary was that recordings triggered by that wake word were being reviewed by human employees, that those recordings frequently captured conversations the user never intended to share, that the data was being used to train advertising systems, and that in at least one documented case, the device had forwarded a private conversation to a contact in the user's address book without any command to do so.
This is the documented record of Amazon Alexa's surveillance capabilities— what was revealed, what was admitted, what was buried, and what it means for anyone with a smart speaker in their home.
The Bloomberg Revelation: Amazon's Listening Team
On April 10, 2019, Bloomberg published an investigation by Matt Day, Giles Turner, and Natalia Drozdiak with a headline that landed like a brick: “Is Anyone Listening to You on Alexa? A Global Team Reviews Audio.” The report drew on interviews with seven current and former Amazon employees who worked on the Alexa voice improvement program.
Their accounts were specific. A dedicated team — located in offices in Romania, India, Costa Rica, and the United States — spent shifts listening to Alexa recordings, transcribing them, annotating them, and feeding corrected data back into the machine learning systems that power Alexa's speech recognition. Each reviewer processed as many as 1,000 audio clips per shift. The clips ranged from benign requests — “Alexa, play jazz” — to intimate domestic moments that had been captured when the wake word was falsely triggered. The employees described hearing arguments between couples, children's voices, and in two cases that employees reported to supervisors, what sounded like a sexual assault and a criminal act.
The employees reported that supervisors told them they were not authorized to take action on distressing content. Their job was transcription and annotation. Whatever they heard stayed in the system. There was no intervention protocol.
Amazon's Response: What They Admitted and What They Didn't
Amazon's statement in response to the Bloomberg story confirmed that human review of Alexa recordings was real: “We use the information to train our speech recognition and natural language understanding systems, so Alexa can better understand your requests.” The company said the practice is “an industry-standard practice” and that annotators are “required to adhere to Amazon's confidentiality requirements.”
What Amazon did not confirm: how many employees were involved, in how many countries, reviewing what total volume of recordings. It did not confirm whether users were meaningfully informed that human reviewers would hear their recordings. It did not confirm what percentage of recordings were false triggers — activations that occurred without any intentional wake word command. And it did not confirm what, if anything, happened to recordings that captured sensitive content beyond the annotator's mandate.
The phrase “industry-standard practice” did the heaviest lifting in Amazon's defense. It was true that Apple, Google, and Microsoft all had similar human review programs for their voice assistants. All three companies faced the same backlash in 2019 and all three made similar post-hoc disclosures. The fact that the practice was industry-standard did not make it disclosed. It made it normalized — which is a different thing.
How the “Always Listening” Architecture Actually Works
Amazon's technical description of Alexa's wake word system is accurate as far as it goes. The device runs continuous, local audio processing using a dedicated chip. That processing is designed to detect the specific acoustic pattern of the wake word and nothing else. Audio is not transmitted to Amazon's servers until the wake word is detected. After detection, the device records the following utterance and sends it to Amazon for processing.
The problem is false positives. Wake word detection is a statistical process, not a deterministic one. The algorithm assigns a probability score to incoming audio. When the score exceeds a threshold, the device activates. That threshold is set by Amazon as a balance between responsiveness (triggering when the user actually says the wake word) and precision (not triggering when they don't). In practice, the balance produces a meaningful false trigger rate.
Amazon has never publicly disclosed its false trigger rate. Independent researchers at Northeastern University and Imperial College London conducted tests in 2020 and found that Alexa devices could be triggered by words and phrases from television programs, other conversations in the room, and even similar-sounding words in unrelated speech. Northeastern University's study identified over 1,000 word sequences that triggered Alexa devices in testing. “Unacceptable” triggered Alexa. So did “election.” So did “a letter.”
Every false trigger is a recording. Every recording enters the system. Some percentage of those recordings — the exact percentage unknown — were reviewed by human annotators. This is the architecture that Amazon described as operating only on user commands. The description was technically accurate and substantively misleading.
The Portland Incident: When Alexa Sent a Private Conversation to a Contact
In May 2018 — a year before the Bloomberg story — a Portland couple named Danielle and her husband discovered that their Amazon Echo device had recorded a private conversation between them and sent the recording to a contact in their address book. The contact was a colleague of the husband who lived in Seattle. He received the recording and called them to warn them that he had been listening to their conversation.
Amazon investigated and provided the couple with an explanation of what had occurred. The device had falsely activated on a word that resembled “Alexa.” The subsequent conversation had been interpreted by the device as a sequence of commands: it heard a word that sounded like “send message,” confirmed the command when asked “to whom?” by interpreting the background conversation as a name in the contact list, and sent the recording when it heard what sounded like “right” as a confirmation.
Amazon's statement called the incident “an extremely rare occurrence” and said the company was taking steps to make accidental messaging “less likely.” What the company did not say was how many similar incidents had occurred and not been reported, how many contacts had received accidental recordings and not informed the sender, or what changes to the underlying architecture would prevent the sequence from recurring.
The Portland incident is notable not because it was a deliberate surveillance act but because it demonstrated that the surveillance infrastructure — the always-on microphone, the continuous audio processing, the integration with contact lists and messaging systems — could produce real privacy breaches through entirely automated processes, without any human intent.
What Alexa Actually Collects: The Full Data Inventory
Voice recordings are the most discussed element of Alexa's data collection, but they are not the only one. Amazon's privacy policy for Alexa covers a substantially broader range of data than most users are aware of.
Alexa collects and stores voice recordings by default, indefinitely, unless the user explicitly deletes them. Those recordings are associated with the user's Amazon account and linked to their purchase history, browsing history on Amazon's retail platform, and — through the Amazon ecosystem — their Prime Video viewing history, their Kindle reading history, and their demographic data from the broader Amazon profile.
When Alexa is connected to a smart home ecosystem — smart locks, thermostats, lights, cameras — it also collects data on device usage patterns: when lights are turned on and off, when the thermostat is adjusted, when doors are locked and unlocked. This produces a behavioral profile of the household's occupancy patterns, sleep schedule, and daily routines that is distinct from and complementary to the voice recording data.
Alexa for Kids, marketed to parents as a way to give children a safe, filtered assistant experience, collects voice recordings of children. The retention and use of children's voice data became the basis of a Federal Trade Commission investigation that concluded with an $25 million fine in 2023 — the largest civil penalty in FTC history at the time. The FTC found that Amazon had retained children's voice recordings for years after parents requested deletion, that it had used those recordings to train its algorithms in violation of the Children's Online Privacy Protection Act, and that it had misled parents about what deletion actually deleted.
This data — voice, behavioral, transactional, demographic — is what surveillance capitalism theorists describe as behavioral surplus: data generated by users as a byproduct of using a service, then extracted and used to build predictive models that can be monetized independent of the original service transaction.
The Patent Record: What Amazon Filed in Private
Patent filings are one of the most reliable indicators of a company's actual intentions, because they require disclosure of specific technical implementations — implementations that companies generally only describe in detail when they plan to use them or prevent competitors from using them.
Amazon has filed multiple patents for systems that use voice data collected by Alexa for targeted advertising purposes. A 2018 patent filing describes a system for analyzing “voice characteristics” in Alexa recordings to infer the user's emotional state and deliver targeted advertising based on that inference. The patent specifically describes identifying whether a user sounds “happy,” “frustrated,” or “excited” and using that emotional profile to select advertising content.
A separate patent filing describes systems for detecting when users mention specific products or brands in conversation — not in commands to Alexa, but in ambient conversation — and using those mentions to trigger relevant advertising. The system described in the patent would identify, for example, a user saying “I need to get some cough medicine” in conversation with another person and serve them advertising for cold remedies.
Amazon has not confirmed that these patented systems are currently deployed. Patent filings do not require commercial implementation. But the existence of these filings establishes that Amazon has specifically designed and documented the technical architecture for using ambient voice data for advertising targeting — and has protected that architecture as proprietary intellectual property.
Amazon's public statements consistently describe Alexa as a service designed to help users. The patent record describes a different product: a data collection system with specifically documented advertising monetization pathways. Both things can be true simultaneously. That is precisely the architecture that makes smart speakers commercially viable.
Ring Doorbell: Extending the Network Beyond the Home
Amazon acquired Ring, the video doorbell and home security camera company, in 2018 for approximately $1 billion. The acquisition connected Alexa's interior audio surveillance infrastructure to Ring's exterior video surveillance network, creating an integrated home surveillance system under a single corporate owner.
Ring's data practices came under scrutiny in 2019 when investigative reporting revealed that Ring had partnerships with over 400 law enforcement agencies across the United States that gave police the ability to request Ring footage from homeowners through an automated system. Homeowners could decline, but the existence of the pipeline — the direct integration between Ring's platform and law enforcement request systems — represented a significant expansion of the surveillance infrastructure that Amazon had built and sold as home security.
In 2022, Ring disclosed that it had provided footage to law enforcement without user consent or a court order on 11 occasions in the first half of the year, citing an emergency exception in its terms of service. Amazon defended the practice as legally authorized and operationally appropriate in emergency circumstances. Civil liberties organizations noted that “emergency” was defined by Ring and Amazon, not by any judicial authority.
The combined Alexa-Ring system means that Amazon now operates, in many homes, a device that records audio inside the house and a device that records video of everyone who approaches from outside. Both feed into the same Amazon account. Both are covered by the same privacy policy. Both have documented pathways to law enforcement.
Children's Privacy: The FTC Investigation and $25 Million Fine
The Federal Trade Commission's 2023 action against Amazon over Alexa's handling of children's data is the most consequential regulatory outcome of Amazon's voice surveillance practices to date. The complaint, filed jointly with the Department of Justice, alleged specific violations that Amazon settled for $25 million without admitting wrongdoing.
The FTC found that Amazon retained children's voice recordings “indefinitely” even after parents requested deletion. The complaint described the internal process: when a parent requested deletion of their child's voice history, Amazon deleted the voice recordings from the Alexa interface but retained the transcripts derived from those recordings in a separate database used for algorithm training. Parents were not informed that deletion of voice files did not mean deletion of derived transcripts.
The complaint also found that Amazon's geolocation data retention practices for children violated COPPA — the Children's Online Privacy Protection Act — and that Amazon had specifically overridden internal recommendations from employees who had flagged the deletion practices as legally problematic.
That last detail is significant. The FTC complaint does not describe a company that failed to understand its legal obligations. It describes a company that had internal awareness of those obligations, received internal warnings that its practices violated them, and continued those practices anyway. The $25 million fine, applied to a company whose quarterly revenue exceeds $140 billion, represented approximately 22 minutes of Amazon's revenue at the time of settlement.
The NSA Parallel: Voice Data in the Mass Surveillance Context
Amazon's voice data practices do not exist in isolation. They sit within a broader documented history of mass data collection that includes the NSA's PRISM and ECHELON surveillance programs, both of which collected data from major technology companies — in PRISM's case, with the companies' participation and in some instances their active cooperation.
Amazon Web Services is the infrastructure provider for a significant portion of the United States intelligence community. The CIA awarded Amazon a $600 million cloud computing contract in 2013. The NSA, the Defense Intelligence Agency, and other intelligence community components have subsequently moved onto AWS. Amazon and the intelligence community are, in a contractual sense, the same infrastructure.
This does not mean that Alexa recordings are being reviewed by intelligence agencies. There is no public evidence that they are. What it means is that the technical and legal pathways for such access exist — through the established PRISM framework, through National Security Letters, through Foreign Intelligence Surveillance Court orders — and that Amazon has a documented financial relationship with the agencies that would be the recipients of such access. The distinction between “is happening” and “has the architecture to happen” matters, but both are real.
How to Check and Delete Your Alexa Recordings
Amazon does provide mechanisms for reviewing and deleting Alexa recordings, though the settings are not surfaced prominently in the default device setup. Here is what the process actually involves.
Reviewing Your Recordings
Open the Alexa app on your phone. Navigate to More → Settings → Alexa Privacy → Review Voice History. You will see a timestamped list of all recorded interactions. Each entry shows the transcript of what Alexa heard and an audio playback option. Many users report finding recordings they have no memory of triggering — these are the false wake word activations.
Deleting Recordings
You can delete individual recordings, recordings by date range, or all recordings. The deletion option is in the same Review Voice History menu. You can also configure automatic deletion at three or eighteen months via Manage Your Alexa Data. Voice-command deletion is also available: saying “Alexa, delete everything I said today” will delete that day's recordings.
Opting Out of Human Review
Following the 2019 Bloomberg story, Amazon added an opt-out for human review of recordings. In the Alexa app: More → Settings → Alexa Privacy → Manage Your Alexa Data → Help Improve Alexa → toggle off “Use of voice recordings.” Note that as of the time of writing, this opt-out applies to your recordings being included in the annotation queue. It does not opt you out of Alexa recording and processing your voice entirely — that is the function on which the device depends.
It is worth noting that the existence of these controls — buried three menus deep, defaulting to maximum data collection, requiring affirmative action to restrict — is itself a design choice. The default state of an Amazon Echo is full recording retention, human review participation, and indefinite storage. Restriction requires user initiative and specific knowledge of where the settings are located. This is not an accident.
What This Means: The Smart Home as Surveillance Infrastructure
The documented facts of Alexa's data practices are not a conspiracy theory. They are a confirmed pattern, established through investigative reporting, regulatory filings, patent records, and corporate admissions. The question is not whether the surveillance is happening. The question is what framework to use to evaluate it.
Amazon's framework is that data collection improves the product, that human review is an industry-standard quality control practice, and that users consent to data collection when they agree to the terms of service. Each of these claims is technically accurate. None of them was what users understood themselves to be agreeing to when they plugged in an Echo and said, “Alexa, play something.”
The alternative framework is the one developed by scholars of surveillance capitalism: that smart home devices are primarily data extraction instruments that provide genuine consumer utility as a mechanism for achieving the scale of data collection that makes the extraction valuable. The device is the product. The behavioral data generated by using the device is the commodity. The consumer who paid $99 for an Echo is not Amazon's customer. They are Amazon's raw material.
That framing is contested. Amazon would reject it. But Amazon has also filed patents for voice-based emotional targeting systems, operated a human review program without prominent disclosure, retained children's data after parents requested deletion in violation of federal law, provided Ring footage to law enforcement without court orders, and built the infrastructure that connects 500 million microphones to a corporate data system that is also the primary cloud provider for the United States intelligence community.
These are not allegations. They are the documented record. The full sourced claim is tracked in the They Knew database. What you do with the device on your kitchen counter is up to you. But it is worth doing it with accurate information about what the device actually is.
See the full sourced claim: Amazon Alexa recording conversations →
Glossary: What is surveillance capitalism? →
Glossary: What is behavioral surplus? →
Related investigation: NSA mass surveillance — PRISM and ECHELON →
Independent research and fact-checking team documenting conspiracy theories proven true by primary sources, declassified documents, and official records.