Comment by alvah
1 year ago
Every time I read some technical description about why this isn't happening, the technical description seems convincing.
However...
A friend tested the theory a few years ago. He doesn't own a swimming pool, doesn't want to, and has never expressed any desire to. He put his and his wife's phone on the table and said to the wife (loudly), "Why don't we look into pool fencing?". She agreed with him. Shortly after, on both of their phones, on a particular social network, they were inundated with ads for....pool fencing.
Think about what this implies. If your phone is listening, it’s listening all the time, right? So like 12-18 hours of continuous audio every day. That’s a lot of ad triggers. Way too many to actually be served with ads during your browsing time, which is a strict subset of your total audible proximity to your phone (plus ad inventory is a strict subset of what you view on your phone).
So how does the phone + ad networks decide which words to prioritize to trigger which ads when?
So for this anecdote to be true, not only would the phone have to be listening, but the targeting algorithm would need to decide to actively exclude all the other audible triggers from that time period, and fill your limited ad impression inventory with the one phrase you were intentionally testing.
How would it do that? Especially if this is indeed an outlier one-off topic of conversation that you cover in a single sentence. There would not be contextual clues (like repetition over time) that might indicate you are actually “in market” for a pool fence.
To me this is the problem with these anecdotal tests. You understood that that was an important phrase in the context of ad targeting. But how did the automated ad system know it should serve you ads on that topic, and not one of the many other advertisable topics you talk about over the course of several days? Or that your phone hears over several days?
1) App stores the trailing two minutes of speech in memory.
2) If the app detects a consumption-related trigger word, the related conversation is flagged for transmission to the server.
3) Flagged audio block is converted to text. Consumption related verbs ("buy", "purchase", etc) are identified. The syntax of the sentence clearly indicates which noun is the target of a given consumption-related verb ("new car", "pool fencing")
4) Serve related ads
Where's the proof that this is happening?
Lots of people run network traffic sniffers to see what apps are doing. Lots of people decompile apps. Lots of people at companies leak details of bad things they are doing.
Why has nobody been able to demonstrate this beyond anecdotes about talking about swimming pools and then getting adverts for swimming pool stuff?
21 replies →
Your voice is unique and can be fingerprinted to ID you (see Alexa devices). Add in things like positive sentiment analysis, changes in vocal inflection/intonation and context surrounding spoken products like purchase inference/intent and you can probably triangulate a threshold for showing products with high likelihood of purchasing intent.
Really smart people have been working on these things at Google for decades and that’s barely scratching the surface of this nuanced discussion. CPU/GPU has only gotten faster and smaller with more RAM available and better power management across the board for mobile devices.
Anything is possible if there is money to be made and it’s not explicitly illegal or better they can pay the fines after making their 100x ROI.
I guess my question to this is, what are the estimates on energy required to do this on a continuous timeframe?
7 replies →
My phone can listen all day every day. It listens for "hey google" and it can listen and passively tell you songs that are playing. It's not outside the realm of possibility to do their audio fingerprinting on keywords and what not. The advertising potential makes it extremely juicy
Your phone can listen for “hey Google” because it’s only one phrase and the model can run at very low power on specialized hardware. If you want to add 1000 keywords the battery drain would be intense.
12 replies →
The system knows to serve you ads about the new topic because it's new. You're already getting ads for the stuff you're normally talking about. The new topic stands out easily.
It doesn't have to be your phone. Could be your TV or any other device.
Most importantly there's just patterns of behavior. Companies are absolutely desperate for every scrap of data they can get on you. Why would they not capture audio from your mic?
You’re so right. We should just trust the computers in our pockets, hands, and nightstands 24/7/365 running proprietary operating systems, firmware, and sensor suites phoning home as much targeting data as they can possibly collect — but not that! What could they possibly gain from harvesting that?
Companies really are using tons of highly sensitive data to target ads, even when we sleep. But they're not generally using microphones to record audio to do it. Both things can be accurate statements.
It's not a strict 12-18 hour window. Instead, it depends on the time frame between specific vocal or conversational cues / signal vs. noise.
[dead]
>So how does the phone + ad networks decide which words to prioritize to trigger which ads when?
The same way they analyze your email and web searches. Basically, statistics.
>To me this is the problem with these anecdotal tests. You understood that that was an important phrase in the context of ad targeting. But how did the automated ad system know it should serve you ads on that topic, and not one of the many other advertisable topics you talk about over the course of several days? Or that your phone hears over several days?
Buddy, so many people have witnessed this happening for at least 10 years and even done experiments at this point that it's common knowledge. I know for a fact that one of my friends now has a phone that is especially receptive to hearing me say things around it, because our conversation topics ALWAYS come up in my searches, ads, and feeds shortly after. Think about that. Someone else's phone sends data to a cloud that I never gave permission to. It then puts that together with data from MY phone about where I was (perhaps even the devices chirping at each other!). The aggregation happens within a week then I see relevant ads. I've seen this happen dozens of times. It's no coincidence.
As far as the article, I'm not even going to read it. It's got to be stupid. We know from leaks, reverse-engineering, and personal experience that this spying is going on. I question the source of this article, but I suppose we should never underestimate the lengths someone will go to in order to feel that they are smarter than the rest of us with our eyes open.
"We know from leaks, reverse-engineering"
I would be VERY interested to hear details of those leaks and that reverse-engineering. I've only ever heard the personal anecdotes.
(If you'd read my article you would have seen this bit at the top: "Convincing people of this is basically impossible. It doesn’t matter how good your argument is, if someone has ever seen an ad that relates to their previous voice conversation they are likely convinced and there’s nothing you can do to talk them out of it.")
10 replies →
>We know from leaks, reverse-engineering, and personal experience that this spying is going on.
No we don't. There isn't any of that. This is flat earthing for technophiles.
12 replies →
"How would it do that? "
AI.
That is entire premise of 'Nexis' from Yuval Harari.
Individualized bot driven surveillance .
https://www.theregister.com/2024/09/16/oracle_ai_mass_survei...
""Ellison declares Oracle all-in on AI mass surveillance, says it'll keep everyone in line
Cops to citizens will be 'on their best behavior because we're constantly recording and reporting'""
How were they doing this in 2017?
1 reply →
Harari is a shill.
4 replies →
Why did they pick a swimming pool? Did they see people in their area installing pools? I think that's often people's best guess, is that the "random" thing people use to test this actually isn't random and subconsciously they already had this topic seeded to them.
Something similar -- while on a family visit at my parents' house, my brother was talking about his upcoming Hawaii trip, Specifically he was going over a snorkeling adventure he signed up for.
For the next week or so, I got many ads on my phone about underwater packages for Hawaii, along with ads for various snorkeling and swimming gear. Now I had never researched any of that on my phone, however obviously my brother has. And the ad trackers saw that both my phone and his had communicated out over the same IP address (my parents wifi) on other random internet connections, so that is probably why they were then targeting my tracker cookie with ads that would be related to his tracker cookie. (This is all technically "easy" for the trackers to do, and seems logical that they would, because "why not").
On an unrelated note, I was making a peanut butter sandwich, started browsing some sites, and started getting ads for Skippy peanut butter. My phone must have smelled the peanut butter in the air.
Until my wife installed UBo, this was helpful in finding presents for her. Because I was using UBo, I could switch it off, browse Facebook and all of my ads were being targetted at her. I could see everything she was considering buying.
It was impressively creepy and a good way of surprising her with something she hadn't said anything about but was considering buying.
You have hit the nail on the head but it doesn't even need to be wifi and it also doesn't have to be that complicated. They see that two devices are in the same area for a long period of time so they serve ads for other peoples web history. Jarringly enough ever since I had this realization I sometimes see ads for things that reveal something meant to be private of those I had just spent time with. Also, seeing peoples ads when they make desktop recordings of their screen can be extremely telling..
Exactly. Years ago, I did due diligence on a company that targets ads using IP address. They buy IP address data from ISPs, target ads based on demographics, and then use cookies to retarget. Not that far off of what you're describing.
[flagged]
2 replies →
I know my iPhone isn’t listening to me. And I know about my friend’s activity influencing the ads I get served, and my demographic, and location, and all of that. And my random idea for a test word being predictable in a shocking way.
But, recently I started thinking about the average user, who will install anything and approve any permissions requested without reading it. And imperfect App Store reviews approving a Trojan horse accidentally.
Am I positive someone hasn’t inadvertently allowed mic access to a malicious party? I wonder if that person’s phone may, in fact, be listening to them.
Ugh how long until a new freemium model comes out where you can either pay with money or pay with microphone access…
using multiple high value advertising targets would be best. like specific brands of mattresses or industrial equipment.
No, they deliberately chose a topic they had absolutely no interest in, to try and avoid confirmation bias. It’s not impossible that what you describe is actually what happened to an extent though, a lot of the recommendations and ads on FB do seem to have a “what people around you / in your network like” factor.
You also need to do the opposite experiment Have two people put their phones in the microwave (don't run it!) / turn them off, discuss swimming pools (or jacuzzis etc), and see if you suddenly start seeing ads for said thing. And then you would have to repeat this experiment several times to rule out outliers
Experiment design is important! I completely believe that this happened to your friends and I also don't think it means what you/they think it does
(That is: you need to completely isolate yourself; music practice room on a college campus where nobody is wearing a watch or phone and repeat the experiment. If it turns out that you still see ads for that thing, then the experiment didn't prove anything)
Confirmation bias makes it hard to extract much from these types of anecdotes. On a daily basis you might be talking about dozens of products. If your lookback period is a few days, that could 100s of products, and you'll get spooky coincidences pop up from time to time from pure chance alone.
And, if not you, your friends and family. “I know a guy who had it happen to him” is almost as bad for confirmation bias as “it happened to me”.
Technilogical causes are much more likely than accidental causes for such effects to appear, in today's world.
Occam's Razor and the answer to the question, "What kinds of companies are at work in the environment?" push that probability in a specific way, because the motives and means are definitely there. Do you think they are the kinds of companies that would waste such an opportunity?
Their Chief Councel's recommendation depends on how slimy they are, right?
What would happen if they got caught? Slap on the wrist would be all, if that, no?
>Technilogical causes are much more likely than accidental causes for such effects to appear, in today's world.
This is absurd. The chances of rolling snake eyes twice in a row is 0.07%. However, that doesn't mean if I do get snake eyes back to back, I should think it's caused by "Technilogical causes" (aliens? CIA remotely controlling the dice?). At best, it's an incomplete argument. The power of the birthday paradox, along with the factors I explained in my previous comment means such occurrences are virtually guaranteed to occur if you're on the look out for them. This can't be dismissed with an off-hand with "Technilogical causes are much more likely than accidental causes for such effects to appear, in today's world".
>Occam's Razor and the answer to the question, "What kinds of companies are at work in the environment?" push that probability in a specific way, because the motives and means are definitely there. Do you think they are the kinds of companies that would waste such an opportunity?
Apple got sued for accidentally recording siri queries, and that cost them class action lawsuit, along with the requisite discovery. Some company intentionally doing this, all the while actively engaging in a conspiracy is far harder, and much easier to fall apart due.
9 replies →
Why can't you both be right? If you talk about 100 products in a week, chances are you're conducting searches about some of them or your demographic data suggests that you might be in market for it.
1 reply →
Something similar happened to me with backpack zippers. It convinced me the phone is listening and serving me ads despite the technical explanations that it isn't.
I was walking to work and my backpack zipper broke getting off the elevator. When I got to my cube I set my phone on the desk and said to my coworker, "damn, my backpack zipper just broke!" 45 minutes later I was in a meeting and checked my phone and backpack zipper ads appear. I had never googled backpack zippers before, never seen backpack zipper ads. Literally the only proceeding thing before getting these was was telling my coworker that mine had just broken.
Maybe your coworker searched for backpack zipper?
Or the ads were there before but ignored. Similar to the number 23 thing…
But this is just selection bias. If a hundred people do that and one gets an ad, it’s proof. Nevermind the 99 others who never saw a thing and wouldn’t bother posting.
The only way to test this would be to have your anecdote together with the complete marketing profiles of your friend and his wife. If such a profile could even be compiled in principle, from it we would be able to tell whether your friend or his wife had generated any non-audio pool-related signals, or whether they had seen other pool-related ads recently. Also, it'd be nice to know how often people in their marketing categories receive ads for pool fencing. Could be an astonishing coincidence.
It’s definitely a difficult one to test in a scientific way. But they 100% had no interest in pool fencing, living long-term in a rental townhouse. They chose the phrase specifically to be something they had no interest or search history in.
> They chose the phrase specifically to be something they had no interest or search history in.
If you feel like repeating this experiment, try generating the phrase randomly rather than choosing one, to eliminate unconscious bias.
The most striking example happened to me while watching a documentary about siberian cats.
We were watching it in Italian, our main language, and I wanted to know more about it, as I typed "g", the first result was "gatto siberiano", exactly the cat I was looking for. Way too specific.
Another time as my girlfriend said she was interested how much a specific model of a watch a friend of him costed, the very same happened, as I typed the first few letters the very watch brand and model appeared.
Since then, I just don't care about how much technical description I can read, nothing's gonna convince me of it being a coincidence.
There’s a breadth of possibilities between coincidence and the phone actively listening to everything via a microphone.
I’m fascinated that this urban legend persists among tech people because it’s so easy to disprove.
Did you know that you can set up a proxy from your phone and capture all traffic from it? It would be so trivial to find the traffic from your phone. There are ways to MITM and inspect the traffic, too.
There are also many people doing static reverse engineering of phone apps looking for security vulnerabilities. To believe this urban legend, you’d also have to believe that none of them have ever encountered this hidden voice analysis code.
If we ignore that, you know there are OS-level security controls on apps, right? iOS and Android don’t make it easy for apps to use the microphone constantly and run in the background to process it.
Finally, if we ignore all of that, how can anyone believe that these companies are recording conversations but none of their employees have ever chosen to blow the whistle? We’ve seen numerous FAANG “whistleblowers” come through with everything down to trivial or baseless complaints, but nobody has blown the whistle on these supposed widespread spying programs?
The whole urban legend is preposterous to anyone who has any experience with apps or phone security, let alone common traffic analysis or reverse engineering tools. I don’t understand why the myth is so persistent among even some technical people.
I'm not sure if the legend is true or not. But this argument doesn't really disprove it. The devices don't need to send full audio recordings. They are powerful enough these days that they can do a cheap on-device audio analysis and tagging, and then upload the (very small) tags. It doesn't need to be Siri quality analysis because it doesn't matter if the analysis is incomplete or sometimes inaccurate. They would just be scanning for certain keywords.
As for whistleblowing... Is there really that much to whistleblow about it? We already know that ad-based companies like Google are collecting our data every chance they get, because they make billions of dollars from it. They're scraping our emails, studying our GPS location, paying attention to who we are in proximity with, etc. The level of surveillance is incredible and people don't really care. It wouldn't be headline news to find out that they are taking advantage of yet another side channel.
>Did you know that you can set up a proxy from your phone and capture all traffic from it?
The phone knows about your proxy. There are phones - actual brands - that were caught on sending secret telemetry to their manufacturer, but only when not listened - definitely only on mobile data, no wifi, and I assume with cert pinning.
I know a person who was researching this and they needed a Faraday cage and a BTS to conduct experiments. So it's not exactly trivial.
The difference is that these were small Chinese brands that were not even that popular in my country - and still someone researched this. Imagine how much research Android and Iphone get, and there's not a single proof of and wrongdoing. Now that is unlikely.
This is just flat earth for technophiles. They don't really want to know the truth, they just want to enjoy their fantasy of living a conspiracy theory.
It is interesting how people always come up with anecdotes like this but none of them try repeating the experiment multiple times.
You might think the pool fencing example might be an extreme coincidence, but far weirder things happen every day. And what made your friend consider pool fencing as an example if they don't like pools? Maybe something they saw recently gave them the idea? Hmm...
Had this happen. “Airport tier tar” was the phrase someone said near me. Saw ads on Instagram for such a niche thing the next morning. Not only did I see ads they were insanely local. I have never needed to buy tar.
Then theres the time a friend told me about a very specific brand of Ramen, I opened up Facebook, and there it was, very first ad.
There is a video of Zuck denying they "recording peoples microphone" -- but how he said it with a smirk I took him to mean "we do on-device transcription and only send back keywords"
You see thousands of ads of every type every day and ignore them. Now you’re doing a test and consciously looking for ads related to pools. Of course you’re going to find something.
its called a noise gate: basic audio gear that triggers a function based on SPL(sound pressure level),which would be a reliable way to trigger a capture event and....the rest, without listening to everything.Change in tempo and pitch could also be good for an "event trigger". the start of digital audio goes back to the 1980's and the full suite of capabilities is trivial for any phone, as they are integrated extensivly to cancel background noise anyway. And with so called digital voice assistants running, I cant be surprised. My main point would be that ,everyone is convinced that there phones are spying on them, its one more thing to make them flinch and grimace, argueing about it will only draw deeper lines.And that, is where we are.
Now tell it over a campfire and maybe shine a flashlight under your face.
The party told you to recount the evidence of your eyes and ears as a campfire ghost story. It was their final, most essential command.
So...... the listening isn't very good? Because recommending a swimming pool simply based on the single word pool is just terrible.
Either they have the most technically impressive spying system that can't do anything right or it's just not happening and people are making connections where there isn't really any.
I’m not sure what point you’re attempting to make here, but they chose the phrase “pool fencing” and were rapidly inundated with ads for pool fencing, which, in isolation, would suggest the listening is extremely accurate.
we haven't done an experiment like that but I've had family know what medium certain topics were expressed and discussed over and those topics that landed in the ads that startled with their topic intersection were topics discussed only verbally vs typed into some search field or connected to some other web interaction etc.
That's kind of the smoking gun when you can create a disjoint set of topics and a disjoint set of mediums of communication delivery and see what shows up in the ad space from those discussion topics strictly expressed verbally.
But why did he pick pools? What if he lives in an upper middle class suburban neighborhood where everybody has pools? And what if he slowed his scrolling just a little too much on a pool ad on Instagram? What if he actually, kinda does think about getting a pool?
Who knows.
I'm just saying, the technical, ethical, and legal implications of creating an ad network that surreptitiously slurps up audio 24 hours a day in violation of the claimed terms of service without anybody leaking anything about it is a conspiracy that seems less likely than people just being more predictable than they would like to believe.
I have similar anecdotes about this and Yamaha guitars.
Whatever made him use pool fencing as his random example is probably also why the ad showed up. Maybe it's the season for that stuff, he saw other ads earlier, or other friends talked about it. He may not consciously remember that, but it could make him more likely to think of it again later. In other words he talked about it because of the ads, not the other way around.
Sounds convincing!
...unless there were actually several thousand people who performed this experiment, got a negative result, and therefore don't remember it or post anything about it.
If true then it's easy enough to produce video evidence of this.
> Every time I read some technical description about why this isn't happening, the technical description seems convincing.
Having knowledge of the technical limitations and challenges myself, I used to be on board for this argument, but now less so.
All of the technical arguments against the listening seem to ignore the "Ok, <DEVICE>" or "Hello, <DEVICE>" initiating phrases for the voluntary surveillance devices people put in their rooms, and offer only a worst case defense ~"how could they process everything everyone is saying?!"
Why is it such a stretch to imagine these devices grab Direct Objects and Subjects and store those singular items for ad keywording?
We have cookies and know how they work, why is it difficult to extrapolate?
simonw is a breathless proselytizer of LLMs and likely is suffering from "a man's salary depending on misunderstanding" and all that.
It bears repeating, "these corpos are raising billions and hiring former alphabet heads to their boards for reasons other than just making you a better programming assistant."
> Why is it such a stretch to imagine these devices grab Direct Objects and Subjects and store those singular items for ad keywording?
How about because Apple say they don't do that, and can and do get sued if they say things like that which are not true?
(Sadly I make basically no money at all from my "breathless proselytizing" of LLMs. I hope to fix that this year, someone should pay me for this stuff!
You know I've written more negative things about LLMs than almost anyone else, right? 121 posts tagged AI ethics right here: https://simonwillison.net/tags/ai+ethics/ )
> the technical description seems convincing.
Isn't the technical hurdle just changing the "Alexa" or "Siri" wake word into a keyword for an ad campaign?
Siri doesn't let you do that, it doesn't give you audio it does pick up, and it doesn't even listen for the wake word with the main CPU.
I assume here you're referring to the user?
It seems like users can have iphone listen to audio keywords, but I don't use Apple so I don't know the details. https://support.apple.com/guide/iphone/use-vocal-shortcuts-i...
Obviously an experimental advertising system would require a special deal with the phone manufacturer and would be covered by Apple's very broad license to advertise based on your data.
That doesn't mean they're doing it, of course. I'm just not aware of technical or legal barriers.
A family member who seems intelligent and sane occasionally tells me stories of her experiences with ghosts.
...and yet no one seems to ever be able to replicate these results.