Stop Listening Please, Google | Hacker News Reader

Stop Listening Please, Google

randomivysaur.bearblog.dev

23 pointsmemorable29 comments

A while back (I believe May 2020) I did an experiment. I kept a phone in isolation for 3 days and recorded all the network traffic. Then I took the same phone and put it up against a speaker that played a podcast archive continuously over a 3 day period. Then I did another 3 days of silence and a different archive of podcasts for 3 days. At the end I had 24 days of phone network pcaps (4 sets of 3 days audio + 3 days silence).

I made sure that the phone was able to transcribe the audio coming from the speakers when I put the keyboard in dictation mode just to make sure that it could "hear" the conversations from the speaker.

I tried hard to find statistical anomalies or pattern differences in the sets but I was unable to extract anything that could yield any type of change detection corresponding to when they happened (silence to podcast or podcast to silence) or anything that could differentiate the content. I had timestamps, hoping to find spikes, say, during commercials where a bunch of brand names get mentioned. There were no such spikes and the traffic was frustratingly uniform and appeared to be completely independent of the experiment I was running.

The phone had google assistant, facebook, instagram, tiktok, and snapchat installed among other applications. All the accounts were logged in and I occasionally adb'd over to the device to make sure the applications appeared in memory in a ps list.

Now of course, absence of evidence doesn't show evidence of absence and maybe I am just too incompetent to see what's going on but I was unable to find any hard evidence using this method.

I of course was really hoping for the opposite and to come forth with some industry shattering announcement, but if there's something there I at least wasn't able to find it in the wire.

I encourage others to repeat this and refute my inconclusive results - that'd be great. It's an easy enough experiment to do. Happy hunting.

It is possible, at least theoretically, that "keyword scanning" only happens when you interact with the apps.

Say, you're searching for something while you speak, lo and behold Google Search's suggestion list pops up and is extremely accurate for what I am about to search based on 1 or 2 characters.

Yes, I thought of that as well but I didn't pursue it. There's ways to MITM the traffic on your own devices if you set up your network and phone right. That'd probably be the next step and then to look into the payload contents (which should be a bit more conclusive). I was just doing flow analysis.

Doing that and getting a hard "no" would be valuable. Alternatively if you can catch the hypothesized spying on the wire, fame will be yours.

> It is possible, at least theoretically, that "keyword scanning" only happens when you interact with the apps.

This isn’t what the popular conspiracy is though. People argue just having their phone with them while they hold normal conversations leads to advertisements later. No app interaction is ever mentioned. I think this experiment is perfect for debunking the conspiracy.

https://en.wikipedia.org/wiki/Straw_man

This was a great experiment. It obviously doesn't rule anything out or confirm anything, but one possibility is that the apps cache this information and only send it out during their regular communication frequency as part of the packet payload.

I'm not sure if there's a way to test for this (or if you tried in your experiment)

There are two aspects that discredit these types of claims: The banal explanation, and the extreme risk Google would be taking by doing this.

The banal explanation is that life is really complicated, and a coincidence of this nature is unimpressive. This could be as subtle as a video you watched said something that was worded similarly to lyrics in one of their songs that stuck around and made you think of it a day later. It doesn't take many people to later search for that song for the algorithms to pick up on it and start suggesting it.

For the risk, Google has had a history of a very open culture. Many people within the company have strong ethical convictions. To do a project like this, you'd have to selectively recruit a sub-selection of employees who don't have these convictions, and manage to provide data streams to other teams algorithms. This is without anyone noticing that, hey, that team sure is using a lot of ML resources or my device sure is uploading a lot of data. This is all on top of the legal risk of saying you're not listening while also doing so. I can't imagine that getting past a legal review.

I think there are a lot of good examples of big companies abusing their power, but this isn't one of them.

I suppose the impeccable ethics of google employees are the reason google is such a reliably ethical company? Uh...

So, google is not that, and has not been that for a very long time. Yes, this case is probably a coincidence, maybe it isn't, but the idea that google engineers would rise up for spying a teensy but more than they already do is just not reality. Alphabet employees aren't ethicists in a position to protect people from the inertia of a trillion dollar company and it is outright crazy to suggest that's the case.

My point was perhaps lost in the pursuit of brevity. It's not that the ethical people could stop Google from doing bad things. It's that there's bound to be people outspoken in the event that Google is secretly spying on you. It's the kind of thing you'd see a leak making it to the front page of New York Times.

Like, Google monitors what you do. A lot. However they also are pretty clear about how they spy on you. There are cases where they misstep (like the recent tracking settlement where they're forced to be more clear about location tracking), but a secret conspiracy to record everything you do with your devices would take a large team and serious resources.

Yes, the world is complicated and there are numerous simpler explanations.

The fact that a few (plural) of their friends were talking about X, means that they probably had a recent contact with X (searched for it, watched the videos, etc). There are anedoctal cases of shared recommendations by location/local network IIUC.

Could easily be a correlation thing, rather than a causation thing.

It would not be hard to believe that the friends were talking about the songs because they're trending as "90's throwback" or whatever, and then YouTube suggested those songs because they're trending.

I think this is probably the case. Humans have a terrible memory for things.

As a counterpoint: I was looking at some hair clay of a specific brand I'd been thinking about getting. When I checked my Facebook account, lo and behold, they were running ads for that product. Not at all surprised since their ad network tendrils are probably into everything, and I was deliberately looking.

A while back I was casually driving home in my car in the wee-hours of the night. This is when I also had an XM subscription in my car so it was that long ago, well before Spotify/Pandora where one can listen to whatever you want at the moment. Anyway, I was browsing all the XM channels and for some reason "The Flame" by Cheap Trick popped into my head and I was in the mood for it. While channel surfing I came across the unmistakable guitar chords of the song just beginning. It was super eerie and to this day an unforgettable coincidence.

"Stop reading my thoughts Please, XM"

The point is - in our long lives, stuff like this happens through shear probability alone several times. An advanced search algorithm designed to predict songs based on past history, and who knows how many other available inputs, isn't even in the same realm of weirdness as feeling a song and it appears on the radio.

The creepy thing is not that Google is listening to you. (They're not. Period.) It's that they can get so good that they can convince a large segment of the population that they _are_ listening to you.

As others have said, correlating your friends' activity with your colocation and other similar habits is enough.

This. 1000 x this. Lets's separate the "how" from the "whether", and consider the outcomes.

I've heard this anecdote from so many people in so many different contexts that I'm beginning to get suspicious about the common "benign" explanations like graphs of friends, their interests, location tracking, etc.

Do we an in-depth analysis done somewhere that proves / disproves the theory of an actual microphone / STT being used either in Android / iOS itself or by a running app?

There's also the benign explanation of "coincidence".

Yep. How many people talked about a song they don't like, then signed into youtube and saw... not that song, so they didn't post about it.

>Do we an in-depth analysis done somewhere that proves / disproves the theory of an actual microphone / STT being used either in Android / iOS itself or by a running app?

For mobile, actual voice processing (i.e., not just hotword detection) would have some kind of telltale higher-than-baseline resource consumption. Whether CPU, disk, memory, battery, or network. Processing voice has to happen somewhere, and it takes resources to either do it on-phone or ship it off elsewhere (either immediately or later).

I think the friend graph explains most of these away, along with coincidences.

If your Spotify friends listen to a specific song, it kinda makes sense for them to recommend it to you as well. They may have also given their friend "the aux" at some point which would train the algorithm on their music, so it's just bringing those songs back up (this one happens to me _a lot_)

Can anyone (perhaps a former employee) confirm this behavior. I can’t tell if I’m imaging connections or the words spoken are being ingested by the ad machine.

It seems like a huge overreach to be listening, the microphone privilege I could see being abused for this, but it just seems so wrong.

Yes, it happens way too often to be a coincidence. I talk about different topics with friends, they end up on my Youtube feed, not sure how. But I use iphone and a mac though.

Google can reasonably infer people that live together based on billing addresses, location, etc. If your roommates pay particular attention to certain media (a video, a song, etc), it's more likely to show up on your recommended media.

In this post, chances are that Google classed each user as being socially close and based its recommendations on that. It's uncanny, but with how much Google data has, eavesdropping is unnecessary for this use case.

A couple of the most popular hits from 30 years ago. Are you sure you didn’t hear them on some advert recently? Seems less likely that Google discovered your obscure love for Nirvana from a conversation you had.

They have enough data to model you in the machine, together with whoever you are talking to. Just like was shown in black mirror or westworld. No need for listening, you already live in their server.

I like to refer to this system that is always listening as The Daemon because I like to assume it has a commen backend. Also lookup the book Daemon by Daniel Suarez it is pretty great

Do you have Hotword detection or Google home in the house. Sometimes they trigger for wrong keyword. Could it be that they got triggered and added these songs?

Realtime collaborative filtering using proximity transmission. :B

I read articles like this and then understand why the average person thinks that Bill Gates implanted a chip in them via the covid vaccine.

To anyone who has worked on mobile devices or ad tech, this sounds as silly as your uncle telling you about the covid chip and how vaccines cause heart attacks.

Reply on news.ycombinator.com