I visited a friend who is watching Breaking Bad. Now clips appear on my YouTube?

55 points by SnowHill9902 ↗ HN
We watched an episode and now clips are showing as recommended on YouTube. How does it work?

(Edit 16:48Z) More info: - using iPhone - YouTube on Safari, no app - got connected to her wifi

130 comments

[ 3.3 ms ] story [ 188 ms ] thread
I went to a wedding and chatted with someone for 5 minutes and the next day they were a suggested friend on Facebook. Never seen that person before in my life.
Was the wedding a Facebook event that you RSVPed for?
I don't even know how to RSVP for a Facebook event.
This one is easy - Facebook recommends people based on geolocation. It's very common to get friend suggestions to people you've met, or just been in the same building with.
Now the problem is how Facebook got your geolocation data.
(comment deleted)
Geolocation data from the phones could have been used to show you both were in about the same place at about the same time.
I believe Facebook recommendations are somewhat bidirectional - if they visited your Facebook page (by searching you, even if they didn't add you), I think Facebook will recommend that you add them.
I've got a million of these, but one of the weirdest ones was when a friend of mine suggested we go to one of the islands northwest of Seattle for the weekend. The next day, I was googling "things to do in..." and it autocompleted to the island's name, even though it's a relative backwater compared to some of the other places around. Best I can figure is he googled it, they saw us in proximity to each other, then predicted that we were planning to go there.
Happened to me as well. I suspect that your Android phone noticed his Wifi network, mapped it to his Google id, and now yours and his YT accounts are partially sharing recomendations. That's the least pernicious mechanism I can come up with at least.
He used iPhone
Then it's more suspicious. Perhaps Google apps with access to localization (such as Maps) could still do the same thing, on an iPhone?
I've had similar things happen when all I did was discuss something with my brother on a phone call, and next thing you know I'm getting ads for it.

I'm willingly to believe it was coincidence, but I am suspicious.

Pure guessing but your phoned showed your location and what you were watching (through your friends data), and thus recommends Breaking Bad to you now.
Not even location, just OP's phone discovered that a router with a unique MAC that just LOVES streaming BB is sometimes nearby. Google's "AI" dictates that means OP would absolutely love BB too, it might be flawed logic but it probably works
I had a 5 minute conversation with my niece at the park about rotary phones and now my feed is full of rotary phone-related articles. This happens more times than I can remember.
Your niece searched for rotary phones?
That's a good theory. I'm not on any social networks, however I do email her, so perhaps Google associates me to her via my contacts list?
The implication is it listens to you? Or she searched rotary phones after you spent an hour or two in the same places?
My gut reaction is that it listens to me, either directly or via some app, but I know Google has refuted this fact. I suppose this is testable too, so I tend to fall on the confirmation bias side of things.
The conversation didn't prompt you to look up any rotary-phone-related content?
I am confident it did not, that conversation was out of my mind as soon as it ended, which was why I was so surprised to see it dredged up again in my feed.
I can't help but think these are confirmation bias events (if the clips were for "Battlestar Galactica" and not BB you wouldn't be writing about it), but similar things happened to me too:

- I walked by an e-bike store, stopped to look at the model displayed on the window, a few days later I got ads for e-bikes on Instagram (I can't remember now if it's for the particular store's brand).

- I opened Instagram at a carwash, a few days later, ads for car-detailing services - here I have a strong feeling it used my location data.

Location. Bike store paid for ads in the city (or county)
You can't be paranoid enough to imagine the disgusting amount of location data being slurped up by unscrupulous apps and traded for any purpose at all. Targeting ads based on having been near or in a store is completely standard.

https://blog.citizennet.com/how-to-measure-store-visits-on-f...

> By using location services on cell phones and 3rd party satellite imagery and mapping data, Facebook is able to tell you if someone visited your store within 28 days of clicking on your ad, all while filtering out employees or people who move past your stores without going in.

> Facebook can't measure all store visits. This is because some people don't have location services turned on in their phone, or are not recognizable by 3rd parties.

https://www.placeiq.com/audiences/

> Reach audiences who visit your location on a regular basis

> Reach audiences who regularly visit your competitor’s locations

> Message audiences who regularly commute past your location

Have you used his Chromecast or connected to his wifi? I’ve noticed lots of misdirected targeted ads after doing this
Your phone is now associated with his IP address/location now. Targeting individual is hard. But targeting people at a location/up is not.

Data is usually refreshed monthly. He may also see ads for things you search for.

Location tracking - they don’t use the mic as many suspect, but definitely know where you were and who you were with, and presumably what you were watching if it was google-linked. Perhaps there’s Bluetooth/wifi mechanisms in there too to make it easier but it’s the same idea. That said I don’t see why they’d do that for YouTube recommended… maybe coincidence
>they don’t use the mic as many suspect

Unfortunately, there's no way to confirm that since the only phone out there with a hardware microphone switch is the Pinephone.

You would expect it to consume quite a lot of power and/or bandwidth if the mic was always on.
You can try fingerprinting, etc. Ridiculous amount of effort to implement it. For little benefit.
They do it for music identification these days.
which is such a halfhearted excuse too- "why are you randomly sampling audio around me and phoning home about it?"

  "uhh.... shazaam?"
I thought that was all done on-device?
I'm not sure. Just pointing out that it doesn't necessarily have to use a lot of power or bandwidth to have the mic always on.
But isn't it always on for "Hey Siri"?
Imagine the computational resources needed to track microphone for all Android users.
We actually can by disassembling the YouTube.ipa, examining the calls, and checking for AVAudioSessionDataSourceDescription objects. Tracing their lifetimes.

It’s the same method Apple uses to check for “unauthorized internal API use” people get busted for.

We can also replace the AVAudioSession class with our own proxy to examine calls.

You'd have to disassemble Google Play Services, not the Youtube app.
OP said he/she was on iOS.
And not using the app at all, just browsers
Ah right, missed that part. In this case it’s rather easy as Safari will respect your settings. Just deny microphone access. Under settings->safari->Microphone which is down under “settings for websites”
it just makes it less likely that a microphone was used instead of other information such as IP and location.
>they don’t use the mic as many suspect

I know they claim they don't, but do you have reliable information that confirms this? Like experience on an internal team at Google/Apple?

Honestly I would love to see it be the norm, maybe even law, that a recommendation is accompanied by a reason for the recommendation. Is this because I watched a similar video? Or is it because you snooped on my microphone when I was at the local cafe?
This might not be possible - there are many inputs into the model; it would be impossible to trace which one tipped the scale.

Better to make listening on the mic illegal, punishable by prison time for the execs.

> This might not be possible - there are many inputs into the model; it would be impossible to trace which one tipped the scale.

I am not a fan of problem-solving via legislation, but I wouldn't lose a lot of sleep if Big Tech was no longer allowed to burn billions of CPU cycles on making opaque, inexplicable guesses as to what ads are most likely to make them money if shoved in front of my eyeballs.

I get the annoyance, but also, targeted ads means both a more efficient economy and information environment.

Mass ads means more money goes into ads and ads are irrelevant to many people. Only big co’s can afford mass ads. Targeted ads means ad money is vastly more efficient, and your information environment is less cluttered. One-man shops can afford targeted ads.

I like seeing relevant ads. I like supporting small businesses targeting me. Privacy isn’t a concern either because the three-letter agencies can find out whatever they want about me if they want anyway.

FWIW, I’m not in advertising.

All the more reason for transparency if no one can explain how it works. Ideally "my model is too opaque" isn't an excuse, but an admission of negligence.
They'll simply say, its not 'listening' if an algorithm does it.
I love machine learning, data, and even targeted ads. But no marketing algorithm should have access to mic and camera data, period.
Fun tangent, I recently went to Rally's (a really hard to find burger joint around my parts). I rarely go so its always a treat with their shake/glorious fries.

Found they've completely replaced the order taker with a robot, with order signage saying "help train our robot".

No opt out, and you can't get to an employee unless it recognizes the word employee. Recording in progress.

The ADA non-compliant parts aside for people with voice related disabilities, its BS what companies are forcing people to do just to buy a burger. I'm not going back after that.

You just put in a pile of all the data you know about the user and session into a pile of linear algebra (some ML-based recommendation system or whatever), and recommendations come out. In other words, I doubt they even know the reason for any particular one.
In systems complicated enough there could be no clear answer
Then should we really be building systems that complicated?
Because it is harmless. The worst that can happen is a bad recommendation.
Or the hijacking of attention by unwittingly flooding the streams with "engaging" content that is politically divisive and spreads misinformation.

Maybe you left that out because it already happened.

Even fairly simple systems might not have a clear way to answer. Consider a movie recommendation system.

One simple way you could build such a system is to come up with a list of things about movies that might affect whether or not someone would like them, where for each thing on the list we can assign each moved a number from -1 to 1 that says how much of that thing the movie has. Call this list the movie's vector.

Some examples of things we might pick are how much comedy is in the movie, how much romance is in the movie, presence of A-list stars, how musical it is, and thing like that. We might also have items for specific stars or directors.

Then we could go through our movie collection and have someone figure out each movies scores for all those things in our list.

Then we could figure out for our users a list that lists for each of those things how important it is to that user, from -1 (I hate movies that have this!) to 1 (I love movies that have this!). Let's call this the user preferences vector. If we have a list for a given user of movies they have already watched and how they rated them on say a 0 to 5 scale then it is some straightforward math to figure out the user preferences vector for that user that does the best job classifying the movies they have already seen in a way that agrees well with that user's ratings.

That user preferences vector can than be used to recommend new movies and should work pretty well if (1) we picked a good list of things to score movies on, and (2) when we manually assigned the scores we got it right.

To predict how well a user would like a given move we just take their user preference vector and compute the dot product of it with the movie's vector. The more positive that result the more we think the user would like the movie.

With this system, it would be easy to tell someone why you recommended a movie. We could look at their preference vector and compare it to the movie and tell them things they really like that the movie has and things they really hate that the movie does not have.

But the system described above has a drawback. It is hard to figure out what factors to include in the movie classification. Should comedy for example be one item, or should it be broken down into several such as physical comedy, insult comedy, bodily function comedy, and so on?

Also, if you have a large collection of movies it is a lot of work to go through them all and score them on each factor. And if you later find out you need to add or remove factors you have to do it again.

It turns out that there is a way to sidestep both the "what should my factors include?" and "how do we get the factors scored?" problems.

What you do is just decide on how many factors you will have. So let's say we decide we are going to have 50 factors. We don't have to decide what they mean. We'll just call then F1, F2, F3, ... for now.

Initially we just assign each factor a random value from -1 to 1.

We also do the same thing for the initial user preference vectors. Just assign each factor in the preference vector a random value.

Then we can do a loop, consisting of these two steps:

1. Using the known 0 to 5 star ratings from users of films they have seen, adjust their preference vectors so that ordering movies by the dot products of the movie vectors with the preferences vector matches the ordering by the user's star ratings.

2. Same thing, except instead of adjusting the preference vector to better work with all the movies a user has seen, adjust the movie vector of each movie to better work with the preference vectors or all the users have have rated that movie.

Keep looping until things aren't changing much. You then end up with a set of movie vectors for you movie catalog and preference vectors for your users that do a good job of ordering movies that user has seen that matches well with how the user rated those movies, and that likely does a good job predicting how well they will like new movies.

This ...

Story time: I have disabled all types of tracking in Twitter, same for Amazon. Now, I bought a few e-books for my wife. Those are very girly books that definitely don’t fit my profile.

Twitter is now showing me kindle ads for those books.

Twitter has a menu offering to explain why I see it. The explanation says: Amazon wanted to advertise this to people in your region.

I say they lie… We need Blackbox monitoring against those behaviors with some legal teeth.

I sometimes wonder if phones are actively recording audio, but more likely these companies are all sharing data or some app on your phone is doing it. No matter how much I change my privacy settings to the max, I still have the same things happen to me where it is extremely unsettling.
(comment deleted)
This is pure speculation, in descending order of probability:

- Random coincidence, which you notice in the one case where it happens but not in the 99 cases where it does not.

- You did search for it or something related. Or, you got one clip due to one of the other reasons, clicked it, now you're getting more.

- Correlation via IP

- Correlation via location

Since we're talking about YouTube recommendations, not ads, I kinda doubt the last two though. That would provide very little benefit and be a huge privacy risk. Location is certainly considered to some extent, but I would expect this to be on a country/region level, not city and certainly not fine enough for your friend to meaningfully influence it.

Search history including Google search history, I’ve found.

I also completely believe that YouTube correlates via IP for at least not-logged-in views (or at least tries to associate to accounts even if they've never logged in); I get bleedover to my iPad from my completely disassociated PC but not my Mac that’s logged into a different account.

(Also maybe I put too much effort into tailoring my own YouTube recommendations, but 99% of the time when they start going awry, I have a pretty good idea what triggered it. Random coincidences don't happen...)

Like others have commented, I too am very certain Youtube uses IP address. My family and I are behind the same cable modem, and oftentimes my family's Youtube "related videos" will start popping up in my Youtube home page, and vice versa, even though I have no interest in or have watched those videos.

Interestingly, I seem to recollect this happening only in the past few months or so, maybe a feature they turned on recently?

I think if the IP is detected as used by one person the recommendation engines relies on it more heavily. For example in an office the same IP is used by many so its weight in the recommendation engine is relatively low. Additionally more and more ISPs are using longer lasting DHCP leases. I know my Comcast IP hasn't changed for over a year. This can better tie a person to the same IP.
There are 6 people using my ip regularly. We all get a mishmash of eachother's YouTube recommendations.

I'm frankly surprised to find some people in this thread don't already know that ip correlation is going on as a matter of course.

>Random coincidence, which you notice in the one case where it happens but not in the 99 cases where it does not.

In addition, there could have been 100 instances in the past but they didn't stick.

> I kinda doubt the last two though. That would provide very little benefit and be a huge privacy risk.

To the contrary, I highly doubt that they don't use it, privacy risks be damned. Using that data is their primary business model after all.

I’ve had enough “coincidences” of the same type that I 100% believe it’s occurring. It’s the simplest answer actually.
Occam's razor I agree.

Easy enough to test, just start talking with another person for about 2-5 minutes straight repeating the words purple dog collar over and over in earshot of the device you suspect.

You'll be amazed when within 30 minutes to an hour all the ads on all devices will suddenly show purple dog collars or pet related items.

Really? I info-dump on my partner all the time about all sorts of specific products. She’s never mentioned one randomly pop up in an ad. I can honestly say I’ve performed your test thousands of times.
As a counter-experiment, try thinking of some product and intentionally DON'T search for it or say or type it anywhere. Make sure it's also not something you heard about recently from a friend or in the media. If you then see ads appear for that product you can be sure it's just baader-meinhof.
Talking with another person won't test whether they're correlating via IP or location.
I would insert for 2nd place:

- The system knows you two are friends (there are multiple ways for that), and you get recommendations based on what interests your friends

They do correlation with twitch though. Went there once to support a charity stream, and I got flying sims and among us in recommend for weeks (or maybe only a week?)
I have found out that google uses atleast 1 of the last 2 "doubt" ones. My sister was watching my little pony on her laptop without being logged in and her reccomendations would "leak" into my reccomendations on different computer with logged in account. It was very annoying and even flagging the videos as "not interested" or "don't show" wouldn't help much. I think I "fixed" it by disabling yt history or something. I am not sure. This was 2...3 years ago
I wouldn't be surprised if it was content ID'd by your smart devices either by mic or by the accelerometer.
How would the accelerometer ID work?
Accelerometers can be used to detect speech, with a high enough resolution they essentially can be used like microphones. The audio data inferred doesn't need to be good enough for humans, just good enough for some AI algorithms to produce a perceptual hash.
Content ID'd by the accelerometer? How does this work?
It depends on your privacy settings, the local privacy laws in effect, and what you did over at your friend's place.

If you watched the episodes on a device connected to your Google account in any way I think the answer would be obvious.

Another reason you might be seeing ads for it is that you might've looked up an actor or character on your phone which Google linked back to the show.

It's also possible that Google noticed your account was suddenly shared by someone who showed interest in the show and "infected" your account.

Another possibility is that Google picked up on the show if you used a Chromecast on the same network to watch an episode; media controls are broadcast throughout the network.

There are more far fetched reasons. For example, if your location history is on and the neighbourhood where your friend lives is seeing a sudden spike in interest in Breaking Bad (watch parties? Idk), you may have been flagged as interested purely by geographical proximity.

It can also be a combination of Baader-Meinhof and your behaviour around the ad (looking at it longer, possibly unintentionally out of surprise) that reinforced the topic in your ad preferences.

Tl;dr, Google can learn a lot about you through indirect means. It's kind of their business model. Compare to the famous case where a supermarket knew someone in a US household was pregnant before the other family members did based on purchasing behaviour. It's scary how much you can tell about a person from their shopping patterns alone and Google has their hands in so much more.

None of this should probably be happening if you've disabled personalised ads (https://adssettings.google.com/), at least not on Google's platform.

I believe it's your phone's microphone
There was a private showing of the next episode of Better Call Saul yesterday at a film festival of sorts, the rest of us won't see that episode until July.

It's probably just trending topic and you fit a similar profile to people that would watch the show/Breaking Bad. My viewing habits are pretty stable and I noticed I get recommendations for The Wire even after resetting all my cookies as well as not logging in to my Google account

how did you watch the episode? many streaming services, devices, or smart TVs track what you're watching.

or more likely, you googled something breaking bad related while watching or talking about it.

or even more likely, it's a coincidence. my youtube recommendations also had a breaking bad clip in them yesterday, despite me not watching it, talking about it, or thinking about it. better call saul season ended recently, and people are searching for breaking bad clips. it doesn't have to be more complicated than that.

Just so you know, over the past couple of weeks I've been getting clips of Breaking Bad recommended to me - which I think it's something that's just trending and it's getting a lot of views.

I think it's random coincidence BB is getting traction now.

(Before BB was The Sopranos clips)

Location-based tracking. In this case, your phone was on your friend's wifi so they know whatever that wifi was doing while you were there, you were in the room.

Unless you've already completely de-googled your life, this is the most straightforward way to link you to the activity. All it would take is for you to have pinged Google's servers with that phone just once, ever, and they've got it fingerprinted. So when it shows up on the same network accessing YouTube, they placed you.

One use of Google Maps, Google Home, logging into Gmail or YouTube is all it takes to compromise a phone.

"got connected to her wifi" thats the reason you are getting the recommendation. Youtube stores the data of videos watched using wifi and then tries to recommend other devices connected on the wifi. Reason 2: have you talked anything about the show ? because it has been conspired and seen that google listens thru our microphone in devices and recommend ads even if you haven't searched for them.

tldr: Reason 1: Wifi cache. Reason 2: Microphone tracking.

Happened to me except even more far fetched:

I shared a car ride with a friend for 20 min, and my colleague, being knowledgeable in trivia, mentioned an obscure historical figure that I couldn't even spell the last name of. I came back home, and on my Google feed there he was, the obscure historical figure with his last name fully spelled out in a biography.

I never connected to my colleague's network in any form (we were in his car); I didn't search for this historical figure (I don't even know how to spell the name correctly). And most definitely not a coincidence as I had never heard of this person before or after the brief mentioning by my colleague. I don't think my colleague searched for this person on his phone recently either as it was some trivia knowledge he randomly recalled.

This happened several times with this particular colleague as he likes to mention random trivia. I have yet to find a plausible mechanism for this phenomenon (unless it is an open mic).

Could it be that your colleague just spent the morning googling these subjects, and Google notices a correlation between the stuff your colleague googles and the stuff you later look up?

e.g: it's using location-tracking to link you two, not an open mic and realtime transcription.

Sounds like certain videos about historical trivia start trending and your colleague watches them before you do.
Yeah, there must have been a reason the colleague came up with that obscure historical person, certainly not random.
My explanation would be fairly trivial: those videos were currently being pushed more broadly by the YT algorithm (propably based on some rough demographics that you two share - age, gender, location, broad interests...). You two probably saw a lot more of the same videos that you didn't talk about.

PS: my understanding is that a lot of social media sites use a strategy where they somewhat randomly push single posts by reliable creators, to make sure that those stay hooked as well. So that could be an explanation why you'd suddenly see something a bit more niche.

Indeed, this is my guess although it means I am having considering my colleague as an unreliable narrator (he is not "randomly recalling" as he claims. He is just regurgitating what he saw on his phone earlier in the day).

On the other hand, the obscurity of the feed subject makes me think YT/Google wouldn't be pushing those on a broader audience, but maybe the algorithm is just that good at nerdsniping us.

Hypothetically there is a common root source further back from your friend. IE a new book/documentary has people talking and the general buzz triggered both your friend to comment and Google independent of each other.
I listened to a podcast which mentioned a relatively obscure topic in passing using the basic samsung music app on my phone which didn't have a SIM card in, location turned on or an internet connection.

Later that day I see an advert for said obscure topic served on a web page.

Is it possible that there is someone transcribing podcasts or at least scraping databases of their RSS feeds and somehow my music player app is broadcasting that I've listened to a particular file(after receiving an internet connection)? The alternative is that the machines really are listening.

You would have had to at some point download the podcast, so at that point before they served the download to you, they would have used your IP to serve you a relevant ad based on that location data and inserted it into the podcast. A lot more podcasts are doing this unfortunately. I don't remember the name of major platform or company enabling this.
I’m not convinced YouTube is using such data for its recommendations engine. We are talking about the app’ that serves you ads in languages you don’t speak whenever you travel abroad.
Visit a friend who watches Breaking Bad? Advertise breaking bad.

Visit a country that speaks German? Advertise in German.

Seems consistent to me, not a contradiction.