Tell HN: YouTube has a spam problem, and it's getting worse

95 points by curiousgeorgio ↗ HN
I've noticed a ton of crypto scams on YouTube recently, many of which are using the names of legitimate channels (like SpaceX), re-posting content from those channels and promoting scams – usually in the form of "send me [Elon Musk, Jeff Bezoz, etc] BTC, and I'll send you 2x back".

It seems reasonable to expect one or two of these to slip through the cracks now and then, but over the past few months, I've had YouTube "recommend" these to me, again and again, multiple times per day. I've often reported them as spam, only to have YouTube recommend the exact same video to me a few minutes later. I don't know anyone who would fall for the scams in the videos, but they've become so common now that it's getting really annoying.

You can't tell me that they're too difficult to distinguish from legitimate content. Any human being or half-decent algorithm would be able to catch these with no trouble at all (heck, I could write a couple of regular expressions that would do the trick and have 0 false positives), which leads me to wonder: why is YouTube promoting these videos? At this point, it feels like they must have some incentive not to shut them down, because it would require deliberate effort to have any form of spam control while also somehow letting these through.

Am I the only one noticing this?

70 comments

[ 5.3 ms ] story [ 22.7 ms ] thread
I've noticed it happening right after any Elon related event
I tried watching 18650 battery reviews on yt for 2021, and almost all the video are text to speech fake videos that just read amazon text or some other blurbs. Prob all automated to make videos. All mostly voted down, but a way to common problem.
I've seen a lot of them too... "top 10 X", then 10 affiliate links to some crap, and text-to-speech reading the first sentance from the description on aliexpress/banggood/...
TTS videos are everywhere. Some of them are good enough you might not notice the voice at first, but the content is always some generic crap. If it sounds like a topic report by a 6th grader, there's a good chance it's auto-generated.

There are even channels with auto-generated pronunciation videos, because apparently TTS engines know how to pronounce things now. /s

the dead giveaway is any url and mention of bitcoin

I can see how ppl fall for these. look very realistic.

Unfortunately, YouTube seems to be trending away from letting people downvote stuff and having the vote mean anything. If I could see the upvotes and downvotes before viewing something, I'd know better than to view the spam.
Supposedly downvoting has the opposite effect of what you'd expect. My understanding is that now downvotes count as engagement, so bad videos with tons of downvotes will actually get boosted up. And of course comments will also have the same effect so you can't leave warnings there.
Controversial or outrageous content is likely to generate way more "engagement" (as people fight over it and call each other names in the comments) than normal, unobjectionable content. Social networks thrive thanks to this.
Litrally just witnessed a perfect example of this a few minutes ago: https://www.youtube.com/watch?v=aMtIpjERmOM

This channels first video ^ titled (the video is 1 week old): "Jacob Collier: Music's Biggest Fraud" Likes: 102 / Dislikes: 469 I think it supposed to be satire, but doesnt come off too well. He makes a decent point about overbearing fans. I was surprised to see it in my recommended, as I dont really know of Jacob and only watched this video once, which is just 33 seconds long with the title of "when your phone rings in a jacob collier concert"

https://www.youtube.com/watch?v=O8PB0EzJ4oA

So your point about engagement really rings true to me. Kinda sucks, especially when you can see the kind of... friction, it promotes (the comments section is pretty mad)

I was "enrolled" in the Youtube experiment where I could not see the downvotes of videos for something like 3 months. I'm pretty sure it's because I am a reaaaaaaly strong believer in dislinking the videos I did not enjoy, for any reason. I always like or dislike a video, there are no exceptions, and the like/dislike ratio is paramount to my decision to watch a video or not. My solution was simple: I created another Youtube account and resubscribed to all of my channels. I simply stopped using my old account. I don't know if it was because of that or not, but 3 months later my old account could see the dislikes again. F*ck Youtube and Google.
There are also tons of spam comments pushing dot fyi or dot xyz domains.

Why can they not just ban urls?

Also a lot that are just

"Yes <link to another youtube video> xyz"

I don't get the point.

Training the classifier that "xyz" is a sign of quality (by posting a benign, non-spam link) so that they can then use it for spam and have it go through?
That would be next level spam. But it sounds entirely possible.
I reported crypto scams targeting a spacex audience many times over the last week. Reported as misleading/scam and do not recommend channel.

Would refresh the front page and would get another scam in first position. Either Youtube is using me as unpaid labour to detect scams because its automated systems are rubbish or their systems are very broken. I wouldn't be at all surprised if they count reporting as user engagement and promote even more shit my way.

Having paid for premium/red/whatever youtube since it was available to avoid advertising I am seriously considering if it is worth continuing as a lot of tech/edu youtubers are duplicating their content to other platforms.

> Would refresh the front page and would get another scam in first position

I have a suspicion that their system simply doesn't reconstruct recommendation lists instantly.

I've been marking a lot of videos (of certain channels) as "not interested" in the past, and remember having the same frustration. The videos wouldn't go away after a page refresh. But after about a day or two you'd stop getting videos from unwanted channels if you're persistent enough with your marking, and if you don't click very similar videos from other channels or videos from the same channel (even accidentally).

They react much faster to what you do watch. The processing of data about what you do not want to watch is seemingly of much lower priority. Good thing, though, that it's of any priority at all.

I suppose it must be somewhat similar with these reports. The buildup is rather slow, but in the end it should work.

Reporting is useless because it takes a very high threshold to get the video removed if at all. Think about how many reports Trump videos or fox news videos get. So YouTube probably does not assign such reports much weight.
Seemed like a reasonable comment but thanks to jschlatter posting a link to the youtube report history in a comment here I have been able to verify every single video I have reported in the last few days has been deleted. So it does make a difference but the volume of these things must be massive.
I saw these a few months ago, maybe a year ago. I was reporting them, and within a few weeks, I stopped seeing them. I had assumed Google had solved the problem, not just that the problem had been moved to a new set of users...
Lower standards because they need more $. Buyer beware, no accountability, short term thinking. The marketing-advertising bubble is perhaps starting to pop.
Whenever my recommendations get worthless I delete my history. It helps for a few months.
Until the law changes so YouTube and other social media is considered complicit of whatever scam they promote they have no incentive to fix this. The spammers and potential marks still "engage" with the site and its ads which is all they care about.

Frankly this is not limited to YouTube even. Twitter has a spam problem for years where scammers would impersonate high-profile users and reply to the real account's tweets (as to be shown right underneath the real thing) with a scam disguised as a crypto giveaway. This is something that simple regular expressions can filter out, but why do this considering all of this malicious activity still counts as "engagement" and their bottom-line depends on this?

Facebook has hosted (and I'm sure still does) cybercrime groups even after being notified about it: https://krebsonsecurity.com/2019/04/a-year-later-cybercrime-...

Forget Twitter and YouTube. Plenty of scams come by email, text, and snail mail.
This scam is way more resiliant. ISPs seemd to come down hard on email scams.
hunt for any youtube cryptocurrency content and instantly get your youtube feed filled with a torrent of garbage containing thumbnails of graphs and amazed facial expressions..
> You can't tell me that they're too difficult to distinguish from legitimate content. Any human being or half-decent algorithm would be able to catch these with no trouble at all (heck, I could write a couple of regular expressions that would do the trick and have 0 false positives),

Of course I can tell you that! :)

Abuse is an adversarial and dynamic problem. A simple regex will never solve the problem for more than a very short time period. Sometimes such short term fixes are called for, since they'll buy you the time for something more principled. But they cannot be the solution forever, because the abusers will adapt. The more different variants you end up blocking, the more likely it is that they somehow interact and produce serious FPs.

And then you've got a PR mess on your hands with people complaining about how "Jeff Bezos is being censored by YouTube". Abuse is pretty thankless that way: nobody ever sees your successes, only your failures, and it's all a job of balancing the two.

What you actually want for fighting bulk spam (rather than other forms of abuse) is robust signals that are broadly useful rather than specific to one campaign.

As for your actual question, I haven't actually been seeing this in my recommendations (but then again, I don't have anything related to cryptocurrency or Elon Musk in my recommendations). Maybe this isn't actually as widespread a problem as you're suggesting? To put it bluntly, perhaps cryptocurrency is an area where the legit content is basically indistinguishable from scams.

We may as well think of the spammers as their own ML whos aim and weighting is to fool the spam prevention. So as good as you make the spam AI, the spammer AI evolves to be better than it since they keep iterating until they find whatever formula does not trip the spam protection.

>Any human being or half-decent algorithm would be able to catch these with no trouble at all

And this statement from OP is blatantly false because if it was true, no human would fall for the scams. But they do fall for them, in huge volumes. Especially if the scams appear to come from a known authority like Elon, the police, your boss, etc.

And if you did find a half decent algorithm that worked, the spammers would just iterate again to make spam that doesn't trip it.

> And if you did find a half decent algorithm that worked, the spammers would just iterate again to make spam that doesn't trip it.

So, the solution is not to even try? How has that worked out for other platforms?

Is it a major problem?

The scams seem super obvious, and if you subscribe to more content you’re interested in I guess they’d go away. Are they really causing you a problem?

I don't subscribe to any channels. Up until this point, their recommendation algorithm has been great, and I'd love for it to stay that way. Major problem? No. Increasing annoyance? Absolutely.

> The scams seem super obvious

Exactly. That's why I can't fathom why YT keeps letting them through.

> I don't subscribe to any channels.

Well that’s definitely where you’re going wrong! Subscribing is a strong signal to the algorithm. It’s having to guess otherwise. And at the moment I guess you’re feeding the algorithm with a stream of engagement with all these videos you’re linking to, digging yourself deeper!

Stop engaging with them, and subscribe to content you like instead.

I never get any videos even remotely like this in my feed - I get videos related to things I subscribe to.

You need to give the system some quality input. It isn’t magic. Garbage in garbage out.

Nope - I'm not going to start subscribing to channels. Their algorithm has plenty to work with based on my watch history and my spam reports (I've reported every single one of these scams I've come across, usually directly from the recommendations page - not even clicking to watch the video itself).

I'm interested in such a wide range of topics that I prefer recommendations based on individual videos I've watched rather than specific creators/channels. I'm glad you don't get these in your feed, but based on others in this thread, it sounds like I'm not that unique in experiencing the issue.

Your watch history includes these videos. Your spam reports are not part of the algo. Because you clicked on them and watch them long enough they become part of what you engage with.
> Your spam reports are not part of the algo

If you have first-hand knowledge that this is true, then I should inform you that this behavior is a bug in your system.

There's no conceivable way someone would want to see more of something they've flagged. And if you can't tell that kind of video from a legitimate SpaceX video I watch on occasion, then that's another issue.

(comment deleted)
I do subscribe to channels and my feed is also flooded with these scams. (Though usually it is not this bad -- it got much worse starting on the day of SpaceX's Inspiration 4 launch.)

I'm assuming the algorithm is showing them to me because I watch lots of SpaceX and other space videos. The descriptions in these videos are even ripped directly from legit videos that I actually watched!

What makes you think they are not trying? Google is constantly working on their moderation AIs. Cracking down harder on spam unfortunately means a lot of legitimate content gets hurt in the process.
I don't buy it. For a lot of spam, yes - I'm 100% in agreement. But in the cases I've seen recently, it's like Google isn't even trying, and this isn't some kind of "grey area" spam where it's hard to make a call (human or algorithmic).

I've worked on moderation systems myself, and there's a difference between easy detection and hard detection. This is the easy kind. Legitimate content wouldn't get hurt in this case. It's painfully obvious and yet it goes on streaming for hours.

You are clearly seeing the failures, and not the successes.

Imagine for just a moment what YouTube would look like if Google were actually not even trying. Nearly two years ago, 500 hours of video were being uploaded to YouTube every minute. YouTube is the higest-value video site in the world. A staggering amount of what is uploaded must be spam, fraud, and otherwise unwanted material, and yet most of us never see any of it.

I, for one, spend no time watching videos about crypto, so even though my watch history is super-nerdy, I don't see the kind of fraud and spam described here.

I don't doubt they're doing a lot to take down spam already. I understand the scale. I'm know they deal with a huge amount of it, and as you say, we only see the failures. I completely agree.

However, the type of spam I'm seeing recently is literally all the same video title, same channel name (maybe differing by a space or capitalization), same btc giveaway scam, etc. 90% is fake SpaceX streams. The streams stay online for hours before being shut down, and I know a lot of people are seeing them in their recommendations. It's not my fault for being interested in space, while also not wanting to see scam videos about Elon "giving away free crypto".

When I said it seems "like Google isn't even trying", I wasn't referring to all spam on the platform – just this very specific type.

Again, I know it would be worse if YT were really doing nothing. But these streams are getting more and more common (because I assume it's working to some extent for the scammers), and clearly YT isn't doing enough to stop them.

Here's the heuristic: if it's a live stream from a channel with relatively few subscribers and a name that very closely matches "SpaceX" or another high-profile channel name, the channel name was recently changed, and more than x number of people (with a solid history of reporting spam that was indeed later determined to be spam) have reported it as spam, you send it to a moderator to review and remove when appropriate. If the volume of videos matching those criteria is really too large for your moderation staff, then start looking at source IPs/locations of those that match the algorithmic part (I seriously doubt they're streaming over tor or coming from more than a small handful of locations), and tweak the parameters as necessary.

Really, with a little ML and the spam reports alone, they could shut these down 10x faster than they do now. I've reported a ton of them (and I know a lot of others have too), and I've never reported a video that didn't eventually end up being taken down. So have your model identify people like me and give our spam reports more weight. Respond in minutes rather than hours or days.

My point is, yes - YT is obviously already doing a lot to shut down spam, and it's probably 99.9999% effective. But the marginal amount that's making it through now could be dealt with fairly simply. I could never write an algorithm to be as effective as what they already have in place, but I could easily write one that's effective in dealing with this crap that's making it through. Would it be effective forever? Of course not, but it'd improve things significantly until it needs to be adjusted.

I agree.

But the only way to attract Google attention is either social media outcry or served a notice from law firm or better from government regulators.

Unfortunately, spamming accounts can be protected in places like Texas under a new law.
Yes, I understand that a regex isn't a long-term solution, but if you've actually seen any of these pop up on YouTube (they're on the front page of my recommendations right now; I've seen probably 30 just today), you'll know what I'm talking about. These are what I'd call "low-hanging fruit" spam – extremely easily distinguished from real content. After months of it happening (and I've seen literally hundreds of the exact same type of video recommended to me), a simple solution would undoubtedly be effective, or at least better than nothing. Get a single intern to double-check the ones that your algorithm identifies to avoid FPs. If the scammers get more sophisticated, cross that bridge when you get to it. But solve the immediate, easy issues now. It's seriously not that hard. This isn't the kind of difficult ML-generated spam that you're probably thinking of, and I agree - at that point it does become a game of cat and mouse. At that point, I'd be understanding when it sneaks through the cracks.

I've dealt with similar issues in my own work (I run a platform with a lot of content posted daily and have multiple tiers of spam control in place). The spam I'm seeing on YouTube is not the difficult kind, yet they're doing nothing to deter it. That's my complaint.

I've probably reported hundreds of these, so at the very least, stop recommending videos with the exact same title, description and content to me. It's ridiculous.

> they're on the front page of my recommendations right now

Can you share some examples? I’ve never seen what you’re taking about. My front page is all channels I subscribe to.

https://imgur.com/a/vzGvyvS

Even if spam detection didn't flag this stuff, their content ID system should be able to identify it as content stolen from other channels (with a fake tweet added at the top), then calling it "live".

I can’t find that video even by specifically searching for it!

Maybe it is suppressed for most users and something about your viewing profile opts you in? I’ve never seen anything like this video or anything else that looks like a scam on YouTube.

Everything on my recommended page is pretty relevant to my interests and seems legitimate.

It's probably been removed since I posted the image, but I think that's what the scammers are doing: posting them so frequently that when they get taken down, another one takes its place - hence the need for something to algorithmically block them before relying on user reporting.

Correction: the video is still up, ID is "qoFPBms-7g4" Others currently broadcasting similar scams:

BbUnrpFd5do

5VOQXRrfibE

fVNbM8mKHUU

M5dBA-01K88

> I've seen probably 30 just today

> Get a single intern to double-check the ones that your algorithm identifies to avoid FPs

I think you seriously, dramatically misestimate the scale at which both YouTube and the scammers in question are operating. By orders of magnitude, probably.

A long time ago social media companies realized that spam, users getting scammed does does hurt engagement, so not worth devoting much money to fix it. YouTube has probably 1/2 billion users. A few thousands getting ripped off by scams is immaterial.

Algos are a scammer's best friend, cause all you have to do is stay one step ahead of the algo . Human moderators would nip this shit in the bud. Such videos would last minutes on Reddit but they last forever on youtube.

>As for your actual question, I haven't actually been seeing this in my recommendations (but then again, I don't have anything related to cryptocurrency or Elon Musk in my recommendations)

I think that's the key, I think OP is finding the spam/scam by searching for content which are prone to exploitation, that said the questions raised by OP are valid.

I think YT algorithm actively pushes these scam to the top just because people who made similar searches clicked on those videos because of clickbait thumbnail/title, through which YT made good money through ads and scammers made good money through... scam.

Maybe they want to show problematic videos to security researchers
The Space X impersonation "live event" is still ongoing, btw... It must have started 2h+ hours ago. That's some very slow reaction. (Btw, do report it as "mass advertising", and the channel as impersonating the SpaceX channel - I just did, and I presume lots of others before me.)
These ppl making so much it's insane

https://www.blockchain.com/btc/address/13NKkx1epf5TcbisVSkDj...

$100-200k/day everyday

makes mafia, drug dealers jealous . no risk either. no tainted crypto

larry lawton brags about having made 4million stealing diamonds, that is like a month's worth of anonymous scam videos except multiply by 10. Former mafia ppl brag about their exploits on youtube. These guys make them look like rookies.

People are starting to realize that traceable and reversible bank transactions were actually a good thing..
Yes I've reported 4 fake crypto scam channels in the last few days.
In the last weeks I have been getting a LOT of spam delivered to the primary inbox in gmail - and its really annoying.
Is the problem as bad as you say and really getting worse? I don't know, I personally have never been recommended anything like this. I also think it's absurd for you to suggest that you could "write a couple of regular expressions" to successfully remove spam from YouTube with 0 false positives. The reason stuff like this will always exist is precisely because false positives are inevitable and they piss people off way more than spam. There are dozens of threads about random YouTube videos getting hit by the algorithm going back a decade on this site, the reality is that this problem will always exist.

With all that said, there's nothing wrong with asking YouTube to do better, but I'd temper my expectations if I were you.

People's expectations are already very low. But there dozens of these videos playing at any time . It's not like this is a one-time occurrence or under the radar. It's everywhere when you search for certain keywords or just browse certain topics like crypto or elon musk or any major event. You would think these highly paid engineers and such a hugely profitable company could do more to fix it.
On my front page right near the top, right now, is a video that looks EXACTLY like a SpaceX live stream, it's "Live Now", the channel name is "Space X", the thumbnail is the exact same one they use for their livestreams, and the title is "SpaceX's all-civilian Inspration4 mission!" and it has thousands of viewers right now.

The only way to tell it's fake is by hovering over the channel name and the URL isn't /c/SpaceX it's /channel/<randomcharacters>

I've seen these on the front page every single day for weeks now, always with different URLs. I tried to report a few but quickly found out from other comments here that it does nothing so I gave up and just ignore them now.

A family member lost $3k to one of these YouTube promotions that ended up a scam.

YouTube and the scammer make money here, so why would it change?

You're not the only one. I've been getting annoyed by these, too.

I was encouraged today to discover that youtube has a report history page that lets you know the outcome of your reports. Most of the videos I've reported have been removed, including most of the ones I've reported in the last 12 hours. So at least _something_ is being done about them, even though the current effort is not sufficient against the unending inflow of new scam videos.

Here's the report history page: https://www.youtube.com/reporthistory

Excellent. Good to see all the ones I reported have been removed. That gives me a lot more confidence in the reporting mechanism. Youtube is obviously dealing with a considerable volume of these. I still wonder if I am being used as a mechanical turk to help identify these. Hopefully the inputs are training a machine learning system.
I think a few of the other comments poke at this general idea, but I've come to be very intrigued by the meta-game of training these media recommendation engines (YouTube, Spotify, Netflix, TikTok?) by picking the right things to get a good feed. This sort of training is becoming more important than ever for a good experience in these platforms.

In my experience, YouTube seems to be one of the more reactive recommenders - especially if you go down rabbit-holes / dig deeper into certain types of topics. I have turned off both web history and Youtube watch history in my Google account and I find that the Youtube homepage very quickly devolves to lists of videos by channels I've subscribed to - a randomized subscription feed. I would really be interested in seeing recommendation behavior be quantified and documented in some systematic/objective way (if this is even possible).

I'd say that it is the internet in general that has a spam/scam problem. Due to bandwidth costs, which apparently only Google is able and willing to subsidize, YouTube is basically the internet for videos. Many opinions I read about YouTube do not consider this point and treat it as just another website. But for me, making Google responsible for whatever is posted on YouTube is akin to making ISPs responsible for whatever is going through their switches.

YouTube has a lot of problems, but I personally have no doubt that its existence is a net benefit to humanity. Calls for tighter regulation (or breaking Google down), which appear to be more and more common these days, make me wonder if would even have an alternative if YouTube went down. It is easy to take it for granted.

Good day everyone reading . If you are seeking means to recover cryptocurrency , tokens, coins and funds lost to any form of online scams in binary options, forex crypto-trading and investment , do yourself a favor and contact Coinwalletrecoup . Com. This recoup company literally saved my life when they helped me recover tokens of over $233,000 from online rippers impersonating group admins on WeBull and Tradestation, my 12 phrase got compromised .They had my blockchain wallet ransacked it and I was left with nothing. I almost lost all hope till I read a review online about this company Coinwalletrecoup . Com that saved my life by helping me recover my stolen funds. I provided details of the incidence and necessary information to complete recovery program and under a week my coins and tokens were recovered by them. All thanks to COINWALLETRECOUP . COM . You Are the Best!!