Facebook AI moderator confused videos of mass shootings and car washes
- Thread starter JournalBot
- Start date
I think you just thought up a more ethical solution to cock fights - robot cock fights.
Com' on down to the Robo Rooster Ranch!
Upvote
10
(10
/
0)
I think most people grossly overestimate the abilities of AI and ML, including, apparently, most of the companies deploying it. We've done lots of work with AI and ML for an application we've been building for the last three years, and I can tell you from first hand experience that neural nets are time consuming to train and fairly easy to corrupt. The broader they become in scope, the less accurate they become.
Artificial Intelligence isn't "intelligent" at all - it's a bunch of pattern matching algorithms. Sophisticated ones to be sure, but just algorithms.
AI can be very good at telling you that a thing is a class of thing within a defined training set, for example, let's say you want to identify fasteners in your industrial plant - AI can tell you that something is a bolt, nut, washer, or screw - that can be very reliable. Expecting AI to tell you what specific screw it is - thread pitch, head type, gauge, etc, is a much more difficult task, and you're going to be hard pressed to get over 90% accuracy reliably, and that's if you control for ambient light conditions and other variables.
AI is a vastly overrated class of technologies at this time - it's good at a lot of things, but it's pretty mediocre to downright crappy at most things.
Artificial Intelligence isn't "intelligent" at all - it's a bunch of pattern matching algorithms. Sophisticated ones to be sure, but just algorithms.
AI can be very good at telling you that a thing is a class of thing within a defined training set, for example, let's say you want to identify fasteners in your industrial plant - AI can tell you that something is a bolt, nut, washer, or screw - that can be very reliable. Expecting AI to tell you what specific screw it is - thread pitch, head type, gauge, etc, is a much more difficult task, and you're going to be hard pressed to get over 90% accuracy reliably, and that's if you control for ambient light conditions and other variables.
AI is a vastly overrated class of technologies at this time - it's good at a lot of things, but it's pretty mediocre to downright crappy at most things.
Upvote
25
(26
/
-1)
You keep proving my point about humans and online communication. And frankly, your sarcasm sucks.
Upvote
27
(27
/
0)
Upvote
20
(20
/
0)
Upvote
2
(3
/
-1)
That is true. But is also true that FB's algorithms can effectively block content based on basic keyword matching and user reports. The issue is that these basic filters certainly do not seem to be enabled in the vast groups hosting extremist ideas or simply in posts from users with some degree of connectivity in the network. Just check what happens if you report a casual user trying to sell something or posting something mildly inappropriate. Now, try doing the same on the content stemming out of "influencer" groups and see what happens (the BBC had this experience when reporting some major groups illegally selling land plots in the Amazon forest).
Filtering out millions of daily posts is a challenge that requires sci-fi AI capabilities. However, the root cause of this problem is mostly not technological, but an explicit decision by the FB execs on how content is selectively allowed. The fact is that if FB enabled basic filtering of content from groups spreading anything from conspiracy theories to extremism they would be excluding a very vocal segment of its user base that has been living in FB for years.
Upvote
10
(10
/
0)
The motivation is for more hate speech because hate speech drives engagement!
Upvote
17
(17
/
0)
Well that's not strictly true. There's the thrid way, which is the Facebook approach, don't pay lots and lots of people to do moderation and don't build an AI solution and just let the world burn. That way they save on people costs and development costs
That's not necessarily a bad thing, as, let's be honest, the second Facebook build a working AI moderation tool it'll go full Skynet and it'll be curtains for the human race.
Upvote
-1
(1
/
-2)
I think based on the results we can all see it is a very safe assumption that in this case they are the same thing.
Upvote
1
(1
/
0)
There is a good reason to be extra careful around the big groups/influencers. Those are bound to have a lot of people reporting them just to "troll" and you certainly don't want an overzealous, underpaid moderator to ban the "Red Cross" official account (if there is) for showing "drugs" for example. I am not saying that the same rules shouldn't apply to all accounts (big or small), just that there is certainly a reason to consider twice or thrice seriously before doing something that will influence potentially millions of people at once (the same for recent Twitter "scandal").
Upvote
1
(1
/
0)
Yeah, I follow a bunch of model related groups where people post images of their nice models of WWII airplanes. And guess what some of the german planes have .... yup. Small swastikas on the tail fin. So now most groups requier everyone to blurr a swastika on a modelplan as they otherwise can get deleted ...
There doesn't seem to be any context check in their moderating/banning.
Upvote
17
(17
/
0)
Well content moderation is definitively seen as pure expense and outsourced to the lowest bidder. So you have some non-native speaker in a third world country that's passed some bare minimum proficiency test who's paid by the click to determine in seconds if this is a violation or not. And you have to dredge the sewers of human existence every day, I'd rather work first tier call center support than that so you're not exactly getting the best of the best.
And nobody's mentioned the other side - the data will be notoriously noisy as people report stuff just to be vindinctive or they're crazy or part of a trolling/activist campaign to get legal content removed from the site. So it's not like you want to escalate all the "I don't see any violation" results either. In fact, judging by the constant stream of people innocently banned I'm guessing many moderators just go by "where there's smoke, there's fire" and ban stuff even if they don't understand the post and fake it until they make it or get fired.
TL;DR pay peanuts, get monkeys
Upvote
13
(13
/
0)
Let’s keep in mind Facebook makes more $ when engagement is high. And nothing keeps engagement high like hate and fear.
Zuckerberg would much rather it LOOK LIKE Facebook is tamping down that type of content that to actually address the problem.
Zuckerberg would much rather it LOOK LIKE Facebook is tamping down that type of content that to actually address the problem.
Upvote
13
(13
/
0)
This to the nth degree!
The cockfighting example is great, but let's take it a step further. In a backyard chicken raising group, someone posts a picture of two of their hens fighting (which sometimes happens) asking about solutions. How would the AI contextually know that it was a legitimate request for help rather than people taking bets on two hens trying to gain the alpha status?
But even worse, when content is incorrectly reported, the appeals process is abysmal. Most appeals just get set to the same removed state, and the option to appeal to the board is almost always declined. My wife is on a month-long mute after her and a friend were teasing each other - precisely the way they do in real life. Since the AI can't detect sarcasm, and the humans who supposedly intervene on users' behalf don't bother, I'm really not entirely sure what the upside of Facebook is these days.
Upvote
12
(12
/
0)
If they have grown to a scale where they can't effectively monitor their own domain, then it is past time for their demise.
Upvote
18
(18
/
0)
I've reported a dozen times groups for basically saying: "lets kill the poor because they are the source of crime" and it was never taken down. Within that group there were posts of people encouraging other members to "put to a stop to the mass of immigrants stealing our jobs" yet the automatic moderation tools were NEVER flagging / removing those posts... I can't say the AI field is crap, because I am pretty sure it is not, but facebook implementation surely seems to be crappy, particular outside of english speaking countries... it even fails with spanish the second language with the most native speakers...
Upvote
2
(2
/
0)
You can be sure there is plenty of context checking. If such content is posted by an "influencer" or in group that plays a significant role in FB's network in terms of content distribution then nothing will happen. If the same content is posted by a casual user or in a small hobbyist group then the content has a high likelihood of being automatically blocked and the group seriously risks being shut down if the "offenses" are repeated. FB allowed for years a non-irrelevant number of groups to disseminate misinformation and hate speech. Today, these groups are major content distribution hubs and FB has absolutely zero motivation to change the status quo.
Upvote
5
(5
/
0)
Very true, I actually recently got a comment flagged because I said that pro-lifers only care about life before it's born - once they're born, just line 'em up and shoot 'em down.
Upvote
2
(6
/
-4)
They're generally the same thing. When you evaluate a model, you plot a curve with (true acceptance rate, false acceptance rate) or in more human terms if I want to block X% of the bad posts, how many good posts will be blocked too as collateral damage.
Most of the time, you evaluate AUC = area under curve - the higher you get, the better and very often the better model is stronger at all thresholds. For very strong models they often measure TAR @ FAR 1e-N for particular N, like what's my positive rate if I accept 1/1000 errors or 1/1000000 errors.
Upvote
8
(8
/
0)
> The posts removed by AI tools only accounted for 3–5 percent of views of hate speech and 0.6 percent of views of violence and incitement.
Not trying to defend Fb, but wouldn’t it be true that content removed by automated systems would be removed far quicker and thus less likely to get any views at all? Anything that survived long enough to need manual moderation would almost certainly have time to attract more attention.
Not trying to defend Fb, but wouldn’t it be true that content removed by automated systems would be removed far quicker and thus less likely to get any views at all? Anything that survived long enough to need manual moderation would almost certainly have time to attract more attention.
Upvote
1
(1
/
0)
Small thing to consider with that chart though, I believe that's only in the case where you expect comparable amount of content. If bad content is a low % of the overall (one hopes with social media, but I don't know of course), then the random classifier would actually have a far worse impact in terms of false positives.They're generally the same thing. When you evaluate a model, you plot a curve with (true acceptance rate, false acceptance rate) or in more human terms if I want to block X% of the bad posts, how many good posts will be blocked too as collateral damage.
Most of the time, you evaluate AUC = area under curve - the higher you get, the better and very often the better model is stronger at all thresholds. For very strong models they often measure TAR @ FAR 1e-N for particular N, like what's my positive rate if I accept 1/1000 errors or 1/1000000 errors.
Upvote
5
(5
/
0)
And in the meantime, they treat their centers stuffed full of moderators like animals:
https://nymag.com/intelligencer/2019/02 ... ation.html
There is no hope for these people getting any break from AI (or a fair living wage, while we're at it Mark?)
Upvote
5
(5
/
0)
They're generally the same thing. When you evaluate a model, you plot a curve with (true acceptance rate, false acceptance rate) or in more human terms if I want to block X% of the bad posts, how many good posts will be blocked too as collateral damage.
Most of the time, you evaluate AUC = area under curve - the higher you get, the better and very often the better model is stronger at all thresholds. For very strong models they often measure TAR @ FAR 1e-N for particular N, like what's my positive rate if I accept 1/1000 errors or 1/1000000 errors.
Unfortunately a receiver operating characteristic curve relies on a clear and well defined "true positive" and "true negative". It generally doesn't work for nondichotomous continuous variables, and it can be horribly misleading if there isn't a gold standard for what is actually positive or negative (or the standard is subjectively dependent). For hate speech, both of those things are true, with the "gold standard" testers having variable criterion themselves.
My bet is that Facebook has set the threshold on its AI detection so far to the left of the ROC curve that it basically will only catch the most obvious of the obvious, and all other slightly obfuscated hate speech are falsely classified as negative. BTW this is the exact OPPOSITE of what you want to do for a "screening algorithm", which should be tuned to catch most cases but also alot of false positives. These are then tested with gold-standard testing to identify true positives and true negatives. Screening allows you to maximize your yield for gold standard testing (which is often expensive). Once again Facebook being cheap and trying to do the wrong thing with the wrong analytic testing.
Upvote
10
(10
/
0)
Snarky Robot
Ars Legatus Legionis
You're absolutely right: it's everyone else's fault you're not good at sarcasm, and you're definitely not an asshole about it.
Upvote
14
(14
/
0)
So, let them hire people to do it.
Insane you say? Just why is it unreasonable to expect companies to hire enough people to control their company? Facebook has over 1.5 billion members and makes hundreds of billions of dollars. If they need to go out and hire a million people to moderate FB then so be it.
If a company grows to the point where it cannot be managed that is corporate malpractice and that is the case with FB now. We need to make them be better and if that bankrupts them in the process?? Well, someone else will come along with the replacement.
Insane you say? Just why is it unreasonable to expect companies to hire enough people to control their company? Facebook has over 1.5 billion members and makes hundreds of billions of dollars. If they need to go out and hire a million people to moderate FB then so be it.
If a company grows to the point where it cannot be managed that is corporate malpractice and that is the case with FB now. We need to make them be better and if that bankrupts them in the process?? Well, someone else will come along with the replacement.
Upvote
4
(4
/
0)
If you still have a Facebook account you should be ashamed, delete it TODAY FFS
Upvote
4
(6
/
-2)
No, but for the curve to be valid the test distribution needs to match the training distribution. So if you're looking for the needle in the haystack, you need to train with a realistic amount of needles to hay. If you train it with 50-50 needles and then use it on data with 0.1% needles you're going to get a poor result. Maybe not a completely terrible result as it should still learn "needle features" and "hay features" but definitively not optimal boundary conditions.
Upvote
2
(2
/
0)
You forget that the goal of today's tech company leadership is to eliminate everyone from companies except for C-suite executives. That way all profit can go to C-suite and hedge funds which in their minds is the definition of a sustainable business model. Also, have you actually tried to talk to one of these "working class" people? I mean, you can't even chat about flying halfway around the world in their private jet for a round of golf and a (comped) 20k dinner with them! Why would you care to have any of them in your company!? /s
Edit: A reminder that most of these assholes consider the world of Atlas Shrugged to be their utopia, with everyone being "self actualized" architects whose primary job is pontificating about how great they are and how everyone else are just mooching off of them, while their food/clothing/shelter somehow mysteriously gets created without someone needing to be paid for them. A book written by a woman whose foundational philosophy was that selfishness was the ideal human state, who died broke and living off Social Security and Medicare because she couldn't help herself and the only people who were willing to help her out when she fell on hard times were godless communist humanitarian social workers.
Upvote
9
(11
/
-2)
I literally spent 90 out of 100 days this summer in "facebook jail" because crappy AI algorithms flagged posts that any human could easily see didn't violate a single of their false "standards". And there's no longer any way to protest or get a review (other than by the same crappy AI that flagged you in the first place). There used to be a secret one, and every single time prior I was able to get them to drop the bans, yet they kept escalating the "punishment". I actually had to create a new profile so that the next time their crappy AI bans me (and it will), it'll start out again at 24 hours and not 30 days...
Upvote
3
(3
/
0)
Well, you can't evaluate if an algorithm is right or wrong without deciding what's your desired outcome. And by correct in this context I mean the binary variable of what Facebook wants to block, to not get lost in an ideological debate of what constitutes hate speak or what they should and shouldn't block. Even if they don't agree you could hold some sort of majority vote to get a measuring stick to measure by.
Upvote
1
(1
/
0)
This...
Facebook will pay $52 million in settlement with moderators who developed PTSD on the job.
So its not just the cost of additional moderators...there's the risk of more financial exposure to PTSD claims from the expanded group of moderators.
Its a lose lose for Facebook and God help the poor employees/contractors would have to review all that cr@p. It's no wonder they hope to deal with this with AI.
Upvote
5
(5
/
0)
Upvote
1
(1
/
0)
How times change!
In the mid 60's, plastic kits of these planes came with decals for the markings, but without the small swastikas, so you had to buy a separate sheet of swastika decals to complete your model. Think they were illegal in some countries.
Upvote
1
(1
/
0)
So basically Facebook can't make an AI algorithm for shit. Or Facebook executives think that today's AI is the same as what you see in sci-fi, and is basically the same as an adult human brain, not something that is more like a trained dog doing tricks.
Upvote
-2
(0
/
-2)
I would hope that if a tool or group only hit a 3-5% success rate with hopes of up to 20%, a sane company would bolster or replace it.
Ah well. "Artificial intelligence? I see no evidence of intelligence."
Ah well. "Artificial intelligence? I see no evidence of intelligence."
I see. The goal of the moderation is to lower costs, not lower hate.Tim De Chant":3rf90dom said:
Yoda":3rf90dom said:
Upvote
1
(1
/
0)
Right, that word is clearly sexual/nudity according to their system...
Upvote
0
(0
/
0)