The algorithms that detect hate speech online are biased against black people Watch

AngeryPenguin
Badges: 18
#1
Report Thread starter 1 month ago
#1
A new study shows that leading AI models are 1.5 times more likely to flag tweets written by African Americans as “offensive” compared to other tweets.

Two new studies show that AI trained to identify hate speech may actually end up amplifying racial bias. In one study, researchers found that leading AI models for processing hate speech were one-and-a-half times more likely to flag tweets as offensive or hateful when they were written by African Americans, and 2.2 times more likely to flag tweets written in African American English (which is commonly spoken by black people in the US). Another study found similar widespread evidence of racial bias against black speech in five widely used academic data sets for studying hate speech that totaled around 155,800 Twitter posts.

This is in large part because what is considered offensive depends on social context. Terms that are slurs when used in some settings — like the “n-word” or “queer” — may not be in others. But algorithms — and content moderators who grade the test data that teaches these algorithms how to do their job — don’t usually know the context of the comments they’re reviewing.

Both papers, presented at a recent prestigious annual conference for computational linguistics, show how natural language processing AI — which is often proposed as a tool to objectively identify offensive language — can amplify the same biases that human beings have. They also prove how the test data that feeds these algorithms have baked-in bias from the start.

Anecdotally, activists have for some time accused platforms like Facebook of policing the speech of black Americans more strictly than that of white Americans. In one notable case reported on by Reveal, a black woman was banned from Facebook for posting the same “Dear White People” note that many of her white friends posted without suffering any consequences.

https://www.vox.com/recode/2019/8/15...cebook-twitter
0
reply
Jamie_1712
Badges: 11
Rep:
?
#2
Report 1 month ago
#2
(Original post by AngryRedhead)
Wouldn’t the logical conclusion of this be that African Americans are writing more offensive things rather than accusing AI of being racist? Queer is considered offensive in almost all contexts by the way
This is correct. It’s the same with African American arrest statistics in America. Only reason they are higher in proportion is because they commit crimes more frequently. Not the way the media twists it into “black man shot by police”.
2
reply
StolenFuture
Badges: 2
Rep:
?
#3
Report 1 month ago
#3
Notice how these Lefties can't accept that their speech may be actually more hateful, so accuse it of being biased instead
2
reply
Napp
Badges: 22
Rep:
?
#4
Report 1 month ago
#4
So from what it says theres nothing innately biased about the algorithms its just that err blacks seem to use offensive language more? I mean does it really matter whom uses the N word (for example) is it not meant to be considered offensive by all?
0
reply
Just my opinion
Badges: 17
Rep:
?
#5
Report 1 month ago
#5
(Original post by Jamie_1712)
This is correct. It’s the same with African American arrest statistics in America. Only reason they are higher in proportion is because they commit crimes more frequently. Not the way the media twists it into “black man shot by police”.
Don't come around here spouting your common sense
6% of a population (US) responsible for 50% of homicides are just a coincidence.
But you never see the figures for crime broken down by class, for some reason.
As a higher % of blacks are working class and working class as a % of society commit most crimes, you would think it would be more front and centre.
I grew up in a rough arsed northern working-class area (UK)a I now live in a very middle class area. The propensity to violence in minor exchanges is off the scale where I came from compared to where I now live.
Two men carrying ale bumping into each other where I came from would lead to harsh words or even a fight. Two men doing where I live leads to them both apologising.
Last edited by Just my opinion; 1 month ago
1
reply
J0n3zviper
Badges: 11
Rep:
?
#6
Report 1 month ago
#6
"African American English" wtf is that? Explain please.
0
reply
Obolinda
Badges: 21
Rep:
?
#7
Report 1 month ago
#7
(Original post by AngeryPenguin)
In one notable case reported on by Reveal, a black woman was banned from Facebook for posting the same “Dear White People” note that many of her white friends posted without suffering any consequences.
😲😲

(Original post by J0n3zviper)
"African American English" wtf is that? Explain please.
Ya know they sort of vernacular common in places with lots of black Americans? That stereotypical black, hood accents?
0
reply
J0n3zviper
Badges: 11
Rep:
?
#8
Report 1 month ago
#8
(Original post by Obolinda)
Ya know they sort of vernacular common in places with lots of black Americans? That stereotypical black, hood accents?
Ah right I'm with you.
0
reply
Stiff Little Fingers
Badges: 21
Rep:
?
#9
Report 4 weeks ago
#9
Not really that surprising - algorithms like this will be biased, because no algorithm is actually neutral like people think, they're programmed by people and thus prone to the biases of the programmer - for instance facial recognition software performing worse on PoC than on white men or how algorithms that purport to determine gender from text patterns will tend to assign academic writing as male - for machine learning in these cases the algorithms will simply not have been trained with appropriate materials or not with enough material. With this, I'd not at all be surprised if it flagged up more PoC for use of the n word, or LGBTQ folk for use of things like queer, when there is a world of difference between the people who historically have been targeted by their use as a slur looking to reclaim the words, against those on the outside perpetuating it as a slur, and whether machine learning can factor that in
1
reply
AngeryPenguin
Badges: 18
#10
Report Thread starter 4 weeks ago
#10
(Original post by StolenFuture)
Notice how these Lefties can't accept that their speech may be actually more hateful, so accuse it of being biased instead
Where is 'left-wing hate speech' mentioned in the article?
0
reply
ThomH97
Badges: 21
Rep:
?
#11
Report 4 weeks ago
#11
I'm pretty sure that the n word is offensive no matter the skin colour of the typer. To say the typer's skin colour matters would be racist, and that a black person should accept being called the n word by another black person is awful.

Without a better breakdown of the results, most notably excluding posts with the n word in them (the study itself admits 'a large part' of the imbalance is due to this), I don't see how anyone can really draw the conclusion stated.
1
reply
PTMalewski
Badges: 20
Rep:
?
#12
Report 4 weeks ago
#12
(Original post by AngeryPenguin)
A new study shows that leading AI models are 1.5 times more likely to flag tweets written by African Americans as “offensive” compared to other tweets.

Two new studies show that AI trained to identify hate speech may actually end up amplifying racial bias. In one study, researchers found that leading AI models for processing hate speech were one-and-a-half times more likely to flag tweets as offensive or hateful when they were written by African Americans, and 2.2 times more likely to flag tweets written in African American English (which is commonly spoken by black people in the US). Another study found similar widespread evidence of racial bias against black speech in five widely used academic data sets for studying hate speech that totaled around 155,800 Twitter posts.

https://www.vox.com/recode/2019/8/15...cebook-twitter
A completely worthless notion. What if this group simply does hateful speech more often?
1
reply
PTMalewski
Badges: 20
Rep:
?
#13
Report 4 weeks ago
#13
(Original post by Stiff Little Fingers)
Not really that surprising - algorithms like this will be biased, because no algorithm is actually neutral like people think, they're programmed by people and thus prone to the biases of the programmer

If a certain group of people, commits crimes more often, they will be more often arrested. Claiming that there is a bias, just because of that single data, is not only worthless, it is potentially harmful.

(Original post by Stiff Little Fingers)
- for instance facial recognition software performing worse on PoC
If a facial recognition software is a neural network and the team that makes it, consists of white people at large, then of course it will performe worse on PoC people, because the neural network most probably has been trained on the team, therefore, white people, therefore, it will beform worse on PoC, because they have different facial feaures, therefore, it will be more difficult for the AI to recognize faces. If Marsians want better facial recognizing software, then they have to make/train such software by themselves.

If you send a bulldozer into a battle, it will perform worse than a tank- as simple as that!


(Original post by Stiff Little Fingers)

than on white men or how algorithms that purport to determine gender from text patterns will tend to assign academic writing as male
And what does that tell us? That the algorithm is a part of a scam, is faulty, or just men tend to be more precise when writing than women?


(Original post by Stiff Little Fingers)
- for machine learning in these cases the algorithms will simply not have been trained with appropriate materials or not with enough material. With this, I'd not at all be surprised if it flagged up more PoC for use of the n word, or LGBTQ folk for use of things like queer, when there is a world of difference between the people who historically have been targeted by their use as a slur looking to reclaim the words, against those on the outside perpetuating it as a slur, and whether machine learning can factor that in
Correct, but a question should be asked: Is it the fault of the software made by one of the most clever people we have in our society, or is it a problem with the groups that look bad as recognized by the software?
Last edited by PTMalewski; 4 weeks ago
0
reply
Stiff Little Fingers
Badges: 21
Rep:
?
#14
Report 4 weeks ago
#14
(Original post by PTMalewski)
If a certain group of people, commits crimes more often, they will be more often arrested. Claiming that there is a bias, just because of that single data, is not only worthless, it is potentially harmful.
What on earth are you trying to say here? It's nothing to do with "do a group of people commit more crimes", it's to do with the programming of said algorithms, and their training. This isn't a worthless or harmful comment, it's an acknowledgment of the issues in neural networks - when written and trained by white men, they're incredibly lackluster in identifying anything that's not a white man. A neural network trained without any input of dialects like AAVE will struggle to give reliable data on it.
1
reply
PTMalewski
Badges: 20
Rep:
?
#15
Report 4 weeks ago
#15
(Original post by Stiff Little Fingers)
What on earth are you trying to say here? It's nothing to do with "do a group of people commit more crimes", it's to do with the programming of said algorithms, and their training. This isn't a worthless or harmful comment, it's an acknowledgment of the issues in neural networks - when written and trained by white men, they're incredibly lackluster in identifying anything that's not a white man. A neural network trained without any input of dialects like AAVE will struggle to give reliable data on it.
I'm pointing out, that if algorythms say that there is a problem with Marsians, it might actually be a problem with Marsians, while you instantly said it must be a problem with algorithms themselves, and used an example of facial recognition, describing a problem in way that suggests the white people who made those algorithms are biased (and therefore racist). This is practically hatred inciting.
And I'm not being hypocrite, because if the only biological difference between humans and Marsians would be different skin color, they would still live in different culture, therefore, behave differently. Therefore, they would eg. be more or less agressive.


No, either a specific tool has been misused, either actually there is a problem with the material the tool was used on.
Scissors for example, are designed for manual operation, nobody says that their designers were biased against people who have lost their hands. There is no point in blaming designers for racial or any other discrimination, especially when their inventions are still in early phase of development. First of all, we need to know specifically, how the tools we use work.
Last edited by PTMalewski; 4 weeks ago
0
reply
OurMoralChange
Badges: 2
Rep:
?
#16
Report 4 weeks ago
#16
What the regime defines as Hate Speech is infact legitimate criticism of Israeli/anti Islamic foreign policy.
1
reply
Stiff Little Fingers
Badges: 21
Rep:
?
#17
Report 4 weeks ago
#17
(Original post by PTMalewski)
I'm pointing out, that if algorythms say that there is a problem with Marsians, it might actually be a problem with Marsians, while you instantly said it must be a problem with algorithms themselves, and used an example of facial recognition, describing a problem in way that suggests the white people who made those algorithms are biased (and therefore racist). This is practically hatred inciting.
And I'm not being hypocrite, because if the only biological difference between humans and Marsians would be different skin color, they would still live in different culture, therefore, behave differently. Therefore, they would eg. be more or less agressive.


No, either a specific tool has been misused, either actually there is a problem with the material the tool was used on.
Scissors for example, are designed for manual operation, nobody says that their designers were biased against people who have lost their hands. There is no point in blaming designers for racial or any other discrimination, especially when their inventions are still in early phase of development. First of all, we need to know specifically, how the tools we use work.
Right, you're still either not understanding or ignoring it to peddle some sort of agenda. This is in no way comparable to scissors, scissors don't need to be trained in how to function. Neural networks are algorithms coded and trained to act like a human would to information, and as such will demonstrate the biases of their programmers (and everyone has unconscious biases, we just work to over come them), usually insufficient training on certain groups. That's not an accusation of malice, it's a warning not to take the reports of algorithms at face value. Going back to the identification of hate speech, is it trained in AAVE? Is it sufficiently intelligent to tell the difference between the n word being used with an a as a term of endearment/as reclamation & with a hard r by people who aren't black? (And sadly plenty of actual humans can't tell the difference).
0
reply
PTMalewski
Badges: 20
Rep:
?
#18
Report 4 weeks ago
#18
(Original post by Stiff Little Fingers)
Right, you're still either not understanding or ignoring it to peddle some sort of agenda. This is in no way comparable to scissors, scissors don't need to be trained in how to function.
It is comparable... If only you choose the paradigm at which it is comparable. Scissors are mmade to be operated by hand, and cut paper, while the facial regognition software we have so far, is made to operate on computers and recognize properly Caucasian faces.
So by your retorics, scissors are biased against people who lost their hands, and biased against steel, rocks and other materials that scissors don't cut.
No, they are biased. They are just not designed for different uses.
If something is good for everything, it is a useless crap- this holds true for lots of things we've ever made.

(Original post by Stiff Little Fingers)
Neural networks are algorithms coded and trained to act like a human would to information, and as such will demonstrate the biases of their programmers (and everyone has unconscious biases, we just work to over come them), usually insufficient training on certain groups. That's not an accusation of malice, it's a warning not to take the reports of algorithms at face value.
If somebody doesn't understand how algorithms or statics, or other data-collecting tools work, he or she will never understand the meaning of collected data anyway.

(Original post by Stiff Little Fingers)
Going back to the identification of hate speech, is it trained in AAVE? Is it sufficiently intelligent to tell the difference between the n word being used with an a as a term of endearment/as reclamation & with a hard r by people who aren't black? (And sadly plenty of actual humans can't tell the difference).
If humans can't tell the difference accurately, then why should we expect an early prototype AI to do better?
Speaking of n words, and things like that, I was bred not to use the 'ugly words'. Now I see even more sense in that than my parents probably saw. So if someone uses them, it is a problem with that person. I can adapt a little bit, to make understading possible, but that other person also has to help me, and adapt a little bit, to make the message understandable. A balanced must be maintain, in communication, relationships between humans, as well as between nations. Both have to adap, otherwise no communication is possible.

Things you call 'biases' are inevitable. We don't have unlimited understanding.
Therefore, tools also will always be 'biased'. You just have to know what your tools are good for.
0
reply
ltsmith
Badges: 20
Rep:
?
#19
Report 3 weeks ago
#19
I don't even need to look at the data to know this is because 'African American English' is likely to contain more swear words and profanity than standard English.

If you have more swear words and profanity in a sentence then the probability of it being classed as hate speech increases significantly.
1
reply
ltsmith
Badges: 20
Rep:
?
#20
Report 3 weeks ago
#20
(Original post by Stiff Little Fingers)
Right, you're still either not understanding or ignoring it to peddle some sort of agenda. This is in no way comparable to scissors, scissors don't need to be trained in how to function. Neural networks are algorithms coded and trained to act like a human would to information, and as such will demonstrate the biases of their programmers (and everyone has unconscious biases, we just work to over come them), usually insufficient training on certain groups. That's not an accusation of malice, it's a warning not to take the reports of algorithms at face value. Going back to the identification of hate speech, is it trained in AAVE? Is it sufficiently intelligent to tell the difference between the n word being used with an a as a term of endearment/as reclamation & with a hard r by people who aren't black? (And sadly plenty of actual humans can't tell the difference).
pls explain to me how I can program unconscious bias into my tensorflow project.

lmaoo
0
reply
X

Quick Reply

Attached files
Write a reply...
Reply
new posts
Back
to top
Latest
My Feed

See more of what you like on
The Student Room

You can personalise what you see on TSR. Tell us a little about yourself to get started.

Personalise

It is really important for me to be involved in helping make my university better

Strongly disagree (2)
14.29%
Disagree (0)
0%
Neither agree or disagree (3)
21.43%
Agree (5)
35.71%
Strongly Agree (4)
28.57%

Watched Threads

View All