Discriminating Systems - Gender, Race, and Power in AI

hi there today we're looking at discriminating systems gender race and power NII by Sarah Myers West Meredith Whitaker and Kate Crawford of the AI now Institute which is a part of New York University or associated with it this is not as much paper as it is a report kind of summarizing current literature and also kind of an opinion piece a slash recommendation giving document yes so we'll dive into it as you can see from the index it's quite a long report and we we don't have time to go into all of it actually most of it I just hope to kind of point out what the main arguments and themes are in the report kind of what it's trying to say pick out some interesting things and summarize it to the best of my ability also give it a little critique so let me actually go ahead and try to state the kind of core arguments that report is trying to make because it's not really clear from reading it and you have to kind of read the whole thing and then it kind of becomes clear what the argument is I feel though they somehow stated in the introduction numerous times in various ways so I might just be not as a tenth of a reader at first son but all right so here's the argument and I really hope I'm representing this correctly we have a problem currently that sometimes AI systems can exhibit what we usually call bias and we don't mean mathematical bias like bias-variance tradeoff we mean bias in a societal sense let's say bias against certain types of people where they shouldn't exist so for example let me draw an AI system and I'll just draw a little computer screen with a little light bulb all right because it's smart this is an AI system and the AI system and they give numerous examples one example they give is for is like face recognition algorithm that is much more accurate on faces of white males as opposed to darker skinned females so let me draw like two curves to represent these distributions are unequal and so the AI system exhibits some bias with respect to some kinds of people with an especially protected attributes and in this report they focus mainly on gender and race so that's what we're gonna talk about the second thing they observe so this observation one is I'm gonna draw some generic people here that represent the workforce of AI so the AI workforce is classified as all the people that work on AI B that University researchers or within companies building AI products or deploying them so this is the workforce and they observe that there is a unequal distribution among the AI workforce and so this distribution I'm also going to do this for unequal distribution there's an unequal distribution in the I workforce most notably it's predominantly males who work on AI and also white people are over-represented compared to the world population at large so that's kind of the two observations they make and now what they claim is that the inequality or the unequal representation in the workforce is causing the bias in the AI systems so they're basically saying the these AI systems are biased because that the workforce is unequally distributed and also they claim in a less powerful sense I feel what they claim there is a loop that this then leads back that because there is bias in the AI system that again leads to an unequal more ani distribution of the workforce so the core argument really is as they set out to do like in the introduction and also claim that they have done in the conclusion is to demonstrate these two directions here in a causal way so the systems are biased because there is an unequal representation in the workforce and that feeds back so the argument is that if you want to fix the bias here that then you will have to fix it via making the workforce more what they call diverse so less unilaterally distributed towards white males that's kind of the final conclusion if you read their report and the recommendations that's mainly what they're going for yeah so my opinion or in my opinion having read the report a couple of times is that as I see it they really don't demonstrate these links so they get they give examples of this and right they show that the workforce I know is unequally distributed they show that AI systems can exhibit such bias but they never actually show these links in my opinion they don't show this so if you make the claim that in order to fix the bias in AI systems you must fix the unequal representation in the workforce I would need a argument that says because there is unequal representation there for a therefore B therefore C therefore bias like a in actual argument to follow that the says because of this that because of that and so on like this is just not there like they simply show parallels they simply show that these two things exist and they just list example after example of that and the the I don't think they make this argument what I think too - also the other direction they don't really make this argument they accept it very in like one frame in one case where you know if you give them benefit of the doubt what I also I think is that it appears like the article if you read it and I encourage you to read it if you have some time it makes a lot of sense if you have already accepted this conclusion like if you've already accepted this then it's like oh yeah because the I feel this is just a text where the confirmation bias is so high just the way it's written that it must make a lot of sense to someone who's already kind of in on this conclusion but to someone who isn't sold yet like myself I am just not finding this convincing at all the second thing is that it very much feels like this isn't like a discovery or something but someone actually set out with the goal to address this here with the goal of I want companies to hire more of these people or of certain kinds of people or to become more diverse or to promote more of a certain type of people and now I'm going to find reasons for this and the reason is like oh look at this bias here this is cost this is caused by this other thing and therefore we must fix this other thing it very much feels like someone setting out with already the conclusion in mind rather than this being an honest investigation but yeah I mean read it for yourself I can't prove the absence of an argument by not reading every single line and I can't read every single line because it will just get very long and boring but read it yourself and I'm I find I think I'm pretty I've read it numerous times with really an open mind to be convinced that there is an argument in there but I don't think there is or a very strong argument for this all right let this

first part here is more or less a summary so research findings and we'll get to these things as they are important then they state recommendations right at the beginning so actually you'd have to read the article first this is kind of more of an abstract section but since it's right here we'll kind of jump right into it so these are recommendations and I've claimed they don't really show a connection but they actually just show examples of this and parallel them and this is reflected in like every single section including here in the recommendations they have recommendations for improving workplace diversity and addressing bias and discrimination in AI systems right so all right it in my case if you make this argument I would feel you'd also make recommendations for breaking these links but or argue why they can't be broken but all right let's jump into some of them and it is really a mixed bag here really um so some recommendations I am really in favor of just from the go not even you don't even need the article for those here published rent harassment and discrimination transparency reports including number of claims over time the types of claims submitted and actions taken so it's known that especially in these larger companies sexual harassment claims often go down in either bureaucracy or are kind of hushed under the table or something like this what you have to recognize is that a human resource department of a large company isn't there to serve the Human Resources it's company providing Human Resources that that's why a sexual harassment claim to an HR department is just a potential lawsuit and that's why they don't want to take it seriously except for it must go away really quickly so I think to kind of force companies or to ask companies to be more transparent to take more seriously these are keys of sexual harassment and assault and also discrimination he's a very valuable goal and I fully support this also the here commit to transparency around hiring practices especially hi regarding how candidates are leveled compensated and promoted that also is in the larger the company gets the less transparent this process usually becomes or the more bureaucratic the more people are able to game it and so on and distort it so I feel it's always good to be transparent around okay this person provides this much value to the company therefore they should be compensated in according like according to that or at least be transparent about it so these are the kind of recommendations I like then recommendations that really go into a different direction is something like this year change hiring practices to maximize diversity and this is kind of reflect I'm not going to go near this reflected in other points increase the number of people of color women and other underrepresented groups at senior leadership levels of AI companies across all departments so these things they are usually within like company diversity goals and so on doesn't really say how to do it but then the I mean as such they're not really recommendations yet they're more like goals but your recommendation 7 I think is the crucial one ensure executive incentive structures are tied to increases in hiring and retention of underrepresented groups so this is it's a bit of coded language but here they talk about executive incentive structure tied to hiring and retention of under represented groups this basically means if you are a manager or someone in charge of hiring or promoting and you hire or promote a underrepresented person and since they're talking about gender and race here if you it means if you hire or promote a person of color or a woman in this case you will be compensated more so at the end of the year you'll somehow have more money like more bonuses or more base comp or more equity or something like you will get more money so this recommendation is a direct call to hire based on race and gender so this is a direct call to racist and sexist hiring basically to discriminate people according to their skin color and according to their gender which I mean how is this okay with anyone like how can anyone how are people even able to state this and in like a high profile report like this and get away with it and not have people criticize them this directly calls for people to be treated according to their gender and race and probably as directly as you can go without getting into actual legal trouble but yeah I'm really against such practices I mean yeah that's I just don't know how this can ever be thought of as a good thing by anyone all right so well yeah in my mind this recommendation and this recommendation kind of our counter to each other because if I commit to transparency our people are okay now I can come I can transparently commit to be racist I guess but if I say okay I'm gonna come and promote people based on how much value they provide to the company then yeah I'd much rather have that than saying I'm going to comp and promote people based on their skin color all right so let's actually jump into the

report I'm not gonna these recommendations for addressing bias and discrimination in systems this these are fairly um general in common so as well

as I said we'll jump most of the things in the report so introduction so they start out with there is a diversity crisis the industry they give like some numbers like 15% of our research stuff and 10% at Google so 15% of Facebook or women so these are some kind of fairly known statistics about how the AI field is gender and race skewed currently so they say they claim involved the diversity problem is not just about women it's about gender race and most fundamentally about power it affects how a our companies work or products get built who they're designed to serve and who benefits from their development so this I find this this word power and this notion of power a lot in this report it's it appears again and again in like power dynamics and power dynamics among groups it's like a worldview it paints like a worldview where these different gender and race groups kind of struggle against each other to gain power over another and whoever's in power will try to remain in power in alliance with their gender and race group and try to keep the other groups down it's I don't I'm not sure that's the correct view of the world in my mind the world is comprised of individual people that want to achieve something for themselves and they would like to prop themselves up whereas in this worldview it's like I'm going to use the power of my group to keep the other groups down I don't know which worldview you subscribe to but I find the world is comprised of individuals yeah and this is not discrediting that some people have it harder because of their gender or race but to see the entire world as a power struggle between these groups to me it's a yeah and I'm not gonna point out everywhere it appears this power wording but it appears a lot and its really shapes how the report reads you have to kind of remember if you're a white male and currently the field is comprised of 90% white males you if you have like 10 hours let's say you have to drive to 10 hours to do something right you can either choose to put down some other groups like put down groups that you're not part of or you can choose to invest these 10 hours in putting up yourself you right so if I like I profit if I'm a white male I profit minimally from keeping the other groups down because guess what I still have to compete with the like one billion other white males there are that's that it's not gonna help me to keep down anyone else and especially like it's moronic I does that who like has Alliance except most French people like to their race or gender rather than to the people they admire and respect and like to work with so I'm gonna find like 10 hours today I'm gonna rather spend this in propping up myself compared to everyone else and I don't care what gender race they are and so that to me that's so much more accurate or I don't know plausible over worldview but just be aware that this report really takes on the language of kind of groups and power between groups and groups trying to you know kind of

gain power and keep in keep power and keep others from having power alright say to date the diversity problems of the industry and the issues of bias and the systems it builds have tended to be considered separately we suggest that these are two versions of the same problem issues of discrimination in the workforce and in system buildings are deeply intertwined challenge and moreover tackling the of bias within technical systems requires addressing workforce diversity and vice-versa so the UM I think this here actually is like how I described the argument and they kind of restated multiple times in a bit different way but I think this is the core and I really think I'm not misrepresenting the article here in that this is what they are setting out to do they're setting out to say okay the diversity the kind of unequal representation in the workforce and the bias in some AI systems are causally linked to each other and tackling one requires tackling the other so yeah if I'm misrepresenting them let me know but I really think I'm accurately representing their argument so what they do as I said is they give examples of one and of the other and also they really they're really unkind of discrediting the kind of issues to solve problems of bias in a different way so they point a little bit to this here in the introduction they say in the face of growing evidence the air research community and the industry producing our products have begun addressing the problem of bias by building on a body of work of fairness accountability and transparency so research concerns these issues for one is research showing that some products are unfair or on transparent and so on and on the other hand it's trying to devise algorithms that are more fair according to some notions or more accountable and transparent which means that the algorithm can kind of say why it made a certain decision rather than it being a deep learning system that you don't really have an insight these fields are active fields of research definitely very interesting to look into so but they kind of it is not already here but they say yeah we have adjusting air systems that and produce a result the deemed fair by one of my various mathematical definitions you can already see in the language here they don't really like this research and they are trying in this report to kind of discredit it or at least claim that it doesn't solve the whole problem because their point is of course you have to address this diversity issue in the workforce in order to fix the problems yeah so to this I just want to say no like if you can I mean you can criticize the fairness and accountability and transparency research field in that they haven't solved the problem fully yet but in principle if I have an algorithm if I'm being delivered an algorithm right and the fairness of literature has been applied to that algorithm and someone tells me I guarantee you here is a proof the algorithm is fair right then I really don't care who made that algorithm as long as it's fair the problem is fixed if the bias is gone and I don't care who fix it I don't care if the person who fix it is black or white or purple then the problem is fixed and they really have to really try to just make the counter-argument here is that no that's it's not enough but I claim yes it if you can actually solve the fairness problem technically then you have solved the fairness problem yeah the only thing you can do is claim that it is not good enough yet but not that it's funda to make the argument that it's fundamentally flawed approach and I don't think they succeed in doing that here yeah so they go on to say we should expand to consider not only how it itools can be biased technically but how they are shaped by the environments in which are built in and the people that build them again this focus like who builds the AI system I don't care I care what it does right as much as if some if I hear an argument for against something I don't care who makes the argument right I care what the argument says this is it's like an ad hominem attack for an entire community that's kind of how this article this report shows or is it appears to me so they say currently large-scale AI systems are developed almost exclusively in a handful of technology companies in a small set of elite university laboratories spaces that in the West tend to be extremely white affluent technically oriented and male so yeah they're their problem that's their fundamental problem here that these spaces are skewed in one direction interestingly enough their problem is not so much that it's that they're all in the same place right that they all live like 20 miles from each other in around San Francisco that's that seems to be not a problem at all as long as we get to like enough people of color and women into these twenty miles but yeah so that that's pointing out T the problem here or the yeah kind of issue they have all right so they go on just kind of wanna highlight again they say both within the spaces where AI is being created and the logic of how AI systems are being designed so paralleling the two things the cost of bias harassment and discrimination are borne by the same people gender minorities people of color other underrepresented groups then they also say similarly the benefits of such systems from profit to efficiency accrue primarily to those are already in positions of power tend to be white educated and male so they again they say the this points to a systemic relationship between patterns of exclusion within the field of AI in the industry driving its production on the one hand and the biases that manifest in the logics and applications of yet knowledge is on the other and they try to make this connection because they say the cost and the benefit of these two things are overlap in the people that wear it costs any benefits and I really again it's just a parallel but I really even don't think that's true because they the kind of argue against themselves later so they always say we have to look at again they should against the take much more than the technically driven problem-solving they point to this so a research requires looking at gender and race as categories within which you must think well in short sorry studies of discriminatory systems we need to ask who is harmed who benefits who gets to decide so it's kind of who bears the cost who bears the benefits and who has the power so that's yeah and again it's we seek to understand how AI disadvantages some we also consider how it works to the advantage of others so keep that in mind that's kind of the lens through how they analyze the this thing again wonder technologist power relationships and centers equity injustice that's they want to see this bigger picture so that's yeah keep yeah you again keep

that in mind so they go into a section called which humans are in the loop how workforces and AI systems interact so this kind of from the title of this section you think okay here's where we get in here's where we make the argument and they start by listing examples of how AI systems can be discriminatory and first they go into an example of Amazon had developed an experimental hiring tool to help rank job candidates by learning from its past reference preference Amazon hoped that the resume scanning tool will be able to efficiently identify qualified applicants comparing their applications to previous hires the system quickly began to downgrade resumes from candidates who attended all women's colleges along with any resumes that included the word women's after uncovering this bias Amazon engineers tried to fix the problem by directing the system to treat these terms in a neutral manner the company eventually abandoned the tool when they were unable to ensure that the algorithm would not be biased against women gender-based discrimination was built too deeply within the system and in Amazon's past hiring practices to be uprooted using a purely technical approach so this just the way is written I find to be quite dishonest but let's I let's analyze what happened here so their final claim is that gender-based discrimination was built too deeply within the system to be uploaded using a purely technical approach so this is one of their arguments they say technical approaches they don't help because the Amazon engineers tried to fix the problem all right but we're when they were unable to ensure that the algorithm would not be biased against women so if you read this you really I mean I really get the impression that's not what happened here what happened here most probably is Amazon built this tool okay and it fed in its past hires and we know of issues of like data set bias inherent in data sets so if your data set is skewed the AI tends to pick up on the skewed data set and become skewed itself okay so I actually would argue that most or all of the examples they stay stayed in here are examples of such biased data sets and not so the cause of the bias is the data set that they straining and not the person that ran the code or built the algorithm to train it on or build the deployment um and so but it doesn't matter your a your ear Amazon you built this tool and you realize oh it discriminates against people having women's on their CV so this is a pretty bad PR wise so you tell your engineers fix the problem so the engineers go fix the problem they come back and say okay we fix the problem and then what you do is you say okay engineers can you insure me that the algorithm would not be biased against women because if only the slightest bias exists if only it doesn't even have to be if one journalist finds one example where there is a down rank because I add the word women's then we are screwed right and the engineers will say no we can't guarantee that it's a deep learning system or something right we can't like give you a proof that it's not biased if you're a smart executive at that point you'll scrap the tool because the potential PR downside are just huge and probably they've also realized it's not that handy to have this two will compare to their recruiters doing their job because their recruiters might actually be you know good and have been doing this for a while so due to the fact that this tool was scrapped is probably much more a result of a PR disaster but also independent of that you say gender-based discrimination sorry gender-based discrimination was built to the play within the system to be uprooted using a purely technical approach it's just I mean what is this is just trying to discredit this kind of technical going about solving this problem but I'm pretty sure if someone comes to me and says here I have this tool and I can mathematically prove to you that it's not biased then it's not then the problem is solved and also I really don't see how the person training the algorithm or the personally researching such an algorithm has any influence over how the algorithm works because they're not the ones making the data set or if they are yeah then they can make a better data set also if a person comes and makes a better data set that will fix the problem and it doesn't matter what skin color the person has that makes the better data set so all of this link is just not demonstrated here or anywhere here at all but this here is the closest Amazon that this report actually comes to making this point so before I draw drew this thing work force AI bias right so this link since it here the AI system is used for hiring the workforce so at least one could make a claim that this link is somewhat demonstrated but I this it's a weak case I would agree but this is the closest they come so there and but then to go this direction you have to somehow argue well of the work force somehow makes the AI system biased no the work force influences the data set i if the AI is trained so if a hiring a I how do you train a hiring AI you optimally train it on the performance so this employee here is going to have a performance over time right and the AR system will look at that performance over time so if the AR system even if it's initially biased because it learns from the risk recruiters it will learn that okay actually if I always forego these women then I don't get as much performant of a workforce so I should correct for that so if you trained AI system on a good metric then and then this problem will Eva even out itself so but again this yeah this is this could be considered like one point in the argument but I think it's a very weak point and only because the AI system is actually used for hiring where I think the point they're making is a much larger one is the general bias in the AI systems contributes to the work force imbalances and there you somehow have to say that okay the AI system somehow influences society at large and society at large then go leads to the work force being skewed I don't yeah that it's just not strong enough in my opinion and the other direction also isn't strong here but again the examples only get weaker from here on they go on they say this is just one of many examples that show how the functional logics of a given technology echo the gender and racial dynamics of the industry that produced it here yeah this that's the claim they're making the echo the gender and racial dynamics and they actually making a stronger claim namely a causal claim they give you the other example of the Amazon's recognition facial analysis service previously demonstrated gender and racial biases worse than those of comparable tools so it failed to see dark-skinned women while being most proficient at detecting likes light-skinned men and they later go into this example again where they basically also stayed yes this is an issue of the dataset being much more comprised of white men and they say but then they you have to kind of make the turnaround argument and say well the dataset is a reflection of society and society you know part of society is the workforce and it's just not and eats it again this argument only works if you already believe the conclusion otherwise it there's actually no argument there or no solid one but what they do here is they say Amazon's initial response to such criticism has been to try and discredit the research behind it this reaction or let's first discuss this so the Amazon yeah Amazon of course being the accused here and a multi-billion dollar company and the criticism is something that is PR wise very bad for them a discredit the research tried to discredit the research behind it it's understandable that this could be dishonest from Amazon side right there I mean they're getting attacked it's like the tobacco company is trying to discredit the smoking research but still I mean that doesn't mean it's wrong it could actually be bad research right do you have to actually go and look at what's Amazon saying what is the research really doing is Amazon right or wrong completely open that Amazon is wrong here but you still have to go look and this citation here I've tried this citation here this one isn't to a to Amazon's response it's to like a medium article and the medium article doesn't even include Amazon's response F of look maybe I haven't seen it doesn't also doesn't link Allison's response maybe it links something that links something or that includes it in some way but basically this medium article only states here Amazon has been denying this or Amazon has been critical of this and if you state such a sentence Amazon's initial response to such criticism has been to try and discredit the research behind it I at least expect the citation to lead me to Amazon's response so that I can verify what they're saying right so this I mean I don't know I'm willing to chalk it up to incompetence rather than malice alright but then they go on and they say this reaction is evidence of the wider problem the research was conducted by two well-regarded AI researchers who are women of color by attempting to publicly discredit their expertise and research methods Amazon is reinforcing the same kinds of prejudice and erasure that the research critiques yeah here you go straight to the idea entity of the researchers you like play the race card straight out I mean this is maximum dishonesty right except if Amazon said something like well these women of color clearly because they're women of color they have no idea what they're doing or something like this is basically it's coded language for saying either saying you're not allowed to criticize people of color because they're a minority or you're basically saying Amazon is racist and that's why they criticize them they just don't take them seriously because they're women of color I mean both are this is just dishonesty really stated here - I mean again I'm perfectly willing to accept that Amazon's critique of this research is wrong and is not well intended because they're the ones attacked but you still have to examine it rather than say well they did shoot against women of color and therefore somehow that makes their counter argument irrelevant or even racist or something that's I don't know I find this dishonest yeah I don't know about you going on so they go on and state a number of examples of kind of examples of bias and discrimination in the workforce and they a lot of times they make a mixture kind of the kind of gender and race imbalance in a workforce and things like sexual harassment not being taken seriously by the companies and also the things like gender or race pay gaps which I'm open to accept the these things exist and or even intertwined but just to tell you what's happening because we're kind of skipping but it's kind of a mixture of these things so they say these issues are cysts there is a close relationship between these workplaces with discriminatory practices and discriminatory tools a feedback loop that is shaping the eye industry and its tools so again here to state I think I've stated it enough now that were demonstrated enough that I'm really representing their arguments as they intended it to namely that there is this kind of causal links and loop between these two things again they shoot against the fairness literature by saying from this perspective sorry from this perspective locating individual biases within giving given technical systems and attempting to fix them by tweaking the system becomes an exercise in futility only by examining discrimination through the lens of social logics who it benefits who it harms and how can we see the workings of these systems in the context of existing power relationships again they say these issues aren't technically fixing these systems won't help if that's the problem and then I agree if that causal link actually exists then technically fixing the system might not solve the problem not even sure I mean if you technically fix a system like this then you technically break the causal link and thereby fix the problem I would oh sure but again this is based on the hypothesis that they've already reached like demonstrated their conclusion which they haven't in which they are not in the entire article yeah so the next section goes into who makes I so I don't know what about you but this section was titled how workforces and AI systems interact and apart from one the AI system being used for hiring the workforce which is I said it's one instance where actually there could be one causal direction from bias to Minister representation to workforce other than that there isn't really anything in there that really shows how these to interact especially in a causal way all right the next section is called who makes AI as broadly about the gender and race imbalances or miss not unequal representation in the workforce and we're going to skip this diversity statistics that kind of that discuss that diversity statistics of companies aren't really accurate or can be you know massaged kind of by the companies which you know is true definitely companies will always try to maximize their profits and even if they

you know give out such a report so that definitely critical thinking is in order alright so the next section is called the discrimination feedback loop right if so if in the earlier section you felt like here we go into the meat then you must feel with this style like okay we're actually going to see how this loop works and how the two things are really linked like how one causes the other and vice versa so let's jump in they say AI systems increasingly play a role in our social and political institutions including education health care hiring criminal justice yes therefore we need to consider the relationship between the workplace diversity crisis and the problems with bias and discrimination in your systems know why I don't see how therefore but yeah so we need to consider the relationship okay if there is a relationship we need to consider whether there's a relationship okay granted so they say fairness accountability and transparency research is playing an emerging role and what they mean here is the aspect of fairness accountability and transparency research that shows that there is a problem so I told you there's two sides one side is showing there is a problem in current systems and the other side is trying to fix them so they're very much fans of the side that shows that there is a problem and they use show some of these problems here we've already seen some but they show some more like Facebook's at delivery systems let users to be shown as for housing and employment in a discriminatory manner so giving 20:19 study found significant racial bias and a widely use commercial algorithm used to determine whether patients will be enrolled in care management programs so the these are just examples of these AI systems being biased and they so they go

into this say taking a contextualized view may enable a more extensive account and the contextualized view they's when they say this they mean anything more than just a technical approach at solving these problems more extensive account of bias to emerge future work could examine the politics of system design study how they are systems in situated reality and study air systems and situated realities ask why a system was designed in a particular way how it was constructed who's interested shaped by the metrics in which its success or failure is assessed rather than solely focusing on improving existing datasets or individual algorithms yeah I agree I mean we always have to pay attention to these things especially like looking at the metrics by which it's success or failure is assessed but a lot of times this is rather straight forward in kind of if you look at the metric most often especially in commercial applications is money right so the metric of like an ad showing system like if I have a system to recommend ads to people show people ads and personalize them and so on I simply want to maximize my revenue so I want to sell someone something and everything I want to know is how likely is it that person is going to buy that thing right I that's basically yeah so in essence sometimes it's really valuable to consider what capitalism is so in capitalism in so capitalism these kind of this the system we're working on is kind of a form of limited capitalism but mostly capitalism and capitalism is very greedy so capitalism all corporations want to do basically is make money and there is and on the other side you have discrimination so discrimination meaning these unequal represent an equal distribution actively so an often sometimes these go hand-in-hand sometimes you can make more money by discriminating against a certain type of people and that's a really bad scenario like that's a very like this is really something where we need to take action but a lot of times these two things stand in opposition to each other so a little arrow here non-compatible that means if I want to sell someone something then I maximize my profit by not caring by accurately assessing how likely is it that person buys that thing if I want to discriminate here start discriminating according to skin color saying like no I don't like that this person with this skin color is able to buy this product I want to kind of keep them down and so on then I forego profit right then I actually even though this person could buy this thing I forego that so often these things are in direct opposition to each other also if I am in charge of hiring and I don't like people of a certain gender but they would actually be really good whatever got good employees so I forego that means I'm getting I pay more for less qualified people just because I'm biased and I'm down ranking unjustifiably these people of the gender I don't like so oftentimes you have to ask yourself are people fundamentally greedy or discriminatory which are they more if push comes to shove would they rather have more money or would they rather keep their own race and gender group in power and with just yeah so the and you have to ask this of corporations people and in my experience and view well-liked people are much more greedy than they are willing to discriminate and give up money for discrimination and so

if we look at metrics by which its success or failure of AI systems are designed then I would argue a lot of the times metrics actually profit incentives and especially if we look at data set construction if there is a skewed data set that makes my a I system be biased that actually loses me money and the company would profit a lot from building a better data set so looking at kind of metrics actually makes a lot of sense to me and very much in favor of that and I think by designing accurate metrics and then getting the best possible information the best possible data sets to maximize these metrics will oftentimes actually eliminate such forms of discrimination again there are situations where they don't we have to very conversant of these they go into this and they say also examine more thoroughly how societal discrimination surfaces in data provenance examining the history process of data set construction and considering how cultural norms and stereotypes were enumerated and represented at the time of data creation this is a big issue yes so the dataset construction kind of at the time of data creation and so on this was a big issue in these systems and a lot of bias and I would argue most of the bias we've seen here arises from corrupt datasets and from data sets that were constructed in an already biased way and the AI system trained on these data set simply replicates this bias so they I think that's very correct here and they go into this example they say the labeled faces in the wild dataset contains over 15,000 images only 7% of images are of black people and this is because these the media landscape of the early 2000s these images were gathered from the news media at the time predominantly featured white men in positions of celebrity in power this exactly so if you train a system on these then this data set the system will inherit this bias yeah so this is a classic example of a corrupt data set also I mean this isn't only with race and gender this is also if you like take pictures from IMDB yes a lot of this currently sell up a dataset that is used in all the gang research is collected from IMDB you probably have overly beautiful like pretty face people on there like that so the Urrea eye system your generative model is only gonna produce mostly pretty face people since movie stars tend to be a lot prettier than the average humans so that the kind of data set construction process I think is currently the biggest source of bias in AI but that also in its interesting that they go into this here and they kind of want to make the point that this is you know this is because society and you know power in society the data set reflects that but I would argue if someone makes a dataset that doesn't have this bias then the problem is solved and I don't care who makes the data set so the link between the work force and the bias is really broken by an argument like this because as soon as we have a correct data set and unbiased data set we can mitigate the bias and they go into this here they say sorry yeah they say down here they say these people this um these researchers have looked at these facial recognition systems and they assess this what we saw earlier higher error rates for darker skinned women than for any other group lowest error rates for light skinned men to measure this disparity these researchers developed a new dataset that is more balanced both in terms of gender and skin color good like problem like make a larger dataset to actually train on and then problem solved like and I don't care at all what race and what gender these people are well done like good people make if you good data set like this and then we've solved the problem what's the problem here like why would you ever care what these people look like if they do good work that's to me that's this just this actually breaks their own argument I don't know why they include it here but yeah to me that to then suggest that there's a link to the workforces if it's if here oh is obvious that if you fix the data set you can fix the recognition system all right so we'll go on here jump a couple more paragraphs except one they say they shoot again against this kind of say to this point a focus on fixing technical systems in isolation without examining their broader context of use and power and dynamics that attends to chuse is not limited in its intervention it can actively cause harm so if you fix the problem in a technical manner they argue here it can actively cause harm and the example they give is that facial and image recognition systems they are often applied in service of police surveillance which disproportionately harms poor people and communities of color so there that this person there's a quote from this person that says it is not social progress to make black people equally visible to software that will inevitably be further weaponized against us we are considered criminal and more surveillance of not magnitude whatever claim to a right of privacy that we may have is diminished by a state that believes we must always be watched and seen so this is an example where by improving the facial recognition for black people it makes them makes a police better at surveilling them which is true and then in it is an ethical problem that the police is able to use these facial recognition systems to surveil people that's a massive privacy problem it's a massive problem in how much the state is allowed to overreach and so on so I think it's a discussion itself but here they argue because at the very beginning I asked you to remember the this whole notion of we always have to look at who benefits from the way the AI system is constructed who is harmed from that who is who benefits from how the metrics are shaped and so on in this case we actually have a perfect example we if the a if their facial recognition system is very inaccurate for black people's faces that actually helps them in the societal context so by logic of this report here that must mean that somehow the bias like the bias works for them and thereby the system is good or something like this and by fixing it you actually make it worse here they say it can actively cause harm so I think this is a pretty much arguing against themselves earlier where they say or we always have to look at who benefits from the system yeah here if the face recognition system can't recognize you actually benefit so I don't think that argument works in any case except in if you only look at it when you want to look at it alright so this we're gonna jump a couple of sections here but the the core thing here was the feedback loop and again the feedback loop isn't demonstrated at all here just examples of systems that are biased and of datasets that are biased or because but there's no demonstration of how the workforce how I mean yeah just take this previous argument so the workforce is supposedly supremely white and it makes a face recognition system that makes that is performing poorly for darker skinned people and that actually in this context of police surveillance helps the darker skinned people compared to the lighter skinned people so that kind of existing an exact counter example to the argument that the but this room is representation in the workforce leads to the biases in the system if we interpret it through the lens who eat costs and who it benefits all right so the next section is corporate diversity beyond the pipeline problem and this is kind of an odd inclusion when I read it first to interpret to go against the pipeline problem here but it kind of makes sense if you know what these people set out to do so is to argue we must fix the workforce right we must hire more people of color more women and so on promote them more so they have a very much have a problem with this pipeline argument what the pipeline argument is the following so at the beginning if you consider like the educational career paths of people again you have like hundred percent of people that's represented at this at the beginning and then most of these people go through school so most of these go on this is kind of the area in here is the population and then some of them pursue higher education like some dropouts so this gets a smaller amount go on so this is here this is time and this is kind of volume of people and then very few going to computer science right and then even fewer going to AI so what you end up is just a tiny sliver of people that actually go into AR so this is called a pipeline and we have various junctions here like where you would go into higher education where you would choose your major in university where you would go into a subfield of computer science where the kind of volume of people drop significantly from one point to the other and now if you compare this and use it say we're not considered all of society but here over here we'll call consider all just men and over here we'll consider all women again they all go to high school and then university and then maybe it's very few go to CS even fewer go to AI what you'll find this and I've drawn it maybe wrong here is that this is smaller than this so if you comparatively look at how many male males end up in the AI field you will find that fewer end up in more and will end up in a field than women if you comparatively look at it so at and this is over time like at the beginning you have 50/50 main women distribution in society almost I guess I think slightly more boys are born but I could be wrong about this and then as you go through time here this skills that I believe so you go through high school and let's just assume like high school is still kinda equal depends on the country then you go to university or is actually more women at university slightly and then you go into computer science and in computer science and this is just relative here that's why kind of norm it at a 100 percent other ones these things would go down all of them at the same time but comparatively you have been much more men than women in computer science all right and then if you see who chooses a I don't know if there's any statistics of specifically choosing AI from computer science I'm just gonna assume that remains the same so if you look into the AI field kind of this will stay the same so in the AI feel do you have much more men than women and presumably because you already choosing computer science as their major or choosing any technical field as their major this is kind of the so-called pipeline argument so where do a AI companies hiring come in AI companies come in here they hire at this point after your university degree presumably there are there's exceptions but just say they hire after your University and therefore they basically have to choose from this distribution and if they just say okay we'll just take the top I don't know ten percent people will hire the good people of this we don't care what gender they are right so the top 10% here at then this will end up being the same distribution as you have graduates all right so this is kind of the company hiring from and let's say an 80/20 distribution without looking at gender we'll end up with an 80/20 distribution that's the pipeline argument of companies and they don't like the pipeline argument because the pipeline argument basically says that the problem is somewhere here right the problem isn't the company's hiring wrongly the problem isn't that the company's here deselect if the problem is somewhere here and because they want to make the argument that the company should hire in a different way they can't have that so they argue against it now to argue against this would actually be very easy if this argument were wrong like they claim the argument is is not good the pipeline argument isn't good if the pipeline argument were wrong what you'd have to do is you would have to say hey companies look at that in your company you have an 80/20 distribution men to women right that's pretty unequal and university graduates the pool you choose from is actually 50/50 so obviously you're engaged in discriminatory hiring because you know the pool is 50/50 there's no reason why it why your hiring practices should cause this inequality and therefore we can clearly show you do discriminatory hiring you should stop it you should definitely hire more women and people of color or more of these more of them Ortiz because your hiring practices are the problem but that's not the case how do I know because if it were the case they would simply state this definitely in this report if that were the case that you could actually show with numbers that the pipeline argument is wrong then they would absolutely do this instead they have to like go back and ramble around it for several pages which will mostly skip but mainly because this is the case it is the case that these companies hire from a pool of unequally represented people and the only argument that you can make is that well if you were to equalize this here then maybe here where the problem is that would fix like so the arguments often made if young girls choosing their majors have no one to look up to like no strong women in corporation CEO roles they will think that it's not a climate for women and they will elect not to go into these fields which is a valid argument like I'm completely open to that argument but it's the only argument you can make and still then even if you determine this as the cause I would still not support racist and sexist hiring practices like do something else like make them clear that the environment can be changed or change the environment like change the if it really is the case that it's kind of a non anti-woman environment change that if it's just the case that they perceive it as such change the perception but do not engage in discriminatory hiring practices because there's always someone losing out unfairly on these practices and that's a something I'm not willing to go into like that's engage in and I don't think people should engage being that actually that's why it's illegal so let's actually look at very few points this is just why the so they claim they go kind of go over these pipeline studies and they yeah they say the term used in the industry to reference the absence of diverse candidates in the hiring pool of to justify the inability of large firms to achieve diversity due to scarcity right so that's they basically agreed of that on the definition that I stated here so they companies that are challenged on their lack of diversity frequently cite pipeline studies as proof of the persistent challenge of finding enough women and people of color to hire yes and the yeah but they say but the evidence suggests otherwise for example in 2016 Facebook chief diversity officer wrote that it has become clear that at the most fundamental level appropriate representation technology or any other industry will depend upon more people having the opportunity to gain necessary skills through the public education system well yes that's something I would agree and that's something clearly that addresses this region here then and where the actual problem is happening so I would say that's a very good statement from the Facebook's chief diversity officer they say but as the Center for Investigative Reporting study of tech company diversity data found ninety-one large tech companies headquartered in Silicon Valley managed to hire higher percentages of black Latino and multiracial employees than Facebook that year well just if just because other companies employ racist and sexist hiring to improve their diversity numbers doesn't mean that Facebook has to do this right this it like just because other companies do this doesn't mean that it's a it's a good thing to do or that's how you should go about it Facebook simply says like if we want to hire without being racist or sexist if we want to just hire the best people then more of the best people have to be in the pipeline like more people have to gain access to educational opportunities so we can then hire them whereas these other companies probably make a big effort to say well even if you are not as educated if you even if you're not as qualified as this other person will hire you because of your skin color I don't think that's an argument in that in the favor of what the report is claiming like I don't think that is evidence that the pipeline argument is invalid all right so they're going to core themes in pipeline research and they do some overview of the kind of pipeline research that often so sometimes the pipeline research examines why for example why women don't choose to go into computer science as much and sometimes they focus on what is their perception of the field water it was what is their perceptions of the stereotypes of the field kind of culture the in the field is it suited to them what is their perception of how qualified they are for the field and is that true is that false and so on so this research examines a whole variety of things and it's very interesting actually to read through this research on the point out this here other studies suggest that gender is correlated with a person's motivations for pursuing a career in the field women and particularly women from low socioeconomic status or 'ti backgrounds are more likely to see computing as a versatile profession that provides an opportunity for secure employment higher pay and better social standing moreover their interests go beyond technical aspects of computing focusing instead on the purpose and application of software however such interests are often de-emphasized in computer science curricula a price technical skill and its ability to industrial in its applicability to industrial settings above all else so I find this really interesting because it's basically saying that women have different interests than men and on average all right that's basically saying that which is almost heresy like to say this and in this context like people will come after you if you suggest something that isn't yet they're just stating it here and remember this for later this is really funny that they're like they're yeah the interests could be different for women than for men and we might have to adjust our curriculum to be more suited to these different interests I mean yeah I'm sure that's uh yeah as I said you're like usually this is forbidden to say alright so they go on I say limitations of pipeline research right these are fairly like common limitations let's say of studies in general social science studies which I won't go into much again they stay they state we have to examine we don't know we don't only have to examine this but the problem they basically say the problem is actually the culture and perpetrators where they're saying I don't remember where this is stated but they again say like we have to examine who benefits from its present construction like who is underserved within the current tech ecology who benefits from its present construction how these dynamics might be untangled and so on so um again stating that in these kind of power relationships from for the different groups which I don't agree is in large part what's happening they say it's worth considering the scope of these studies and the by and large the recommendations the issue are limited targeted at the administrators of university computer science programs seeking to broaden diversity of their student body yes that's exactly where we saw the problem appears to be right so that the reason they have a problem with these studies is that they actually focus on the point where this discrepancy appears to happen because they want to claim that no no you should focus on a different point namely hiring in these companies hiring and promotion they say though important also at least they acknowledge that that's an important problem this is a narrow frame through which potential solutions to barriers to inclusion it does not address the companies that hire computer science students the pier is responsible for promulgating stereotype views or engaging in hostile behavior or the broader social conditions that may influence students success and later science programs actually the research and even some of the examples they've included of this research addresses all of this like they research often addresses the kind of stereotypes and the how the peers act and how the companies act and also how the companies hire and how the kind of people have something to look forward to or nothing and how that influences their decisions yeah again they say the studies are frequently cited by those within corporate environments to justify their own lack of diversity as they situate the locus of change outside of the corporation itself as such pipeline studies are disproportionately emphasized as a part of the broader research agenda on diversity in technology again they state companies use this together and of course like companies of course there can have use this to get out I mean I agree at least with that I agree that companies are gonna try to use this to get out of responsibility certainly all right so the last section here is the pipeline dreams after years of research again this is on this pipeline studies basically say they say the pipeline research hasn't shown like hasn't borne fruit it hasn't led to meaningful change in the field even though we've researched this the reason they stay the number of reasons they tends to place the onus to solve issues of discrimination Silicon Valley on those who are discriminated against rather than the perpetrators I find this word choice really interesting perpetrators right like again the group of white men is trying to put down everyone else that's the perspective that the article takes and again and it's not even true this research a lot of times it actually says the reason why for example women don't choose to go into computer science is the male-dominated culture within these corporations is the perception of you know this of this not being a woman friendly environment is the people here of sexual harassment and so on so I it's not even true but moreover I just wanted to point out the choice of world here perpetrators I don't yeah I don't know how you get to this word it really shows kind of a world view of the authors in my opinion alright so they go on and say ok this pipeline studies haven't been beneficial and companies haven't done much or that it hasn't been successful they're going to work or LED initiatives which I'm going to skip here it's just a kind of a reporting of what happens at companies where the workers themselves organized and then the last section here is the pushback against diversity so in this section they're kind of documenting and arguing against people who have basically stated counter arguments to their recommendations mainly so their recommendations being let's change the hiring let's change the promotion and so on to be based on race and gender and the pushback here yeah characterized in different ways so we'll go through this the last section I know it's a long video already if you're still here like the one person who's still here hi I hope you're doing well keep hydrated yeah so they say it's a critical time we now see diversity itself being weaponized so they say this growing awareness accompanied by demands for inclusion equity has led to some change but there has also been resistance especially among those implicitly privileged by the status quo so again jumping straight to attack on the person like I don't care if who makes an argument against me I want to go on the argument and I'm gonna go on the content of the argument but these people straight first thing they stayed is that's but that's just by the people who are you know benefit it's just by the white men basically straight to that identity of the person that's dishonesty right there yeah so those questioning and even rejecting the idea that racism and misogyny and harassment are problems within the AI field and the tech industry have appropriated the language of diversity to argue that efforts to improve inclusion are in fact exclusionary and addressing the deeper structural challenges posed by racism sexism inequity is misguided and yes definitely efforts improve inclusion can be exclusionary like just because so this is a thing just because you're say fixing that's a just because you're fixing a problem doesn't mean the method you're using to fixing it is justified and is itself good right methods to improve inclusion can be exclusionary and some that have been proposed or exclusionary definitely it depends on the method it doesn't mean these people are against these efforts it means that the measures for example implementing racist hiring policy I can definitely see that this is going to lead to more equal representation within the workforce but the tool itself is really bad and exclusionary and discriminating so yeah I would say that's it's it's accurate that they can be exclusionary they say for example somewhere researchers greeted the announcement of black and AI workshop Europe's leading machine learning conference by questioning whether the event was necessary arguing that it would be discriminatory well can't they question whether the event was necessary like that would I would hear I would need a discussion what is it for right what why is this event happening and what is it doing and is it discriminatory like it could be like any event can be discriminatory does it discriminate based on race or gender or anything is it you know does it do so unjustly and old so I don't just don't see why question could still be wrong like you could question and then you can be wrong but you should be taking on your argument but the argument here is just already questioning this is already on the wrong side of the argument and I don't agree with this though these people that question this workshop and have a particular opinion on these things but I have the opinion that you have to take arguments at their argument value and not just and who makes them or whether or not they're against a particular viewpoint all right so they say such pushback often centers calls for cognitive diversity or viewpoint diversity the idea that individual differences in the ways people think and understand the world are distinctions that should be counted alongside or instead of other identity categories such as race and gender well yes that's it I mean isn't that is that very reasonable thing to say isn't it very reasonable to say that differences in the ways people think and understand the world there are distinctions that should be counted alongside of identity categories such as race and gender they say a dozen white men so long as they were not raised in the same household and don't think identical thoughts could be considered diverse that's I don't know if this is a sarcastic statement or not but really it's kind of the counterpoint they're trying to make care that but yes I would totally agree with this statement in a way a white man growing up in San Francisco a rural Idaho and white man growing up in Florida a white man growing up in Western Europe one in Russia and one growing up on the road with its circus his circus parents in Mongolia would definitely be that plenty diverse right I mean they criticize this here but this is actually how can you not see this that yes these are valid differences and people are going to think differently independent of how they look people are going to have different thoughts and it's important to recognize other people think differently and therefore you should you know include them if it's relevant and the counter-argument to this is of course what the authors are saying basically is that twelve at a dozen people as long as they are don't look the same could be considered diverse even if they all were raised in the same place and basically all live in San Francisco and think the exact same thing yeah that's I mean it sounds to me it sounds as absurd as the other way around it to me so here's my here's a my thoughts on this I am NOT going to pretend that I know what life is like as a woman all right I'm absolutely sure that four areas of life it is definitely valuable to listen to the experience of a woman or multiple women an aggregate of women because the life is just different as a woman life is also different as a black person I absolutely concede that there are things that I might not be able to draw from my life experience because I am NOT of that skin color that different problems that people face and that's why it's important to have an opinion of that at the table but I'm also absolutely certain that I have no relation to someone who grew up as a child popstar from the age of 12 and then had that life I have no relation to someone growing up under a communist regime in kind of a Buddhistic religious tradition I just don't and I don't care how they look they have different experiences they have different bodies of knowledge to draw on and I don't think why we should make the difference along the exact lines of race and gender yeah well that's what they that's of course what they argue here those arguments work by centering identity while flattening or ignoring power relationships here the VP the Facebook VP of engineering said that the ultimate goal is cognitive diversity and cognitive diversity is correlated with identity diversity that means it's not just about getting women in tech it's about broad voices broad representation right so the this is exactly what I would say the reason why we want different but a woman or a black person at the table is because they have a different knowledge different thought because of their different life experience they have different thoughts that they can bring in so actually by including these what they call bodies it is about cognitive diversity even in itself but the authors here really see this from a different angle they really see this in terms of power relationships between race and general groups and I yeah the arguments of the authors don't make sense if you don't view it through that lens and that lens to me is just such a it's such a I don't know it's just sad look on the world and also I think is a very inaccurate look on the world and it's I think a very dangerous look on the world um yeah again they say instead of looking at historical patterns of marginalization calls for cognitive diversity argue that all differences are equal no we're not like no calls for cognitive diversity or don't argue that all differences are equal well aware that some people have it harder differences are bigger worse or better that's absolutely well aware all they're saying is that race and gender shouldn't be the like only things to consider and shouldn't be in itself be considered diverse just because someone is of a certain skin color it doesn't mean anything right it doesn't actually tell you anything about that person so why not consider people as individuals and look at what was their life like until this point and what could they contribute to the discussion we're having rather than looking at the color of their skin I mean if played a role in their life then obviously that would manifest in my suggestion as well but to just look at people through this kind of group lens is so foreign to me and yeah I feel it's quite dangerous yeah so again and this could argue that all differences are equal I mean the point where you have to start misrepresenting what the counter-argument is saying that's really how you know you're dealing with it with not a well-intentioned person on the other side of the discussion this is really politics now this isn't a well-intended argumentation it's really someone - trying to achieve some goal because they have to misrepresent the other side and this only gets worse from here they say recently was exemplified in the controversy over Google's appointment of Heritage Foundation see ok calls James to its advanced technology external advisory council Google's reasoning for the appointment of James was ostensibly to ensure diversity of thought by including a conservative viewpoint on the council all right so google has a technology advisory board council sorry of external people and they've included a conservative and she is by all metrics let's say a standard conservative so this is not a far-right neo-nazi type I don't know but this is someone who has similar opinions then half the US country and in generally in at least in the Western world generally half of the of the country's population tends to be conservative more or less I mean that there's differences but yeah so this is a this is an opinion that a large portion of the population shares so it would be I don't know it would be suitable to include at least someone of that opinion in an external advisory council to have that on board it you don't have to listen to her like it's not like she's made King it's simply that she will have the opportunity to input her voice at representative of kind of that large very large percentage of people they go on and say James is also a black woman thus adding racial and gender diversity to the panel so even further right this is it's a conservative black woman all right but the pushback following James's inclusion focused on her policy position siding specifically her vocal auntie LGBTQ and anti-immigrant views and highlighted why cognitive diversity is a particularly limited lens and the pushback here was very much spearheaded by one of the authors of this article so I um this isn't just reporting I will also I'll also criticized the this pushback here since its you know it's kind of argued for in this article it's not just reported and also because the authors are the same so here they say they have vocal anti LGBTQ and anti-immigrant views and I haven't actually gone specifically and looked at what this person particularly has said but given that she's a standard conservative and has been in public office I believe under george w bush she can't like i have trouble believing that she has like extremely hateful opinions like these people shouldn't exist or like something like that nature like often people like conservative people have issues with forcing people to adopt certain pronouns for people or issues with which bathrooms do people go in and you know generally are tougher on immigration especially legal immigration and so on i mean these are views that people hold and a large part of people and these are discussions to be had so including this person would be very sensible move but they say in a letter opposing the appointment a group of Google workers calling themselves Googlers against transphobia and hate we had transfer peon hate responded to the idea that diversity of thought justified James's addition to the council this is a weaponization of the language of diversity by appointing James to the IAEA see Google elevates and endorses her view implying that hers is a valid perspective worthy of inclusions in its decision-making this is unacceptable here it says again the author was one of the organizers of that and that's what they're saying here this the views if you don't have our views it's these are unacceptable views right it's a valid perspective worthy of inclusion it's what they're saying basically is you don't even talk to these to this person like talking to this person considering their opinion you can still evaluate the opinion but even considering their opinion is already wrong and that given that the person is a black woman so basically they are called the author's idea of diversity is people that look different that are from race and gender groups that have don't have much power or perceived that what they call power right now as long as they all think exactly as we think right then that's fine as long as they they share our thoughts as long as they don't have dissenting opinions we want the different looking people but don't dare talk to anyone of a different opinion yeah this I don't see how I mean this these authors in my opinion they really live in a bubble the in a tiny Silicon Valley or Silicon Valley influenced spaces because this is half the people they're basically saying half the people in their greater community in their country aren't even worthy listening to their opinions aren't even worthy of inclusion in of consideration so yeah well done might as well discredit them at once I'm sure I'm sure that's gonna fly well with these people all right yeah hmm let's start calling them deplorable and see what they do maybe they'll return the favor and elect a just to stick it in your face and that's what happened so the idea of cognitive diversity is mobilized by some support in support that the AI field and the tech industry are already diverse goals for us to support claims that not including identities like widened mail constitutes discrimination yes it can like if you include every single identity except white and male that constitutes discrimination that's I mean yes even if they're in the majority is still constitutes discrimination like no one can help being born white demille no one white male chose to be born like that don't mostly don't choose the melanin content of your skin you can modulate it a bit by going to the Sun which computer science people statistically don't do very often so there's not much leeway there so yeah this is to not include identities like that if you include every other one can't constitute discrimination true a July 2017 memo written by James d'amour a software engineer at Google is illustrative of such pushback titled Google's ideological echoed chamber and published in an internal mail list memo critiqued the company's diversity policies arguing the biological differences between men and women rather than bias and discrimination help explain gender disparities at the company I feel the you can leave out the rather than here I think the memo simply stated that biological differences can help explain the gender disparities the most objective writing the memo was to make the case that policies designed to achieve equal representation are unfair divisive and bad for business well some are yes especially the recommendations that you've given at the beginning number seven is unfair divisive and I would also argue bad for business yes so supporters for the mores point of view at times even drew on rather odd part of the pipeline to make the case that diversity initiatives are in fact discriminatory they argue incorrectly that if there aren't qualified candidates in the pipeline then hiring those who are unqualified of the basis of identity discriminating discriminates against those who are qualified no I would say hiring anyone on the basis of identity discriminates I mean inherently so against I think that's the larger argument that these people are making which is not incorrect is very correct yeah so in an update to the memo the Moore himself asserted that he values diversity and inclusion but his primary concern was cognitive diversity he says diversity inclusion is not denying that sexism exists doesn't endorse using stereotypes and in specific I've read the memo and it directly says these are population level kind of statistics and there is more overlap than difference and you absolutely can't say anything about an individual by looking at these statistics that's almost a quote from this memo so he was very much concerned with considering people as individuals but also if you like he was basically making the same argument as earlier I told you to remember hey look the this one stays described this one study that found that women's interests might be different and we might shape the curriculum that's basically what Moore said he said women's interests might be different we'd have to maybe shape the way we do work like change software engineering to attract more of them that's what was one of his points so is exactly the same thing but of course he's a misogynist because he suggested that this could be due partly because of biological differences and I think it the way he was dragged through the mud is just crazy and this they shoot here very much against this kind of biological what they call biological determinism will see this very briefly they say diversity becomes an empty signifier stripped of the histories and experiences of systemic discrimination repurposed around the ideology rather than bodies yeah I'd say diversity has nothing inherently to do with bodies like as such I think that only that's only the case if you are already convinced of this that's a yeah within hours of the memos publication harassment targeting minority advocates who push back against the claims in the memo began with a particular focus on queer and trans workers book that's bad but also I think the pushback against people who voiced support is also was also pretty bad because one of them was fired as you already stated Google's vice president of diversity even locked down her Twitter account shortly after the Morris firing responding to the barrage of threats describing her as a police Nazi well yeah if you fire something I mean undoubtedly Google fired this guy because they thought it was less of a PR disaster if they also fired him now then I mean this wasn't there probably wasn't an ideological decision much more a PR decision but yeah if you fire someone after stating like after stating something like this it very much looks like you're firing them because you don't like their ideas and you don't like what they're saying which generally not in favor of censoring freedom of speech but yeah that's being said harassment is bad don't harass people also that being said criticism isn't always harassment and don't conflate the two the mores memory memo also stated at the distribution of preference abilities of men and women differ in part due to biological causes and that these differences may explain why we don't see equal representation of women in tech and leadership this certain hinges on a flawed assumption that identities like gender and race are essential and fixed biological attributes and that inequalities are at least in part the product of such irreducible differences well I mean if they're not fixed biological attributes certainly gender and race have a 0. 99 correlation with biology or invent since your biology is first and it's determined when you're conceived that we that demonstrates a causal direction right so even if they're not exactly fixed they're like overwhelmingly fixed and to suggest that is a flawed assumption that these inequalities are at least part product of such differences what you'd have to do if they simply state it's a flawed assumption what you have to do in order to show this is a flawed assumption you have to show that gender and race as far as they're biologically determined have no influence whatsoever on these differences that's what you have to show right that's the counterclaim because the claim is they have at least in part something to do with it and that's also I believe what the molar stated and what did predominant opinion it like it's very like all the research points to for example there is a large difference in interest between genders as far as for example career selection goes and so on now we can talk about why that is but there is also a large consensus I believe that this s at least partly determined to however degree but it is at least partly determined by biology in order to show that this is flawed you need to show that it does not have it can't have any influence right you have to basically prove them the impossibility of this having an influence which no one has done so far much to the contrary so they simply state this is a flawed assumption kind of shows to me that they've already they are there they're in the bubble and they're expecting to speak to people in the same bubble yeah so they go on and kind of discredit this as call their biological determinism which yeah I don't think that's a correct use of the term biological determinism but you can judge for yourself all I think these people are saying that biology might have some influence and we could adjust for that it's not even right it's not even yeah this comes up here so conclusion finally I think it's been two hours sorry conclusion throughout this report we've outlined the scope and scale of the problem tracing how the diversity crisis in the industry and the problems of bias and air systems are interrelated aspect of the same issue no in the past these topics are commonly examined in isolation but increasing evidence shows that they are closely intertwined No there you've shown that they're parallel you have absolutely not shown that they're interrelated aspects of the same issue and you have not shown that one any one of these causal influences the other that there is any feedback loop you have not shown that fixing one leads to fixing the other I mean you could also like take a company that extremely is focused on or like for some reason has a different work force and then show how their products with the same data sets as the previous companies don't end up being biased probably not so easy but again they're not none of that is in the report there are many things you could actually do to show what you wanted to show but it's just not the case in this article our analysis surface two prominent responses to the diversity crisis on one hand a worker driven movement which we've skipped on the other hand we observe a small but vocal counter movement that actively resists diversity in the industry I mean what again what dishonesty actively resists diversity I mean this the thought that these people stray around like no I don't like to other working people it's just so absurd all they're saying is that either we don't understand the problem in the correct way or our tools aren't appropriate to solve the problem I think everyone has the same goal of the workplace and the AI systems being as fair and as non-discriminatory as possible - I don't the misrepresentation of the other side is something that really bugs me and it's something that these authors do a lot so yeah I lose my polite side maybe and uses arguments from biological determinism to assert that women are inherently less suited to computer science into AI what a load of crap sorry but uses to assert that women are inherently less suited to computer science in it no one okay not no one but no one that I know asserts that absolutely no one that makes these argument okay sorry not no one you can always find a sexist douchebag that makes that argument but this is not a serious argument made and this is not there this counter movement most people in the argument that most people in this country movement make not at all and to represent them as such is just so dishonest that yeah this this basically this is the it's nice that it's in the conclusion because it finally like at the end it completely destroys the credibility of me taking seriously these authors I felt they had so that the parts we skipped over I mostly would say mostly okay with day mostly show parallels between the how that AI systems are biased and they also show that there is unequal representation they also show examples of discrimination harassment and so on problems in AI companies and universities that all you can read the report for this that's it's pretty interesting to read but the points I've addressed I'm not happy with yeah so that was it for now sorry this was took so long but I felt a for otech was necessary have a nice rest of the day