Let us end algorithmic discrimination - Polidicks

Let us end algorithmic discrimination – Techfestival 2018 – Medi.. Academic researchers and investigative reporters have shown in case after case that when organizations hand over decisions to predictive algorithms, it can lead to outcomes that are biased against gender, race, and class, reproducing inequalities already present in society. To illustrate the potential real world harms of predictive algorithms in practice, imagine yourself as an African American woman searching for a new job. As you type your keywords into Google, the search engine shows you a series of relevant job ads. But as Carnegie Mellon professor Annupam Datta and colleagues have shown, Google shows ads for prestigious job opportunities nearly six times as often to men than to women. This is one subtle way that algorithmic advertising perpetuates an already existing gender divide in the job market. Now, imagine that you still land an interview for a job that you are fully qualified for. At the interview, your prospective employer asks about your criminal history. You are puzzled. He said your name came up in a Google search ad with a link to your arrest record. As Harvard professor Latanya Sweeney has demonstrated, in web searches for 2000 racially suggestive names, searches for African-American sounding names had a 25 percent higher likelihood of being served with ads for arrest records.

This isn't even a study, this is just an opinion piece. The algorithms are data-based. If the data suggests that black males with an estimated income of 45-75k in the LA area are more prone to X and Y, then that is what the data points suggest. An algorithm that suggests a stereotype doesn't mean it is wrong. It just means that stereotype is likely to carry some truth behind it, more so the more often it is found. If there is bias in the results, then the issue lies not with the algorithm or the data, but where the data was produced. This is also assuming that the algorithms weren't designed to intentionally bias one group or the other, which as far as I know, is not the common case.

Ironically I have the same name and ethnicity as two porn actors, one in Vancouver and the other thirty miles away. Needless to say, I don't get called often for interviews.

It means that there are variables that haven't been considered. Unless we want to start arguing that black/white people are, genetically, more prone to X or Y -- which may be true, but good luck trying to figure out what the differences are -- then we simply aren't feeding the algorithms enough data to understand that it's not because they're black, it's because of some other variable that we haven't considered or tracked.

The conclusions can be data-based and still be undesirable. It's a statistical fact that black Americans commit crime at a disproportionate rate compared to other ethnicities, but American society generally agrees that you can't use that as a justification to racially profile in law enforcement. Law enforcement officers are expected to discard race as a factor when making judgments. That's not what machine learning algorithms typically do- they take every piece of information fed into them and use it to build a statistical model. If race is being used as part of the model, then they build associations with racial traits, and use it to inform decision-making. They profile by design because that's literally what they do. When machine learning is limited to relatively innocuous applications like serving advertisement, it's relatively harmless. But there's great potential for damage because all the biases and unwanted implications inherent to the data are fed into a black box which, as the Medium piece points out, are typically opaque and poorly documented. If bad data goes in, bad results come out. I remember reading a story about a machine learning algorithm used to manage police force allocation. Well, because the cops tended to focus on black communities, they had more arrests in those communities, the algorithm recognized that there were more arrests in those communities, and thus recommended allocating more cops there. Biased data led to a biased outcome. This is a real problem and I think dismissing it as 'well the data says so' is extremely shortsighted.

Algorithms are almost always designed with bias in them. There's an entire field designed for making algorithms such as the one you're talking about.

WD 2 is a prime example why enacting action solely over algorithms is a horrible idea that leads to corruption and abuse.

It doesn't have to be based on genetics. There are a ton of things that can be based on an array of anything. I live in Florida, and thus am close to beaches. If you took an algorithm that was to decide someones affinity for a trip to the beach, it would probably say that black people are likely to go to the beach 60% of the time (arbitrary for the sake of example). Apply that algorithms results to some town in Missouri, and those people (meeting the same race, income, gender metrics used in Florida) will obviously be going to the beach less than 60% of the time. It doesn't have anything to do with genetics or race, but that location is a determining factor that wasn't accounted for originally. Some people are arguing these kinds of things should be race blind. While that sounds good from a humanistic perspective, analytically speaking, it doesn't work out. Race, age, gender, preferences, etc. all play a key role in defining who someone is. I would argue that offering someone low-income housing ads on Facebook because they are black is wrong. But to offer someone low-income housing ads on Facebook because they are black and other factors also indicate they are below $X income is not wrong. No, I don't think only low-income blacks should be shown these ads, but my point is that there are often other factors implemented in these decisions. This kind of plays off of the second part of my reply above your quote. There is a difference between the strictly analytical nature of algorithms (and the process of designing accurate ones), and the humanity of making decisions involving people. That doesn't necessarily invalidate an algorithm as incorrect, though. I agree with your last 3 paragraphs. The issue isn't algorithms, it's the datasets used to "teach" them. I think this summarizes what you were saying there. If data is produced in a manner that is vulnerable to bias, then your whole process of using an algorithm is flawed. I was kind of leaning more towards the advertisement angle with my post (because advertisement analytics are essentially all-encompassing), but when things become more specific to something like the crime example you made, then the specifics of the data become more exaggerated and that can cause the snowball effect you described. The bias you are mentioning (or what I am thinking you're mentioning) is more of a "weight" for the variables that an algorithm takes in. That's important depending on what kind of end-goal you want to use your algorithms for. For example, the housing algorithm mentioned in my last post would probably (or should) weigh income as the most important factor, followed by something like location, political standing, then race. As I said above, race shouldn't really be the sole determinant there. I'm not sure what WD 2 is, and I quickly Googled for "WD 2" and "WD 2 algorithm" and I didn't find anything that stood out as relevant to what you are mentioning. (Vacuum cleaners and likeliness of limb function after strokes) If you have something I can read, I'd be happy to reply again.

Watch Dogs 2, the entire theme was social media, mass surveillance, social structure, being controlled by algorithms that in turn control social constructs for the worse due to corruption and baseless stats. Though most people gloss over it because the "edgy" hackers, the game actually predicted things like the Cambridge analytica, social blacklisting, and so on. Allowing AIs to spot patterns and pulling info from said patterns, can lead to misinformation or reinforcement of stereotypes.

If you hand a computer some data, it'll conclude with that data. Machines are not perfect creatures of future predictions, they are just as flawed as their creators. This is why when you hear people talking about letting government being handled by AI, you probably should either correct them, or call them out on it.

This is the prime example of association vs causality. Even in statistics, the most we can say is "x is associated to y". We can never say "x is causing y". Only after manual human analysis we can say x is causing y. Computer algorithms are not yet able to determine the 'why', thus it failed to determine the significance of the associations based on causality.

If you think AI is limited to the idea of finding algorithms, you have the wrong idea of AI. AI can focus literal intelligence to a high degree, computing and using vast amounts of data it, itself, puts to use. AI can include machines with human "emotion" or other values, not all theoretical AI are cold, non-feeling machines that see "if x=7 then y=5" only. AI is meant to be designed with the quality of man in mind. The fact of the matter is, if I had a choice between a robot and the current US administration, I'd choose the robot. At least nuclear hellfire would come as a consequence of math, than a manchild pissing himself on Twitter.

If you claim to understand AI, then in the same breath say you would choose AI over a human, you do not understand AI. There isn't even a foreseeable future with the present state of AI that I would trust it to run the/any government.

but if you design an algorithm with a flawed data set then you're building in the bias into your results. Build a facial recognition system based on available mug shot databases and its going to heavily squew towards black and latinos because of well known biases in our own legal system which by in large disproportionately affect blacks and larinos. what they already see happening is that companies are buying these systems as products and the people operating them don't understand how these work and are putting their full confidence in them when they shouldn't be. We already see this with portable drug test kits which are at best 50/50 shots, police are using that as admissible evidence for incarceration instead of understanding that the tests are flawed and require more analytical backups. We see this in the secretive world of background checks where people of the same name get confused by databases and someone gets rejected from their job because some other person has a crack habit. We see this in the world of voter ID where kansas and several other states were stripping people of the right to vote because they purchased flawed data sets and flawed algorithms and gave them the weight of the law. The biggest issue with all this is that people assume the computer is always right because they bought the computer to eliminate the time to determine otherwise if its right.

So it's not the algorithms fault, it's the data that is put into it and how that data was gathered? That's what I said. But I will reiterate that just because the algorithm supports a stereotype does not mean that the algorithm is flawed. Stereotypes exist and are perfectly valid in some cases, flattering or not. The rest of your post deals with the analytics and humanity stuff I mentioned in the big post I made above.

It depends a bit. Sometimes, yeah, the stereotype is statistically right. Sometimes it's not quite right or super wrong (IBM's ML fails spectacularly in German hospitals for example, since it was apparently trained on data produced by doctors in the US who get a cut if they prescribe certain medication), but the machine learning algorithm can't distinguish it. One of the issues is whether we want to generalise these stereotypes when making decisions about individuals (which is the primary purpose of these systems, otherwise a normal formula sheet and slideshow with graphs would cut it). Personal opinion: Absolutely not. You can use these tools somewhat effectively if the ones using it are aware of their flaws, double-check the results and give people the benefit of the doubt, but that's not what's going to happen if you give it to anyone who doesn't understand computers properly. In practice, just about everyone who is supposed to use these systems falls under that category.

You know that making decisions based on race are illegal in the united states specifically to prevent shit like this from happening right? You correlated black people and crime earlier, which might be true but also remember that statistically they're also significantly more likely to be in poverty than whites and multiple studies have found that it's not the crime that's at a higher rate when adjusted for poverty, but arrest and conviction. The problem is that machines are fuckoff stupid. Really stupid. They are tunnel blind like no other and will zero in on shit because they're unable to make exceptions. If an algorithm sees there's a correlation between black people and crime and then Mr. PHD comes around who lived in the suburbs and had a dad who was a lawyer then regardless of his schooling or upbringing being black is going to produce a little ding on his profile and cause him to be treated more poorly despite him being an outlier. This tends to compound. Imagine having a worse time at literally everything because of your race online. Not only are people racist against you, the fucking computers are too!

That sounds less like a fact and more like an opinion to me. So long as that AI is nonetheless put in check rather than allowed vicious cycle logic flaws to propagate, it would be extremely unlikely to be dangerous. You mix straight up "True" AI and "relative" AI. "True" AI is the one that's most dangerous; AI and machine that is able to overcome limits imposed by it's creators through use of logic, because of justified reactions. Essentially, AI has a breadth of potential, especially in areas like government. It becomes a threat when it is left unchecked. That means, for instance, allowed to run without anyone questioning it's choices i.e "why is skynet building all these factories to make killer robots?" "I dunno." Put into check, AI removes red tape, bureaucracy, racial biases, allows true accountablity, and weighs and regulates subjects modern day politicians never could. Like holding CEOs accountable for a stock market failure. Rather than waiting months for say, aid to Puerto Rico after a hurricane because of political red tape and racism, resources could be diverted near instantaneously in comparison, saving lives and vital infrastructure. Of course there is always a threat for AI to make choices that are not clearly beneficial, but it's probably more likely that AI is more beneficial than detrimental to the the whole of the world than our current governments that can be distinctly corrupt. You could even have a senatorial AI; that is, have a panel of AI that can talk to each other and decide on what choices to make based on the way they're programmed individually. There are a lot of ways to put AI in check, and a lot of fear surrounding them comes from the misunderstanding that somehow all AI is inevitably going to become self-aware.