AI bias at its worst.
AI bias showcases suppressed prejudice in the name of political correctness.
Beauty AI, created by a group called Youth Laboratories (backed by Microsoft), organised a beauty competition which was judged by AI, the process was simple, you had to download their app, upload a selfie without a beard, no make up and no shades, and the AI algorithms would do their thing and release a list of winners.
Over 60,000 people from 100 countries participated, and eventually 44 were declared as finalists, the catch? None of the winners included a person of colour, a few, 6 to be precise, were Asian and only one of them had dark skin.This sparked widespread controversy and debate, with the company’s chief science officer saying that the main problem was that the data the project used to establish standards of attractiveness did not include enough minorities.
The algorithms seemingly evaluated user-submitted selfies on the basis of youthfulness, skin quality, symmetry, and appearance relative to databases of models and actors. Winners were broken down by age and gender.
The argument that the training set was biased and in turn the resulting algorithm turned out to be biased can’t be taken seriously, they surely must’ve thought about this way before they started the experiment, there’s no way they just ran an algorithm on a bunch of images and deemed it fit to judge a global competition, and if that was the case, it is even more concerning.
Datasets are biased, and this has been the case for a long time, consider the Boston Housing Dataset, one of the most commonly used datasets by beginners to start their data science data exploratory analysis journey, this data consists of housing price data in Boston. This dataset contains a column which is labelled ‘B’, it represents results based on a formula run on the number of black people in the respective town, thus implying that more black people in an area would negatively affect the price of the house in that area, this dataset was one of the built-in datasets found in scikit-learn library, it has since been deprecated and removed all together.
This might seem harmless now, but we should be mindful of the fact that the algorithms that might be substantially powerful in the future are being trained on racist datasets and being developed by people who are unintentionally programming further bias into these algorithms.
An investigation found that software used to predict future criminals is biased against black people, which can lead to harsher sentencing, adding to the already existing bias against black people in criminal justice system.
The bias in algorithms isn’t just limited to beauty contests, there are numerous cases that have come up, a few years ago Google’s photo app was found to have labelled black people as gorillas, or Facebook algorithms being accused of being political biased, or going absolutely beserk or Flickr’s auto tagging system coming under scrutiny after it labelled images of black people with tags such as “ape” and “animal”. The system also tagged pictures of concentration camps with “sport” or “jungle gym”.
The interesting thing to note is that eventual deliberation, Google decided to drop the gorilla tag altogether, Google with all its proficient programmers and access to a huge variety of data, could not fix the issue, and these are a few biases are openly highlighted and adequately reported, what about the issues that are still hidden and are found after its already too late?
It is not just AI that is incorporating racial bias or bias of some kind for that matter, For years, Kodak used a coating on its film that favoured Caucasian skin tones, making it more difficult to shoot darker skin. Nikon and other consumer camera companies have also had a history of showing bias to white faces with their facial recognition software.
Thus, the question comes down to the fact that we’re suppressing bias in humans without adequately addressing it, racism is taboo and nobody wants to talk about it, some even failing to acknowledge the existence of it.
With social media becoming extremely sensitive without outrages over the smallest of things and the very nature of social media is very polarising, with the existence of recommendation systems that suggest news articles or content pieces that make the user interact the most, irrespective of validity of the piece. This is leading to people becoming overly politically correct on social media platforms, and this conceals the actual extent of the problem which then manifests itself in various ways, some of them more harmful than others.
One of the solutions is to make all sorts of discrimination part of mainstream discussion and education, biases are a result of stereotypes and deeply personal experiences which can almost never be generalised.
One more thing that must be done is the analysis of racial bias or bias of any kind for that matter in datasets before they are used in training of algorithms irrespective of the application of the trained algorithm.
We’re currently at the stage where the technology is at the cusp of rapid advancement. In 10 years the world will be very different from today, most things that function manually today will be run by algorithms, and for proper functioning of these algorithms, we must make sure that these algorithms are trained properly, and we have adequate enforceable regulations in place for that. Amongst all this uncertainty, one thing is for sure : we have a long way to go, both humans and robots.
