7. To bias or not to bias

A while ago we discussed the idea of eliminating that which makes black boxes opaque. There we found out that the whole premise of a black box is what creates the impossibility of removing that which makes it not transparent. Today we talk about another hypothetical entity in the desires of developers: The entirely unbiased AI. In a very similar fashion, we find out that creating AI without bias would result in completely fair models which are not useful anymore. Worse even, chasing the thought of a model without any bias leads us away from the path to actually making AI less biased. Let’s get into why that is and what we can do to make the technological world a better place.

To start off, let’s talk a bit about motivation here. This piece is an attempt to deconstruct the idea of entirely unbiased AI, so that we can focus on being constructive in the bigger picture of creating fairer AI. Where this seems close to paradoxical, it is often the case that what seems contradictory ends up being the most helpful when looking from a different angle. Removing the enchanting idea of an end goal to which no dispute is possible, also means removing the paralysis that comes with a goal so far away and unreachable. There is no such thing as unbiased AI and the sooner we realise it, the sooner we open up for discussion and creativity in dealing with the possibilities that we do have. Almost more importantly, this also leads to the recognition that there is no answer to be found in just an algorithm and that machine learning engineers may be as clueless towards developing fair AI as we are. Once we realise all this, we can start working on the actually important question of how to create AI that we find better, rather than trying to discover what would objectively be the best there is.

Talking about bias in machine learning starts with talking about bias in humans. In my opinion, bias is very closely correlated to objectivity, an often discussed topic in the history of western philosophical thought. One obvious antagonist of the objective view in many areas is Friedrich Nietzsche, of whom I like the following quote for its relevance to today’s topic:

“The only seeing we have is seeing from a perspective; the only knowledge we have is knowledge from a perspective. ”

Here, the evident origin of human bias is well summed up. We are products of our past and of our position, and therefore we act from our perspective at all times. We may work very hard on minimising the impact of our present bias, but this does not change the situation which is that we have a present bias. We can only see, know and therefore act from our perspective. With this I am not trying to say that we are all actively biased agents, but there still exists bias within our actions and perceptions. Consider for example something as abstract as recognising different accents from within another country than the one we were born in. Whereas my British friend might tell apart a Manchester accent from a London accent effortlessly, my abilities stop at distinguishing the Brabanders from the Friezen within my own tiny country. I might even feel more compelled to help out a Frisian person, even though this is on a level of “provincialism” that I would not try to actively invoke. However, my passive bias persists despite my best efforts.

Consider machine learning now. All our models are created by humans and trained on data sets from our human world. As we have seen before, even including the most possible data that there is results in accurately biased models. This makes us arrive at the necessary conclusion that merely allowing a model access to more data will still result only in accurate human bias, rather than eliminating bias completely. Therefore, we must move beyond our own human scope of objectivity if we are to reach the desired unbiased AI. However, there is something inherently useless to this idea.

To remove bias from a model commonly means to eliminate its reliance on certain variables. For example, consider a goal to remove a bias against people that are genetically predisposed to developing certain illnesses in a model that calculates insurance prices. This is achievable as we possess the methods to figure out which variables can be directly or indirectly learned to the problematic factor and remove them. However, this does directly influence the data which the model can access. Removing more bias will correspond negatively to the accuracy of the model.

Similarly, consider the discussed problem of removing gender biases when considering certain words - e.g. professions - in a language model. Profession-gender bias is something that we would maybe like to remove from the model, but there may be more variables that are worth considering the removal of. Dresses will most likely more often be used in combination with a female person than with a male person, even though we would never say there is anything wrong with a man wearing a dress. If we are to remove this and keep going for similar cases, we will eventually end up with a language model which has no more biases in the area of assigning relations to gender. However, it would also be a model that is much less accurate than that model which ends up being untouched by us. Aiming for completely unbiased AI removes exactly that accuracy for which we are using machine learning in the first place.

To say it once more for clarity, this does not mean that removing bias is a bad idea in my eyes. It is very important to do so even. That just does not mean that we can reach the goal of a useful and completely unbiased AI. Once we reach the human level of bias and want to keep improving from there, we must sometimes make concessions in the area of accuracy. This is similar to the dilemma that developers face when creating accurate models that end up being black boxes: Improving on human decision making skills will mean that it is harder for humans to understand what exactly is going on in the model. There is no way around such trade-offs and pretending there is makes it increasingly difficult to move forward. Now let’s see how it is that we can move forward.

Without the clear-cut destination of fully unbiased AI, the question becomes which direction we should take to create fairer AI. This can be done by still allowing for a high level of accuracy, there just need to be choices made. Think of discussions like whether we should go by equality or equity, or whether it is actually alright to discriminate based on age when considering a life-or-death scenario which is not unlike versions of the trolley problem. This is where the ethical debate comes in, and our considerations are important to help shape the future of the technological society.

As is often the case, the ethical debate is put on the map through legal documents. It is difficult to slow down possibly unethical progress without appealing to the law - although that is a different discussion - and this is recognised by large governmental instances. The EU AI Act is an enormous step in the legislation of responsible AI, and the White House designed their own blueprint of a guideline for AI development. These documents are aimed at the regulation of the creation of machine learning models, exactly because of the necessity of discussion here. When we face the fact that there exists an inherent trade-off between accuracy and fairness, it becomes necessary to consider where the line should be drawn. With these regulation documents, global organisations put a stop to the inevitable conclusion of capital-driven companies to reach more accurate and powerful models at the cost of the decrease of ethical values. That is a cynical view of course, but the 21st century warrants cynicism on the topic of capitalism.

What I find the most important thing to take away here is how current this situation is. As I write this meetings are being held, papers are being written, conferences are being held for the sake of creating a fitting AI Act for the European Union as well as for many other government organisations. The lines are being drawn and the balance is being found between moving forward and holding back. All of this will shape society, but it is also influenced by society. Artificial Intelligence is part of our world, of our nature. Therefore it is so necessary to understand what is going on and what the regulations will be that form the foundation of our society in the years to come. The first step to influence is debate, the first step to debate is discussion and the first step to discussion is creating an understanding. That is why I write this here, and hopefully why you are interested in reading whatever I put out. This is not the future or the past, it is happening right now. As machine learning continues its upward surge of amazing intelligence, it is us that should consider what its limits and objectives should be:

With machine learning comes great power, but the great responsibility is still ours.