AI ethics and biases, and the mindlessness of deep learning

Bias, as most people know, continues to be a challenge in machine learning and artificial intelligence (AI). Recall the 2015 Google scandal when its face-recognition systems tagged black people as gorillas, or the faulty camera AI software that labelled Asian faces as ‘blinking’ when they were smiling.

Understanding these kinds of bias requires both a demystification of the processes underlying them, and the political will to enforce structural and institutional changes, as has been argued by Timnit Gebru, a former Google employee who was dismissed for pointing out the lack of fairness in AI model training.

These biases, which are often reduced to race and gender, in actual fact extend even to aspects like the homogenisation of language norms because of a reduced linguistic online footprint by smaller or poorer countries.

In other words, biases can take many forms which, in computer-speak, include overfitting or underfitting, recall bias, observer bias, selection bias, exclusion bias, association bias, measurement bias and outliers. So how do these occur?

Contemporary approaches to AI

Contemporary approaches to AI differ significantly from methods that were used before 2010, which were oriented towards the symbolic representation of human knowledge via the use of explicit rules relying, for example, on semantic nets, logic programming and symbolic mathematics.

After 2010 deep learning, or deep neural networks, one family of machine learning (sometimes called AI) methods, became the dominant paradigm. Technically, AI and deep learning are different, even though they are often conflated.

To be clear, artificial intelligence is a cross-disciplinary method that makes use of computational, biological, mathematical and other theories and applications to model and replicate cognitive processes in the creation of ‘intelligent’ machines.

Deep learning or deep neural networks (DNN), on the other hand, is one approach employed to achieve this and works, characteristically, by using multiple or ‘deep’ layers of units, or ‘neurons’ in a layer, alongside computer architectures (or systems design) and highly optimised algorithms.

The revolutionary back-propagation advance

A revolutionary advancement in DNN methods occurred with the advent of back-propagation, a neural network training method in so-called ‘supervised learning’ used by a neural network to update its parameters. In other words, by propagating the blame for errors observed at the output units backwards, the initial parameters can be weighted to make a network’s predictions more accurate.

The neural network is, in some sense, ‘learning’ by using a form of reasoning that people also use, namely feedback and error correction. So in the same way a student might use feedback from a lecturer to update their knowledge parameters, so too does a neural network – and this is why we speak of ‘algorithmic reason’ or ‘deep learning’.

But what the neural network ‘learns’ through this process of back propagation is not necessarily what a human would, because the statistical methods it relies on for learning rely on the datasets it is fed and if these are biased to begin with, the neural network will blithely learn those biases and, more worryingly, represent those back to us.

This dynamic is evident, for instance, on our personal social media feeds and results in the production of filter bubbles that skew our views of reality.

Deep-learning and prejudice propagation

This can even lead to extreme political biases, often accompanied by a proliferation of conspiracy theories, because the surveillance capabilities of these new technologies can be used in sophisticated ways to harness and weaponise existing prejudices in societies and datasets.

One of the ways in which this occurs is when a network too rigidly clusters patterns it finds in the training datasets.

Imagine, for example, that the learning model has to identify and analyse pictures and photos of cats, but the model was trained largely on cats with longer hair, leading it to classify hairless cats as another kind of animal, for example a dog.

What this means is that the machine learning model has unsuccessfully extrapolated patterns from the dataset and has, as a result, failed to effectively generalise what it has learnt. This is, of course, precisely what happened with the camera software and Google’s face-recognition systems discussed at the beginning of this article.

The obvious answer for fixing this may seem simple: have better datasets. It seems self-evident that including more information from Asian, African and Global South countries, or even having Africa-specific or Global South-specific datasets, would be hugely beneficial for creating more accurate and less prejudiced AI systems.

The more complex answer, however, is that even though this kind of inclusiveness would positively affect machine learning, it does not actually address the underlying systemic predispositions in our societies, nor does it change the concentration of power, information and wealth vested in a few ‘tech giants’ – companies like Apple, Facebook, Google, Microsoft and Amazon.

It also fails to adequately address the exploitation that accompanies the correction of bias in datasets. Amazon Mechanical Turk, for example, is known for exploiting workers in various ways, including paying them incredibly low wages for the repetitive and tiresome work of labelling images that are used in datasets.

Some of these tasks include, as mentioned below, exposure to explicit and violent images, often leading to work-related post-traumatic stress disorder.

Perhaps needless to add is that while much of the development of AI and machine learning is concentrated in Global North countries, much of the raw labour for this expansion is outsourced to Global South countries.

Content moderators who review graphic media that expose them to suicides, murders and rapes, among other gruesome content, are often located in places like Nairobi, India or the Philippines, where they get paid as little as US$1.50 per hour.

The AI ethics imperative

Much of the work being done to address these issues of bias, as well as others like opacity and the concentration of wealth, are covered by what is collectively known as ‘AI ethics’.

This work is extremely valuable, but it does not sufficiently address the deeper individual and collective consequences of digitalisation on our societies, our psychological health and even our thought processes.

The late philosopher of technology, Bernard Stiegler, found these more subtle effects deeply troubling. As a result, he spent his entire career diagnosing these elusive ‘disorders’ of digitisation, which he described in terms of a kind of “generalised arrested development” that materialises as symptoms of widespread disaffection.

These include, but are not limited to, a reduced ability to experience pleasure; depression and hopelessness; a lack of concentration due to cognitive saturation; novel dysmorphias such as Snapchat dysmorphia – a body-image disorder characterised by an obsession with perfectible appearance by erasing perceived flaws through the use of filters and other feature enhancements on social media platforms; and new social phobias like ‘hikikomori’, or acute and prolonged social withdrawal, which was first discussed by Tamaki Sait in his 2013 book Hikikomori: Adolescence without end.

The point I am driving at is that while we must continue to address issues like fairness in AI and machine learning, we must also, and more importantly, address the underlying problems giving rise to these matters.

Instead of just asking how we can eliminate biases in datasets, we should also be asking why those biases are reflected in datasets in the first place, and what this tells us about the power and politico-economic arrangements that keep these from changing in our societies.

The problem, from this perspective, is more about what we think a good society is, and what the place of technology should be in that society, than simply what a good algorithm is.

As Dan McQuillan argues, the ethical challenges emerging from AI are not a result of machines or computational processes, but arise from the ways in which machine learning processes extend already existing prejudices and other tendencies in our societies.

Chantelle Gray is a professor in the school of philosophy, faculty of humanities, and chair of the Institute for Contemporary Ethics at North-West University in South Africa.