AI researcher and artist Joy Buolamwini’s video poem ‘AI, Ain’t I a Woman?’

One suggested remedy for white privilege is to be actively anti-racist. For the 1998 head-tracking system, it might seem obvious that the anti-racist remedy is to treat all skin colors equally. Certainly, we can and should ensure that the system’s training data represents the range of all skin colors as equally as possible.

Unfortunately, this does not guarantee that all skin colors observed by the system will be treated equally. The system must classify every possible color as skin or nonskin. Therefore, there exist colors right on the boundary between skin and nonskin – a region computer scientists call the decision boundary. A person whose skin color crosses over this decision boundary will be classified incorrectly.

Scientists also face a nasty subconscious dilemma when incorporating diversity into machine learning models: Diverse, inclusive models perform worse than narrow models.

A simple analogy can explain this. Imagine you are given a choice between two tasks. Task A is to identify one particular type of tree – say, elm trees. Task B is to identify five types of trees: elm, ash, locust, beech and walnut. It’s obvious that if you are given a fixed amount of time to practice, you will perform better on Task A than Task B.

In the same way, an algorithm that tracks only white skin will be more accurate than an algorithm that tracks the full range of human skin colors. Even if they are aware of the need for diversity and fairness, scientists can be subconsciously affected by this competing need for accuracy.

## Hidden in the numbers

My creation of a biased algorithm was thoughtless and potentially offensive. Even more concerning, this incident demonstrates how bias can remain concealed deep within an AI system. To see why, consider a particular set of 12 numbers in a matrix of three rows and four columns. Do they seem racist? The head-tracking algorithm I developed in 1998 is controlled by a matrix like this, which describes the skin color model. But it’s impossible to tell from these numbers alone that this is in fact a racist matrix. They are just numbers, determined automatically by a computer program.

The problem of bias hiding in plain sight is much more severe in modern machine-learning systems. Deep neural networks – currently the most popular and powerful type of AI model – often have millions of numbers in which bias could be encoded. The biased face recognition systems critiqued in “AI, Ain’t I a Woman?” are all deep neural networks.

The good news is that a great deal of progress on AI fairness has already been made, both in academia and in industry. Microsoft, for example, has a research group known as FATE, devoted to Fairness, Accountability, Transparency and Ethics in AI. A leading machine-learning conference, NeurIPS, has detailed ethics guidelines, including an eight-point list of negative social impacts that must be considered by researchers who submit papers.

## Who’s in the room is who’s at the table

On the other hand, even in 2023, fairness can still be the victim of competitive pressures in academia and industry. The flawed Bard and Bing chatbots from Google and Microsoft are recent evidence of this grim reality. The commercial necessity of building market share led to the premature release of these systems.

The systems suffer from exactly the same problems as my 1998 head tracker. Their training data is biased. They are designed by an unrepresentative group. They face the mathematical impossibility of treating all categories equally. They must somehow trade accuracy for fairness. And their biases are hiding behind millions of inscrutable numerical parameters.

So, how far has the AI field really come since it was possible, over 25 years ago, for a doctoral student to design and publish the results of a racially biased algorithm with no apparent oversight or consequences? It’s clear that biased AI systems can still be created unintentionally and easily. It’s also clear that the bias in these systems can be harmful, hard to detect and even harder to eliminate.