(How) do you bias?

These days, AI is the magic hype word. A simple statement like, "We are heavily investing in AI" may boost the company's stock value. A correct AI application may very much do that and create a competitive advantage. Or, it may just as easily have severe adverse effects on your business if implemented without a proper foundational strategy.?

When the Institute of Electrical and Electronics Engineers (IEEE) interviewed a range of big brain technologists and futurists to weigh in on the future of AI by answering the question, "When will we have computers as capable as the human brain is?" their answers ranged from "this will happen in 2029", through vague "soon", to "within a small number of decades" and "not in our lifetimes".

Could it be that most of them were wrong? Perhaps, but it also depends on your definition of AI.

What does "as capable as the human brain" truly mean?

Does it mean being capable of doing complex computations, connecting multiple dots, selective memory, making flawed assumptions, biased decisions, irrational decision-making, etc.? That is how humans rationalize and make decisions, for the human brain is infinitely complex and, at the same time, flawed, or to be more precise, individually subjective.

This individual subjectiveness is perhaps the most human-like trait and the very foundation of our decision-making process. That is where things become slightly controversial and much more challenging because this is the source of one of the biggest problems of today's commercially available AI models - AI's susceptibility to bias. The modern, most prevalent, approach to mitigating biases in AI models is to use as diverse a training team as possible. But is this truly the best approach?

Bias is unavoidable, although not necessarily a bad component, of every human-like decision-making and behavior, and as a result, it is inevitable in all AI training models, supervised or unsupervised.

That is also true for the large language learning models, like #GPT-4 and similar, which are used for conversational AIs, like #chatGPT and #Bard.

In a conversation with IBM's Vice President of Data & AI Product Management, Shadi Copty, I learned that the biggest hurdles preventing AI adoption in business decision-making are:

·???????? Costs of data & training – i.e., cost of compute power needed for AI training models,

·???????? Shortage of skills – i.e., people with advanced math skills,

·???????? Trust in AI – how credible is the output that AI produces?

While the first two are more tech-financial problems that we know how to solve, the third problem is techno-behavioral.

It is, therefore, important for managers and AI adoption teams to understand and to keep in mind the very nature of bias and what can be done to minimize its negative effect on the AI training process and subsequent decision-making.

But why only the negative effect of bias? Why not bias overall?

Bias is not necessarily bad. Cambridge Dictionary defines bias as "the fact of preferring a particular subject or thing." As Robin Hauser noted in her bias-related TED talk, "I cross sides of the street when I see a big scary dog coming my way." At its core, bias is a survival technique. Moreover, in the form of a subjective opinion, bias is the root cause of a decision being right or wrong. In other words, when we rely on someone's expert advice, we do not want their objective opinion. Quite the opposite. It is their subjective experience that makes them experts. This subjective experience is based on a personal, biased, perception of the inputs that an individual has processed and shared as their personified output.

The problem is when our bias interferes in a harmful way with the way we interact with society. In business, this interference leads to inconsistencies, resulting in a faulty decision-making model.

It is extremely difficult to un-bias our personal views, and we should not do that, but we should filter the bias to only those subjective constructs that are purposeful for the current decision-making process.

Examples of strong AI susceptibility to bias show the importance of human nature as a critical training component.

If we add a human factor to our AI learning process but do not account for human nature and biased behavior, we are on our way to losing control over AI. Remember the infamous Microsoft Twitter AI chatbot Tay that started tweeting messages harmful messages? Or Amazon's failed attempt of feeding the AI with candidate profiles and letting it decide who the best hires are, resulting in AI selecting predominantly white males? How about when we trusted chatGPT before we learned, some of us the hard way, of its ability to hallucinate and then confidently persuade us of the correctness of its innately incorrect output?

But, AI bias is not only about racial, cultural, or demographic profiling. These are just biases that we are aware of. While we can employ quality control mechanisms to verify AI's outputs, the real question is whether we can avoid the harmful effects of AI bias altogether. If the input is flawed and biased, those biases will be replicated and contaminate the output. To use the phrase: bias in, bias out.

This is where things get complex and challenging. We, humans, carry essential traits like caring, empathy, innovation, sharing, and ingenuity, all of which are hard for AI to grasp. We are able to recognize harmful biases by engaging empathy, social intelligence, or other forms of higher cognitive skills incomprehensible to AI.

AI cannot replicate human decision-making because the human brain is vastly more complex than anything we have ever programmed, and science we still do not understand how human decision-making works, replicating human-like decision-making that involves empathy, improvisation, innovation, and other essential traits of the human brain is out of reach for commercially available AIs. Moreover, AI cannot apply previous experience and training in solving new problems. Unlike AI, learning from experience is something we humans are exceptionally good at; we are impervious to the so-called change in variance, while AI is not.

The implications can be far more dreadful knowing that the modern AI is deciding who will get the loan or who will get the job. In China, AI is profiling citizens based on their interactions with other individuals by assigning them scores, and in the US, the criminal justice system is using AI to assign risk scores to defendants in order to automate the risk assessment that the defendant is posing to society.

If you think of it, building an AI model is similar to raising children. There are trainers (parents) and there is the society, i.e., the environment from which the trainee gets inputs (datasets). While trainers' job includes establishing a baseline for rights and wrongs, society and social interactions are essential for every individual's social and psychological development. This process takes years and we still do not fully understand all causal effects of an individual's social development. The nowadays prevalent unsupervised AI learning further emphasizes the problem of the inadequate dataset in the training process by allowing AI models to self-learn about the causal relationships from the internet.

While we grow and develop as individuals, we paint the picture of the world around us by analyzing environmental inputs, e.g., society's values and learning through personal experience. Our understanding of the world changes and evolves as we continuously analyze new inputs over time. Much like us, an AI model evolves through input diversity. However, the result of our input analysis is subject to our personal biases – our previously built and established understanding and value system.

If our own decisions are, by their nature, subject to bias, how can we ensure that we are making the right call? One way we humans do this is by applying various environmental control mechanisms. These mechanisms include consulting others for opinions and advice, and subjecting our input to people with different backgrounds to get a fresh perspective. The process repeats until we come to an output that we find the most acceptable, which is, again, a biased and opinionated decision.

Translated this problem to the AI training process, the increased diversity of the training team that is supposed to set boundaries will inevitably bring in new biases, so the question remains: How can we ensure that our AI training model is as resilient to bias as possible? While trainers may teach AI to look past the obvious biases such as skin color, gender, etc., they cannot train AI to ignore biases they are unaware of. It may take days or even years before we realize that our AI model is fundamentally flawed.

One solution to this problem is the Deliberate Reduction of Input (DRI) method that my team and I worked on – feeding AI with only data types that are strictly relevant to the situation for which we require AI's input. In other words, instead of feeding AI with large quantities of data varieties, the DRI purposefully reduces data sets, focusing them on a particular problem to achieve more precise model definitions. Below is the schematic representation of the loan-granting AI model example with and without DRI.

The bottom line is that there is no cookie-cutter solution to the problem of AI bias, as we cannot exclude bias from our own decision-making cognitive processes. This is why decision makers must not overly rely on AI; instead, the AI should be a source of structured data and data turned information that decision-makers can purposefully use as a baseline for their decision process.??

?

?