The Predominance of Biases in AI

The year 2022 may be remembered as the year that generative AI burst onto the scene, producing enormous ripples in the digital community. Criticism and reaction are unavoidable with the introduction of any new game-changing technology, in this instance, the incredibly sophisticated algorithms powering picture and text generating systems such as ChatGPT and DALL-E. Nonetheless, I do not feel that the challenges and debates surrounding these algorithms should halt their growth.

The recent discussion of OpenAI’s outsourced content moderation is a good illustration of how hazy the road ahead may be.

On the one hand, human oversight is unavoidable; on the other hand, human moderators exposed to severe content should have all the assistance they need to carry out this painful yet crucial element of the moderation process. In all AI-related settings, the ideal strategy is to improve our standards while producing optimal versions of these algorithms.

Interestingly (or not), 2022 was also the year in which I published extensively about how various cognitive biases and other processes impact human reasoning, particularly in the context of full (both classic and deep neural networks-based) AI solutions.

Inadvertently (or not), the biases introduced by these generative algorithms are the primary issue that must be addressed if we are to drive generative AI in a more productive path.

What can we learn from these prejudices in the future? I’d want to use this time, while 2023 is still fresh in my mind, to go over the papers I’ve written about cognitive biases and how we might address them especially when working with generative AI.

Bias in Confirmation

The cognitive route created into our subconscious based on our views, which then focuses attention to arguments and pieces of evidence that reinforce this system of thinking, is a concise definition of confirmation bias. One of the most common biases in social media sites. This is not an easy one to overcome completely: it needs consistent practise and deliberate effort to see how they slant our reasoning.

Survivorship Bias

Survivorship bias illustrates our inclination to select only examples of success, or “survivors,” while ignoring any bad examples of a given case and what they might bring to the dataset. I’m sure you’ve seen and heard about famous entrepreneurs, actresses, sportsmen, and the tales behind them, but for every success story, there are many unsuccessful efforts. This type of prejudice is also fairly common, not only in social media, but in all media since the birth of the communication age.

False Causality

The principle that best describes false causality is that “correlation does not imply causation,” however even with this understanding, it’s still common to create a strong link between two variables not necessarily connected to one another.

Bias in Availability

The mental shortcut we employ to make rapid decisions, known as the “availability heuristic,” causes availability bias. We have a tendency to utilise the information that is most immediately accessible to our thinking, making it simpler to miss the broader picture; more often than not, you do not have the complete picture available to make your conclusion. This is due to the fact that experiences we witness or recall vividly have a significant influence on our subconscious.

Generative AI vs. AI

Many of the above-mentioned biases occur under certain cognitive situations, and identifying them provides us a significant advantage. So, while working on an AI project, how can AI specialists successfully remove them from the decision-making process?

Fundamentally, the solution to cognitive biases lies outside of one’s own cognitive processes, which means you can’t completely trust how you see the problem and how you alone will solve it. In reality, this involves placing some distance between yourself and the problem before coming up with a solution.

Obviously, collaboration is essential in this situation. Getting a second, third, or even fourth perspective will help to cut through each individual’s prejudices and strive towards a shared, objective vision. The more the diversity of the team, the better, since the value of diverse human scrutiny cannot be overestimated.

The instance of generative AI, however, offers some more challenging issues because to the tech’s grandiose aims. The algorithm need a large quantity of training data to be as robust as possible, and here is where cognitive biases exist. Each entry has the potential to be tremendously prejudiced. To mitigate any of them would need an endless amount of labour.

In the case of biases, these highly strong tools necessitate extremely powerful methodologies, such as working on the data prior to training or through some type of data filtering.

To reduce prejudice at its roots, the justification for pre-training mitigation is straightforward. The AI model’s configuration is determined by how the training is carried out and what data is used; after all, these are all human decisions. One of the safest approaches is to provide the ethical quandaries before beginning training-based development, but this substantially decreases the pace with which a corporation can produce such a tool. The usage of free, unrestricted training data was what allowed these algorithms to be so resilient in the first place.

What about data filtering for mitigation? There is a case can be made for directly applying technological breakthroughs to bias concerns and training the algorithm using filtered datasets. One recent example is connected to the fairness of representation, which is one of the most pressing issues to be addressed. Researchers at Google are working on a technology dubbed “Latent Space Smoothing for Individually Fair Representations” or LASSI. While admirable and ethical, implementing this type of data-filtering solution successful on the ground is proving to be a difficult task.

Whatever remedies we devise, it is critical that we do not lose sight of the ethical grounds for doing so. If human-created material for training and human moderation appears to be an important component of the project, the attention should be kept on the procedures that are prone to human error. The technologies being developed have incredible superhuman powers, but it is the human component that will make it relevant to our needs and cognizant of our flaws.