The ghost in the machine: AI biases and hallucinations | Opinion

By Ali Al Mokdad

The ghost in the machine: AI biases and hallucinations | Opinion

Over the past few months, I have been exploring the world of artificial intelligence (AI) and using different tools and platforms on a daily basis. In doing so, I have gained a better understanding of both the potential and limitations of AI as well as the public’s concerns about it.

I found myself sharing similar concerns to many other users regarding data protection and privacy and the ethical use of AI. However, what worries me the most is the alarming lack of representation and diversity in the design and rollout of AI.

I have noticed that there’s not enough diversity in either the data or the teams that create and deploy AI, and this is causing current AI systems and those of the future to increase existing biases as well as access constraints.

In this article, I will highlight this issue to raise awareness about these risks and hopefully inspire you to become part of the solution.

Representation in AI: A reality check

Representation, inclusivity and diversity in AI design are not simply a matter of fairness— commonsense dictates that AI systems should be designed to serve us all, which presumes that AI needs to understand and reflect the diversity of all users in all countries and from all backgrounds to achieve its purpose.

However, the reality is that the AI industry is largely dominated by a narrow demographic. This lack of diversity is not just a theoretical concern—it is a problem that has real-world consequences and we are already seeing the impact of these biases in many areas from facial recognition software that struggles to accurately identify people of color to AI chatbots that give stereotypical answers and generate false data or analysis along with the lack of cultural sensitivity and language diversity that exist in AI systems.

You would be surprised if you looked at the data that reflects diversity in the AI industry. Even though the giants in the industry, such as Google and Facebook (Meta), employ up to 22% women, the general overall picture is very sad. I read a 2021 report by Wired that indicated only 12% of AI professionals globally were female and I felt frustrated just imagining the numbers behind racial or ethnic diversity in this industry (the Wired report does not provide those, unfortunately).

Another study published in Nature revealed that an AI algorithm used in U.S. hospitals to allocate healthcare to patients was systematically discriminating against black people because the datasets used to train these systems were not sufficiently diverse. I believe this is just one example of how a possible lack of representation in AI design can lead to real-world harm.

A report by CBS News highlights how the AI chatbot, ChatGPT, has been criticized for its output, some of which is nonsensical, factually incorrect, and even sexist, racist, or otherwise offensive.

I have found myself watching many videos and reading plenty of articles about the future of this technology. Yet, throughout this journey of exploration and reading, I have often felt a disconnect because the voices I heard and the perspectives I read did not truly represent me, my country, my culture, my colleagues, my friends or the sector that I work in.

I encourage you to do the same: explore, watch videos, read articles, and ask yourself, “Do I feel represented?” More often than not, you might find the answer is “no”.

This lack of inclusivity and diversity in the AI content in the media is not just about who is speaking—it’s about who is being heard, who is being considered and, ultimately, who is being served.

AI hallucinations

When I explain AI to my friends, I often compare it to a child (I even call Microsoft Bing “baby Bing” and I call ChatGPT “a teenager”) because it is growing and it learns from the data it is fed and, just like a child or teenager, sometimes it can make things up.

Experts call this ‘hallucinations’ which is when AI makes a claim and generates information that simply is not true. It’s like a child making up a story based on its beginner-level understanding of the world and this happens because AI lacks the context for the questions asked or the statements made. And, just like a child, it will understand and learn what you teach and the data you give.

See also: Should AI training be put on pause? Why or why not?

What I am trying to say here is that these hallucinations and biases are not inherent flaws of AI, but rather reflections of the data it is trained on. If the AI reads a lot about management, and the text mostly mentions male managers, then its answers will assume that most managers are men. And if the AI’s data on humanitarian intervention primarily comes from Western perspectives discussing the Middle East or Africa, then the AI will generate data with the assumption that humanitarian intervention involves Western individuals supporting people in Africa or the Middle East.

My advice to you is to make sure to check and review the data that you receive from AI or chatbots before considering it to be fact because it could be hallucinations or maybe even biased.

What’s next? How do we deal with this issue?

The clock is ticking, and we need to act swiftly. The field of AI is rapidly evolving, and it is important that we stay informed and engaged by establishing processes to mitigate bias in AI and engage in fact-based conversations about potential human biases.

This is not a task for a single individual or organization—it requires collective action and collaboration and we need to consider how humans and machines can work together to mitigate bias and fix this problem. This may involve investing more in bias research, providing more data, and taking a multidisciplinary approach or it may mean that you as a user correct Chatbots when they are being biased.

We also need to diversify the field of AI itself. A more diverse AI community will be better equipped to anticipate, review, and spot bias which means encouraging tech companies to prioritize diversity in their hiring practices, particularly for roles related to AI development.

The companies leading the industry should be aware that it is not just about creating more advanced AI. It should be about creating AI that understands and serves us all.

I think, in the end, the question is not whether we can afford to ensure representation in AI, it’s whether we can afford not to, so let’s work together to ensure that AI becomes a tool for all, not just a privilege for some.