Forbidden fruit of AI

SERVICES > FISH Academy > Post 1. ~ 5 minute read.

Artificial Intelligence (AI) is changing the way we do things. It can help us be more efficient, pick up errors, and help us get started with new ideas.

There are a whole bunch of topics we can unpack with AI. What is AI? How do we best use AI? How are other people using AI? What tools are best? What happens to my data? What about privacy in AI? What about AI governance?

There are many things still being developed in AI, and the true impact of AI is just around the corner. I won’t be going into those today… Instead, I will focus on Unconscious Bias in AI.

AI sometimes makes me laugh

There is a video of a 12-inch sub sandwich being cut in half by "AI". The AI tries to make a cut along the length, creating two thin 12-inch subs (as opposed to two 6-inch subs). The instructor says, "No, turn it around," and the sub is rotated 180 degrees. The flat bottom part of the bread is now on the upper side, but it is still in the same position. Again, the AI tries to cut it in half, producing two thin 12-inch subs. The instructor sighs. The audience laughs. Instructions and communication are important and AI is only as good as the instructions and training data.

A thought experiment

Today I wanted to talk about a simple topic: the Forbidden Fruit of AI. Well, not forbidden, but something we can minimize: Unconscious Bias.

If I ask you to think of an apple, chances are you will think of a red one. You probably know there are several varieties and colours of apples. But why do we think of red ones first?

Now go to an AI (really, go try this) and ask it to generate an image of an apple. Don't mention colour. Was the apple red? Why this happens is complex: culture, history, books, films and more. Many times, the “forbidden fruit” from the Bible (Adam and Eve) is depicted as a red apple. But nowhere does it say it was an apple, or that it was red. That is where my thought experiment came from.

Discussion

People who think of a red apple, and AI that generates a red apple, are not wrong. But it is technically incomplete. Statisticians call this unconscious bias. Perhaps it is not that important if we are grading apples in our backyard. But if we entrust AI to make decisions that impacts the community, it may include biases and have a significant impact on outcomes.

You might say, "But I have never asked AI to make a decision." If you have used AI, you probably have. You asked it to generate an image of an apple. The AI "decided" (not really a decision) to construct a red apple of that size, with that background. That is at least two decisions. AI is trained on existing information. If the training information is out of date, incorrect, or not created by experts, the AI might make odd, biased or generic recommendations. Just like AI doesn't "know" how humans prefer to cut their subs in half; it tries to follow instructions of the user. However, AI can be trained and it is improving weekly.

Subject Matter Experts (SMEs) are often aware of these issues and can work around them or check for them. A non-expert might not notice the issues. It is the reason why you go see a medical doctor instead of looking up your symptoms online and self-diagnosing. Again, AI may have been correct or partially correct.

My experience with AI

AI is fantastic, except when it's not, and caution is only natural. "Proceed with care," I would say. But don't throw out AI altogether. As a statistician, I use AI, with care and caution. I have had instances where I was researching a topic and I asked AI how it came to a conclusion, and it told me that it had made up data. Another time, I asked AI to analyse a small survey of 20 responses, and it kept saying the sample size was 19 and all the averages were incorrect by a little bit. But luckily, I could fix it. Like the sub problem, the issue was obvious to me. Other times, AI has saved me hours, if not, days of work.

Good statistics doesn't necessarily eliminate human (or AI) bias; it helps us identify where bias may be entering a decision-making process and test whether it is actually affecting outcomes. AI can be used for decision-making. But engaging an SME might be able to reduce unconscious bias (and other bias) and improve decision making.

Beyond AI, consider addressing your own personal unconscious bias in any decisions you make, with or without AI.

See Unconscious Bias in the Workplace for further reading.