Beyond Biases: The True Dilemma of Tomorrow's AI?

The other day, I was asked about the potential for biases in generative AI, especially concerning diversity and inclusion (D&I) perspectives. My response? While I recognize the concern, I believe it's a challenge that's well-understood and actively being tackled. Let's look deeper into this issue and also segway into my main concern.

We've observed instances of generative AI reflecting biases—often mirroring biases present in the data they were trained on, which originates from imperfect human sources. A frequently cited example is OpenAI's DALL-E. When prompted with “photo of a CEO”, DALL-E3 typically displays four diverse images, representing different genders, races, and sometimes ages.

However, upon closer inspection, each image's prompt specifies gender, race, and age. If we remove these specifications and simply ask for "one photo of a CEO without specifying gender or race", the output is often a young Caucasian male.

This reveals two things: 1. The underlying model does have a predisposition towards Caucasian male CEOs, and 2. OpenAI is aware of this and is actively striving to reduce such biases. Are their efforts perfect? Likely not. But are they addressing the issue based on user feedback? Absolutely. And this might very well be the most biased iteration of this tool that we'll encounter.

So, should we be wary of using such models? This reminds me of a concept introduced by Ethan Mollick in his article here: the Best Available Human (BAH) standard. It involves asking the question "Would the best available AI in a particular moment, in a particular place, do a better job solving a problem than the best available human that is actually able to help in a particular situation?". The answer will vary. If you're deeply rooted in D&I advocacy with a wealth of expert resources, AI might not be the best choice. However, if you're an L&D content creator racing against the clock, AI could be a lifesaver and do a good enough job. Plus remember, the current state of generative AI is the worst we will see. And - OpenAI is aware of it and has both the means and motivation to deal with it.

So, what does keep me up at night? It's not a Terminator-type AI scenario. My genuine concern is the lesser-discussed topics, where solutions aren't straightforward and were it's not clear who to turn to. For instance, how will we utilize the immense time savings AI offers us? Peter Scott has eloquently spoken about the need for a balanced tech approach, where human well-being trumps mere efficiency and speed. If you haven't tuned into his insights yet, I highly recommend doing so here.