I’ve been thinking lately about the fact that if you use generative AI content you are deciding what degree of bias you are comfortable with.
Didn’t we always do this? Why is it any different with AI? Yes. We did and this post is all about why I think it’s different with AI.
It is a challenge to be efficient in any field if you try to do things on your own. I spent a long time stuck there. I still struggle a bit with that. I’ve always wanted to customize everything I do because as I’m fond of saying, the default version sucks. This goes for pretty much everything I can think of by the way–Microsoft Office, Learning Management Systems, WordPress, vendor curriculum, continental breakfasts, operating systems, phone settings, everything! Whatever you might name, I’d probably say that too. I have spent a lot of time trying to tweak things to suit my preferences. Not everyone has the option for that degree of customization, many people just take what they are given and live with the consequences of using something out of the box.
As far as I know, I’ve been pushing back on things since I was born. In my professional career, it’s why I’ve gravitated toward alternative settings, open source software, open educational resources (OER), and now AI. It gives me some degree of control over how I, or my audience, will engage with what I make. And am I biased? Hell yes. I have all kinds of them. I realize that everything we do and create contains bias. Using something like curriculum that was largely targeted toward the customers with the deepest pockets comes with some pretty severe biases that are explicitly developed for those groups/districts/locations to keep them happy and buying their stuff. That rarely worked well for students in an alternative setting, or those with predominantly underserved populations from marginalized communities.
I get why a teacher or curriculum department goes along with the vendor stuff though. Making your own is a lot of work. Not everyone is up for that, nor has the time to figure out how to do it well. So, they accept the bias of the content they are given and they deliver that to their learners. Some make some changes and supplements to provide other perspectives, but it’s always within some limitations of the content and tools that are purchased. In some ways, that’s a lot easier to have someone to blame when a problem arises with something that doesn’t work the way we thought it would.
Here’s an example: During the pandemic, my sons were in an online school using vendor curriculum that the teachers could not edit. They saw horribly offensive lessons on how some slaves actually enjoyed their servitude, or things about Deaf culture that were horrible generalizations. Every time I saw something like this I complained, but I have no confidence that any changes to curriculum were ever actually made. I’ve been a part of a number of district challenges to content that I believe violates student rights and I’m always told that it was the fault of the publisher. Schools and teachers never had to take accountability.
One of the things that I enjoy about Large Language Models (LLM) like ChatGPT and Claude is that it makes it easier for me to customize the things I always have and it saves me many hours. In some cases, LLM allow me to make changes to things I would not have been able to before because of my skill level (PHP, CSS, web hosting).
Here’s why all of that is related…
Now that so many are in the driver’s seat creating their own materials, they are forced to choose. Which biases do we accept and which ones are important to push back on. When you have a machine do it for you, the decision is in your hands. While the implicit of LLM is frustrating and challenging, with AI, I believe we can do better. Unfortunately, I do not think most will (hopefully that’s my bias). I also don’t think there’s any way to eliminate it since values can contradict each other depending on the situation. There are simply too many to overcome.
Let’s put the big stuff aside for a moment that I typically write about as it relates to racial, cultural, gender, ability biases. There are ways to overcome that very often. What I want to get to the heart of when we think about using these tools is the fact that there are so many points of view we do not consider in our day to day lives, and those things are different depending on who we are.
If I am developing a lesson with an LLM, what do I tell it? I know that if I tell it I am a man vs. a woman I will get different results. I’ll get different results if I design a lesson for a Chicago audience based on a Birmingham one (or Dallas, or San Francisco, or Albuquerque.) I will get a difference based on my age, the age of my learners, my education level, my area of expertise…I won’t belabor that any longer, you get the idea. Every aspect of myself is something I’ll get a sycophantic response from a LLM about. I can set up a few different scenarios and have wildly different responses but some will contradict. So which biases do I accept and which do I push back against?
Some cultures aren’t into the whole separating content by subject area and grade level, some segregate genders, some prefer a rural perspective over an urban one. Think about how we program self-driving vehicles. If there is a choice to run into an elderly couple or a young mother pushing a stroller, which direction does the vehicle turn? Some cultures value elders for the experience and they are revered, some would see the potential of the young mother and baby worth preserving. The values in that case conflict. Just like the values conflict if I am explicitly trying to elicit multiple racial and cultural perspectives when teaching a topic and that diversity of opinion has been legislated against in a particular state or community.
We can do better than we ever have before through collaboration and back and forth with our AI assistants. I don’t have a solution, just questions…for you. Which biases are you okay with when you use AI? How do you arrive at that conclusion?
These are the kinds of discussions I’d love to help you facilitate in your organizations. For more on my work on AI go to Inclusive AI Strategies!
For further reading:
- Outcries When Your Name Triggers Generative AI ChatGPT To Treat You Differently;
- Google, Microsoft, and Perplexity Are Promoting Scientific Racism in Search Results;