AI Recipes

Artificial Intelligence (AI) is everywhere. It has slipped into our DMs on Facebook Messenger, appeared as an uninvited guest in our phones' list of apps, and infiltrated the social media platforms that once enabled us to connect with people we know and care about. AI is helping us to drive more safely, with enhanced crash detection and lane-keeping assistance. And it's annoying us on every website we visit, popping up as the helpful AI chat bot that no one requested. In all realms and sectors, it's omnipresent.

In education, everyone is grappling with this disruptive technology. It is changing everything we thought we knew about how to effectively teach students, how people learn, what skills are and will be valued in the workplace, and how to assess competency. There are some who want to ban it. Others are attempting to proactively engage with it. It's a hot topic for debate at many tables. Recently, in an effort to learn more about this pervasive technology and keep up with the pace of innovation, my colleagues and I have begun exploring the power of various tools, like ChatGPT and Copilot, to determine when and how they might be used in an ethical fashion to work smarter, not harder. 

We've had mixed results. Partly, the quality of the results depends on the quality of the prompt, so there's some learning to be done by users in that regard. AI appears to be incredibly effective at revising draft text or summarizing an extensive body of data. It also has some epic failures -- even when the prompt is highly specific. Now, admittedly, in my free time, a lot of my experiments have been revolved around making cute pictures of cartoon bears drinking coffee using Copilot. But as it executes a seemingly simple task (compared to mastering scientific literature, for example), I've been entertained to see the results and the limitations. For example, it produced a great image but misspelled the word "chocolate" as "choccolat." I asked it to fix the spelling while keeping everything else exactly the same -- and after five attempts, it still got it wrong (or fixed the spelling but changed the image). I had an abstract that was 2 words over the limit and asked it to shorten it. It told me it had done so, but then output the exact text I'd given it with no changes. Not so intelligent, now, eh?

During one of our coworking sessions recently, we started talking about how much we missed the energy bites that once appeared at every meeting. In these days of fiscal prudence, the stockpile we once had in the freezer has long since disappeared with no replenishment in sight. We wished we could get the recipe from food services, but figured there was no way they'd give it to us. I mentioned that I've looked at many recipes online, but haven't found one that is quite right. They all have been missing one or more of the ingredients we know are in them -- chia seeds, oats, chocolate chips, peanut butter, coconut. My brilliant coworker mused, "I wonder if Copilot could write us a recipe if we told it the ingredients." We temporarily put that in the metaphorical parking lot, finished the task at hand, and then rewarded ourselves with a quick Copilot prompt. In no time it produced a recipe.

Scanning the ingredients and the method, it looked like it would work. In fact, there was only one change that I would make: instead of mixing the dry ingredients first and adding the wet to them, I would mix the wet ingredients first and add the dry to them (only because peanut butter and honey can be a challenge to combine). We considered trying the recipe sometime in the future. For me, sometime was about 45 minutes later.

What can I say? I was curious. It was also a Wednesday -- and Wednesday nights are devoted to trying new things, especially in the culinary realm. To round things out, I had all of the ingredients on hand. It was meant to be.

I followed the recipe, making only one modification in procedure (mentioned above), and was impressed by what I saw. The mixture had a good consistency and the resulting bites looked a lot like the ones we had enjoyed at work in the past. The recipe was a success in the sense that it worked. But how would they taste? 

I'm happy to report that when we tried them the next day, they were delicious and a very close dupe. In my opinion, the amount of chia seeds could be reduced. My colleague observed that you can't normally taste honey when you eat the version made by food services. Upon reflection, given the cost of real honey (and maple syrup), we concluded that they probably use a cheaper sweetener, like corn syrup. We both thought that mini chocolate chips would be better. Regardless, they are "a keeper," as Trooper Ron would say. 

While this experiment worked, I'm not sure that I would necessarily recommend AI recipes to others. Presumably, Copilot and ChatGPT have trained largely on free online content (as well as books). The quality of recipes online, though, is variable (to say the least). Sure, some are delicious. But far more yield disappointing results -- in some cases, they are just okay, while others are total disasters that don't work. You're definitely taking your chances. 

But if you're trying to find an elusive recipe, it might be worth the gamble.