Learning to talk to AI

I don’t know about you, but I’m feeling a bit of AI fatigue these days. I think the fact that “South Park” is actively critiquing our societal relationship with ChatGPT (Season 27, episode 3) tells us a lot about the cultural moment we’re all having with AI. But lest you think I’m a complete AI cynic, I can say that like a lot of technology, there are upsides and downsides to the AI revolution.

In order to unpack the larger topic of AI and delve into some of the nuances, like what works and what doesn’t, we need to start with a concise definition of what we mean by Artificial Intelligence. Modern AI learns patterns from data and uses them to predict the next word, label, or action. What happens next is complicated and not even fully understood by the mathematicians and Ph.D. neuroscientists creating the models. Most popular tools are trained on a large, fixed snapshot of information from publicly available data sources with cut-off dates (ChatGPT was June 2024), and they are powerful pattern matchers that can be at once impressively helpful and confidently wrong.

What this really means is that most of us are using AI wrong. We’ve all been trained on Google search, and so most of us tend to replicate the same habits over to our use of AI, treating it like a smart search engine, and that’s where things go wrong. For one thing, AI models don’t know anything about current events (see cut-off date). But if you ask ChatGPT about current events, the model recognizes it needs current information and searches the web, then attempts to blend those search results with its trained knowledge to predict what you want to hear.

Did you catch that? “What you want to hear.”

Let’s go back to what question you asked, what parameters you gave the model, and how many details you provided the model to get a useful answer. Did you ask Chat GPT a question like you usually ask Google, or did you elaborate on your goal, your scope and constraints, and your quality and verification requirements? This is where we differentiate between a search and a prompt. Treat it like a junior teammate. Give it a role, a goal, some constraints, and figure out how you’ll check the work.

AI is the epitome of the garbage-in, garbage-out engine, and the field of prompt engineering is what minimizes the noise, focuses the AI model on what you need, and helps avoid hallucinations or completely off-base answers. And the problem is, when you’re researching something you don’t know really anything about (isn’t that why we’re supposed to be using ChatGPT in the first place?), it can be hard for the human in the loop to discern the accuracy of the answer.

The good news is, AI is pretty good at writing its own prompts.

For example, on a recent business trip to Philadelphia, I needed to figure out how long it would take me to get from my hotel to the airport on public transportation, so I could gauge my departure time in an unfamiliar city. Now, for that a quick Google search would tell me some things, but it would take several steps, and I had other requirements. In my case, I wanted to find the best Philly Cheese Steak I could while en route to the airport, and then figure out what time to leave. But in order to get what I really need, I’m not going to ask ChatGPT the question itself; I’m going to first ask a question about how to ask the question.

This is what I asked ChatGPT: “I’m currently at the PA Convention Center in Philadelphia and have a Delta flight departing at 1:30 PM. I need to work backwards to figure out when to leave the convention center. Write me a prompt to accomplish this.” This one simple step will turn your AI model from an ill-informed advisor to a mission-focused assistant and concise problem-solver.

Yes, the AI fatigue is real and, in my experience, largely driven by our collective frustration with the usefulness of the results. One fix is to stop treating AI like a search bar and start treating it like a junior protégé. And while you’re at it, take a deep breath, walk outside, and remember that you can enjoy the view without asking ChatGPT for the weather report. Sometimes the best things in life are free.

Thad Dickson is CEO of Xpio Health, a Gig Harbor company focused on security and compliance for healthcare organizations. He lives in Lakebay.