#40: AI + Thinking, Fast and Slow

AI shouldn’t automate everything (yet)

One of my favorite behavioral psychology books is Thinking, Fast and Slow by Daniel Kahneman. If you haven’t checked it out before, I’ll give you a quick summary. Kahneman and his research partner Amos Tversky break down the way humans think into two distinct systems. System 1 represents fast, intuitive thinking that is mostly driven by our subconscious mind. On the other hand, System 2 handles slow, deliberate decision-making that leads to conscious action.

Let’s explore System 1 and System 2 a bit more before we dive deeper into how AI comes into play.

System 1 Thought Process

I make my bed every morning. It’s a small win, but I’ll take it. Yet, I can’t remember the last time I thought about each step of making the bed (from fixing the sheets to pulling up the comforter and ultimately, repositioning the pillows). It’s automatic for me, as I’m simply executing on a series of ingrained procedures. My subconscious mind is making the decisions for me, which is a tell-tale sign of a System 1 process.

System 2 Thought Process

I recently booked a flight to Europe, and my brain was on overdrive trying to lock down this trip. I compared flight times and corresponding prices across a few different airlines, confirmed I had enough credit card rewards points to cover the flight, and then input my credit card details to cover the random fees that don’t get covered by rewards points (if you work at an airline, please fix this and allow all fees to be covered by points!). I consciously completed this task. That’s System 2 at work.

That brings us to a question that fascinates me: how does AI fit into this model of fast and slow thinking?

At a high level, artificial intelligence (AI) is the process of machines completing tasks typically requiring a human-level skillset. AI is able to mimic processes that are usually reserved for humans, which is impressive because these processes require learning, reasoning, and self-correction.

I think it’s worth asking ourselves: if in today’s day and age we are allowing AI to automate processes on our behalf (expecting AI to be a substitute for humans), how does AI think? Does it think fast and slow, like humans do?

Believe it or not, humans make 95% of their decisions from their subconscious mind. Terms like “muscle memory” and “autopilot” are commonly used to describe our subconscious mind taking the wheel and making the turn. I refer to these kinds of actions as micro decisions. They’re choices we make without actively thinking about them. We just do them, no second thought required. Actions like stopping when you see a red light when driving, walking to the end of a line when a big crowd is up ahead, or saying, “hi, how are you?” when you see someone familiar are all actions that tend to happen in an automatic fashion. They’re quick, habitual, and require little conscious effort; hallmarks of System 1 thinking.

AI is great at these types of tasks (System 1). As I mentioned in AI + Brand (Part I), I used AI to help choose an outfit for a work dinner. Picking an outfit out of my closet is something I do every day, and is rarely something I consciously think about doing. But why not see if ChatGPT can do the task too?

Because my prompt was both specific and relatively simple (“suggest an outfit for a work dinner with this vibe”), ChatGPT gave me exactly what I was looking for on the first try. The simpler the prompt, the better the result. Fast, intuitive output for a fast, intuitive task.

One day, I expect a robot to be able to make my bed for me. What feels habitual to me can become habitual for a machine too. AI can knock out those micro decisions so that we have more time to focus on the bigger, more strategic decisions in our lives.

And that’s because AI isn’t fully there yet on System 2 thinking. Since AI is trained off vast amounts of human-generated data (as well as data generated by technology), it’s excellent at pattern recognition. But multi-step, ambiguous questions pose more of an issue for AI today.

Revisiting the System 2 example, there are plenty of AI tools that allow you to book a flight based on the parameters you give it. However, my European flight search had a lot of nuances as I didn’t initially know which days I wanted to travel. In my head, I factored in details like constantly changing itinerary plans for a multi-city trip, weather in various cities, fluctuations in dollars versus rewards points, and much more. For what it’s worth, I did ask ChatGPT for help and I wasn’t impressed. With so many details, it was difficult to provide ChatGPT with inputs that it could leverage its elite pattern recognition abilities to go out and find the ideal flight. Personally, I find AI most helpful when I know what the end solution will look like. It’s not super helpful for solving ambiguous problems.

So far, I’ve found AI thinks fast more than it does slow. It nails the subconscious System 1 tasks that we humans do without even thinking about. Yet when it comes to those thorny, multi-pronged questions that we strategize on (System 2 thinking), AI struggles.

But this is great for now. AI can handle our monotonous tasks, while we use our brain power to focus on the bigger-picture problems that we know AI lags on solving today. We’ll be able to become more efficient and effective, with time freed up to focus on these more strategic (and usually interesting) problems.

I expect researchers to continue to have breakthroughs and improve the System 2 thinking capabilities of AI. Even in a year from now, there’s a scenario where I’m looking back on this article and thinking, “dang, AI moved faster than I thought”.

However, for the time being, let AI think fast, while you think slowly.