We wanted to go on a bike tour recently. Nothing complicated. A nice route, suitable for the family, ideally somewhere scenic. I asked AI and it immediately came up with some great proposals. All routes sounded nice and were within the distance I had asked for.
The first suggestion described a scenic road around a small lake. Sounded like a perfect match to what I was looking for. I was already about to settle for the first suggestion, but I just wanted to check quickly whether this route was also fine for kids. Just to be on the safe side.
The AI told me that the route was maybe not ideal for kids, so it updated the suggestions. Good that I asked, I thought.
One question leads to another
What I did not realize at this point was that this started a long research process where, after an hour of back and forth, I still did not have a good bike tour.
Knowing the area, I thought that I’d better ask about the path that was suggested. Would there be any steep hills? And can we actually cycle the whole route?
To my surprise, some of the suggested routes were not really suitable for bicycling. As always, the AI was very friendly and talked to me like I was asking a colleague at work to help me with a problem. ‘I had another look at the proposal. I just realized that the suggested route contains stairs. Let me reconsider the proposal.’
And off we went with a completely different set of proposals. Again, I asked a couple of follow-up questions. Would we be cycling on a busy road? The AI responded friendly and reconsidered its proposal.
I was getting more and more frustrated. It seemed that I was far away from finding a good bike tour. But at the same time I was intrigued by how the AI was working and how it affected my decision-making. I eventually realized that the AI wasn’t really doing what I assumed it was doing.
What I was expecting and what the AI was doing
The answers always sounded great. It is just very good at writing convincing answers. So convincing that I sometimes find myself going along with the first proposal without really questioning it.
Looking at the bicycle example, there are definitely points where I did not tell the machine all the relevant things. I was implicitly assuming that it would be doing the right thing. I never told it that I do not want to cycle besides a busy road. How could it consider this point if I never told it that it is important to me? For some people it seems to be perfectly fine to do so.
At some stage, I started wondering whether I was simply using an outdated AI model. Maybe a newer model would have handled the task better. But I wasn’t sure that this was really the point.
As frustrating as this was, the interesting part for me was to see how the AI was working. Somehow, I had assumed that it would consider all the available information on the Internet. Everything that is needed for a perfect bike tour. And all the different possible routes in the surrounding area.
But this is not how it works. The AI had found one plausible source and built a very convincing recommendation around it. It adjusted the recommendation whenever I introduced information that didn’t fit.
This became apparent when it recommended a route where we had to cycle on a busy road where there was not even a bike lane. By this point, I was skeptical, so I looked at the map to validate the suggestion. This is when I saw the road and asked directly about it. Just before, I had told the machine that I wanted to avoid big roads. I was puzzled.
Then the AI told me that this suggestion was based on a posting in a forum that was five years old. Back then, the road had been closed to cars for a year due to a landslide. But today, the road is open again. ‘Sorry, let me reconsider…’
The engine had taken this one source from five years ago without validating it. And then it wrapped it nicely in a well-written answer. The problem was that I mistook a fluent answer for a well-founded decision.
AI and decision-making
This made me think about how we use AI for decisions. The dangerous part isn’t necessarily that AI gives us a wrong answer. It’s that it can give us a very convincing answer before we’ve asked the questions that actually matter.
In quite a few posts, I have argued that AI cannot replace human judgment. To be honest, the answers from the AI are so convincing that I am often intrigued to go along with the first proposal without questioning it. Essentially, I was letting AI take the decision for me.
The experience with the bike tour showed me that this can go terribly wrong. So, I take this as a good reminder that AI can help you make a decision without actually being able to make the decision for you.
AI is very good at answering questions. But decisions are not just collections of answers. A decision requires you to determine what matters, how much it matters, and what you’re willing to trade off. Do I prefer the beautiful route, with some sections on the road, or do I choose the route that is completely car-free but less scenic?
A decision often contains something that isn’t in the data. It contains preferences, values, risk tolerance, responsibility and sometimes simply the willingness to choose an option that is good enough.
The decision is still mine
The irony is that AI didn’t make the decision easier. It made me realize that I hadn’t actually defined the decision properly. Which, in retrospect, was probably the most useful answer I got.
After many iterations, AI eventually proposed a route recommended by the local tourist agency. I settled on this route because I felt confident that it was suitable for cycling.
In the end, we enjoyed a nice day with the occasional detour. Nothing to do with technology. This time, the problem was entirely human.
