Can you trust AI?
Trusting AI is not a yes or no. It applies to a particular model and a particular kind of task, and you cannot read your way to it. It is a craft you build by testing your limits where it is safe to fail.
After a talk at a conference, someone in the audience asked the question many people seem to be carrying around right now: can you trust AI?
The answer was roughly: yes, I would say so, but you have to try it yourself. Get to know it, experiment and become confident.
I think he was right, and that the answer says something important about what it means to work with AI. You cannot read your way to it. It is a craft.
But “experiment” is easy to say. What does it mean in practice? Here is what I have learned myself.
You get to know a model
Trusting AI is not a yes or no. It applies to a particular model and a particular kind of task.
That became clear to me in a conversation with a colleague. If I am used to one model and happen to start a session with another, I become unsure. Suddenly I want to review the work in the session.
The work has not necessarily got worse, but I have lost my sense of how to read it. It is like working with a colleague: after a while you know what you can rely on and what you need to double-check. You only get that knowledge by working together.
AI always finds something
More reviews do not automatically give better results. Tools increasingly let one AI review another AI’s work, and we have noticed that the reviewer nearly always finds something.
So a list of remarks does not say much about whether the work is good enough. Ask an AI to look for faults, and it will look for faults. Someone still has to decide what actually matters.
That is where the human comes in. You need to know what to take into account, what you can ignore and when it is good enough, and then steer the work accordingly. That judgement is the core of the craft.
It becomes especially clear when you work in an area where you are not an expert yourself, such as law or finance. When I went through my own tax return with AI, it concluded that I should submit a correction. So I did. When the reply came from the Swedish Tax Agency, I realised it probably had not been needed. It just created unnecessary work, both for me and for them.
The problem was not that the AI found something. The problem was that I could not judge whether it was worth acting on, because I do not have enough expertise in the area. That judgement still has to come from a person, and sometimes from someone who knows the subject better than you do.
How to become confident
Confidence comes from testing your limits, and you can do that without risking anything for real. Two things have helped me.
Ask the AI first
At the start I was quite bad at testing things myself. I got stuck and thought I had to ask someone who knew this stuff.
Now I see that thought as a signal: put the question to the AI and reason with it. It can sound confident even when it is wrong, so it does not replace experts. But you get further on your own, and if you still need to ask someone, you will have a better question.
Solve real problems where it is fine to fail
You need to set aside time, but I find it hard to just sit and tinker with something for the sake of tinkering. What works for me is to pretend I am solving a real problem, but on test data. Then I have learned something I can later use for real.
If you are worried about touching real data, you can ask the AI to build a replica of your environment to practise in. Use test data instead of customer data. And remember that a replica is only as good as what it knows about. Something working there does not mean it is ready for production.
So, can you trust AI?
Yes, once you have learned what you can trust it with. That is why nobody else can give you the whole answer.
Set aside time, pick a real problem and start where it is safe to fail. Test your limits, get to know the model and train your sense of when it is good enough. That is how you build the craft.
We are Hanna and Tobias, and we write here. We try out everything we write about ourselves, and help teams that want to do the same. Have a question after reading? Send it to us.
Yes, we wrote this together with Claude. It would be a bit odd otherwise, but we know what we want to say, and we have read, poked and discarded until the phrasing, opinions and choices are ours.