24/02/2026
The more I use AI, the less I trust it with complex things
By Robert A. Le Veneur
AI is certainly great for simple things and can save a lot of time. But with complex things, even if I enter a prompt the length of a short article, it still can't grasp the complexity of the world. It simplifies it terribly.
For example, I'm currently looking for a new supplier. I know about 30 possible suppliers in this area, I have written to at least 20 of them, and I have personally negotiated with about five. I wondered if Gemini 3.1 Pro Deep Research would surprise me.
The results were appalling. If I didn't have in-depth knowledge of the field, the result would have looked "impressive." But because I do have that knowledge, I often see "fluff" in the AI output.
Despite a very detailed prompt, the AI was unable to perceive the complexity of the situation and subtle nuances. When I told it that it was wrong and entered Deep Research a second time, it adjusted the result a little, but it was still at a level of, say, 5 out of 10.
The reason I'm lowering the score so much is that it didn't "think" about the results, even though I told it how to think. It's like asking a junior secretary to make you a list of suppliers.
She will make it for you. But only a senior person will be able to cross it off immediately. For example, the AI described that one supplier has a certain advantage - which in reality is an immediate elimination rule.
Only when I divided the big task into small subtasks did it start to become really useful. But it was very one-sided.
Don't be fooled by various official tests of different AI models. Their weakness is that they have a clear-cut solution to evaluate. But reality sometimes has ambiguous solutions. Or hidden solutions. Or no solution at all. Or it requires solutions for which we have no data. And as soon as AI has no data, it fails.
We all know from high school that parallel lines will intersect at some point. But we can't point to where. AI tends to show you that point.
AI developers are, of course, well aware of this problem. That's why they don't give you the ideal or optimal result, but the closest possible approximation, which is also limited by computing power.
I have also noticed that it very often plays it safe. Instead of recommending a risky maneuver, it discourages you from certain moves. But you can't do business without taking risks.
AI is also unreliable because it often uses Google searches in its work and only looks at the main results. But you have to look for the best suppliers on, say, the fourth page of Google, where the robot will not go in the vast majority of cases.
At our strategic consulting firm, Le Veneur Sàrl, we repeatedly test whether AI is suitable for drafting contracts. It saves time, but it cannot replace a senior or even a mid-level lawyer.
In one of our tests, we asked it to create general terms and conditions using a long prompt. The first result was quite general, despite the amount of data available to it. The problem is that by default, it generates a contract that is perhaps two pages long. But what if you need something more detailed?
Only after many iterations and several hours of telling it, "You forgot this and you forgot that," was the result excellent.
So yes, in the end, it was able to create excellent general terms and conditions and save 60 to 70% of the time. But only because it was managed by a senior person who still had to slightly adjust the result.
That's why I'm not at all concerned that it could replace people who work with high complexity.
Is it suitable for obtaining certain information? Absolutely yes. Is it worth listening to? Absolutely yes. Should you trust it and blindly follow it? Absolutely not.
I recently tried to vibe code a solution for emails. Programming advanced solutions for emails is very demanding.
After many hours of futile attempts to debug the code with ChatGPT, Gemini, and Claude, I finally bought a ready-made solution for $15.
And one last example to finish. Gemini Pro 3.1 recommended that we drive to Castillo de Almodóvar del Río with our young children rather than walking, as the climb would be too difficult for them.
The AI had information about the children's ages, their physical abilities, and what they enjoy and don't enjoy.
The children decided to ignore the AI's recommendation and raced up the steep paths, pretending to conquer the castle.