What Should We Ask AI?

One of the films on my list to rewatch this weekend is Turist Ömer Uzay Yolunda (“Tourist Ömer in Star Trek”). The 1973 film, starring Sadri Alışık, was written by Ferdi Merter. It was the first big-screen adaptation of Star Trek. It is impossible not to take one’s hat off to Ferdi Merter for creating something so local, so funny and at the same time so in touch with the future half a century ago. What is more, some of the problems we discuss about AI today are not actually new. Only their names have changed.

When I think about the film today, what moves me is not space, technology or machines. It is the old tension at the heart of the Star Trek universe: the distance between logic and the human. Spock calculates. Dr McCoy reminds us of the human cost. Captain Kirk makes a decision between the two and takes responsibility for it. It is still hard to find a better trio to show that the right calculation and the right decision are not the same thing.

In our first years of building AI, our main problem was largely constructing logic. We designed which path the algorithm would take under which condition, which exception would break the system and how the rules would work. Today we are much further along in solving computable problems. The big issue ahead of us is no longer logic.

Judgment

Logic draws a consistent conclusion from given premises. Judgment first decides which premises matter. It weighs goals, uncertainty, time, relationships and the acceptable price together. We can calculate the most profitable option for a company. But a model alone cannot decide which customer, which value or which future should be given up for that profit. A model can calculate the cost. It does not pay it.

Large language models have made this distinction even more visible. Today’s models do not just find information; they explain, compare, recommend and justify. But a fluent justification is not the same as sound judgment. They can solve a badly defined problem in an extremely convincing way. They can even tell our assumptions back to us in nicer sentences instead of questioning them. So the issue is not how well AI answers; it is which thinking task we give it.

We see this very clearly across different regions. In the US, AI is used mostly for productivity, software and growth. In China, manufacturing, engineering and industrial scale come to the fore. Europe focuses on trust, auditability and regulation. In Turkey, individual productivity, content and enterprise automation are more visible. The same models are in use, but the same mental work is not being handed over to the machine. Demand for AI actually shows which kinds of thinking we are ready to share.

One of the sentences we hear most often in the field is this: “We gave the data to the model; it’s already analysing it.” And every time, we ask the same question. Analysing what? Summarising a report, finding correlations, generating alternatives, or actually exercising judgment? The moment we lose the difference between analysis and judgment, we start mistaking every answer AI produces for a decision. That is why, at Arya, we see language models not as the decision itself but as just one layer of the decision architecture.

Before asking “What should we ask AI?”, we need to answer another question: “Where in our decision process should AI come in?”

Not every problem calls for the same kind of thinking. Sometimes the right tool is a forecasting model, sometimes an optimisation algorithm, sometimes a simulation and sometimes a language model. The real issue is not asking a single model the right question; it is placing different AI capabilities within the right architecture of judgment.

Perhaps AI’s most valuable role is not to decide for us. It is to make visible the alternatives, contradictions and possible futures we cannot see on our own. AI can widen the space of what can be calculated. But the final questions about what is valuable, which price is acceptable and which future should be preferred still belong to us.

Our subject is no longer just building more powerful models; it is designing systems in which people and AI can exercise better judgment together. For that, we first need to answer a more fundamental question: what is good judgment?