Human-AI Interface#
User interface is about control. The right amount of control. It is granted to users to allow software to serve its purpose properly. A system of considerable complexity must provide an abundant number of ways to interact with it. This is how sophisticated capabilities are made available.
Artificial intelligence was introduced to the world at large through quite a modest user interface: a prompt area to “ask anything you want”, the response area for intelligent responses, and simple ways to organize conversations. This contradicts the expectation based on experience of using modern software.
It Is Unique#
As soon as we realize that artificial intelligence is a unique phenomenon that functions on principles totally different from what constitutes conventional software, we shatter not only the idea of user interface but also the idea of control itself.
The reason the user interface for AI is so simple is not that the underlying system is simple. The complexity exists and lives elsewhere. We can use natural language. Not highly formal expressions, such as in programming languages, that you have to learn specifically. Everyday language with all its irregularity, ambiguity, and vagueness will suffice. AI will understand you and respond by answering hard questions, analyzing complex ideas, even constructing software along with a sophisticated user interface.
It Understands Us#
The scenario of constructing software or any other system is a special case. Although we don’t have to be as formal as in programming languages, we must use clear and unambiguous language. What if we don’t? The system won’t be built? Surprisingly, it is quite the opposite: AI always delivers. It just won’t be the expected result.
OK, no problem: clear and unambiguous. Got it.
In fact, the clear and unambiguous is a serious challenge. Because it is not what we consider clear and unambiguous. It is what is translatable into a set of instructions that can be interpreted in one way only. Since AI always delivers, our control is getting looser at every point where there is another way to see the tiniest bit of our request. We can’t figure out each such point ourselves.
We Must Be Specific#
The problem of making instructions effective is well known and widely discussed. There are several ways to address it: properly constructed prompts, curated examples, specification driven development, skills, context management, and so on. And they work. What is difficult about any solution is to apply these methods to a growing system with all the uncertainty that comes with changes.
We must discipline ourselves to use the language as a tool and apply a well defined protocol for controlling the AI by restricting it to a specific role, by analyzing and giving feedback, and by seeking ambiguity and vagueness across instructions. The discipline works as long as the language is precise, AI stays within the boundaries of the assigned role, the feedback is complete and trustworthy, and all instances of there is more than one way to see it are reported.
Here, we must remember something that any AI client interface provides: AI can make mistakes. And it will if there is no method that warrants feedback trustworthiness and completeness of the instruction problems report. Mistakes in a thoroughly crafted set of instructions are much harder to find than in a sloppy request.
The harsh truth is that we cannot control AI doing its job.
We Control It#
Our interaction with artificial intelligence is of a unique nature. Unprecedented. The forms of making effective instructions have limitations because they have grown from human-to-human interactions. Although artificial intelligence is great at reasoning, it is not identical to a human: it connects the dots differently.
For this reason, .. include:: ./partials/featured.perspective.528.txt
+ References#
+ Page Info#
Version |
1.0 |
Revision |
2026-10-16 |
Written.by |
Borys Belinsky (borys.belinsky@protonmail.com) |
Assisted.by |
Anthropic Claude, Opus 5.5 |