Notes from the workbench

I have always maintained that AI is always inferior to human intelligence despite how it may impress us. AI is made in the image of man, and man is given the agency and responsibility over the things that we are given to take care of. This week, my principle has been further affirmed once again, in which I write this to remind myself and to share with you the critical importance of not handing agency over to the machine, and how I design and build my products.

It happened while I was working on a draft of the Orbital Softworks site in Qanvas (opens in new tab). The AI had written a line about one of my products that did not describe it accurately, and another line that implied a commitment I had never made. None of this went live, because I had already written down what the site should say, and when I personally read against that, the problematic lines stood out.

What is interesting for me is that the system did exactly what I built it to do. Qanvas checks every change the AI proposes against a schema before it is staged. Since those lines were well formed, they passed as they should. The schema can only tell me whether something fits. It cannot tell me whether it is right. That is a judgment, and that judgment belongs to me because I am the one who has to stand behind what is said.

This is how I deliberately design my products on purpose. In Qanvas, the AI has no write access to component code. What it proposes is staged as a draft for me to see, and it remains a draft until I save it. With BuildQrew, which I use to coordinate my coding agents, an agent can propose a decision but only I can accept it. BuildQrew does not run the agents either. I talk to each of them myself and decide what happens next. There will be times when I disagree with what the AI suggests and override it nonetheless. The system has to allow for that. The system should ultimately adapt to me rather than I adapt to it. This is the correct principle for products powered with AI.

I don't think a better system moves the responsibility away from me. If anything, the better it works, the easier it is for me to stop reading closely. I therefore want the point of decision to be built into the process instead of depending on how attentive I happen to be that day.

If you are building any solution or system with AI, do ask yourself whether a person still decides in your own process, and whether that step would survive a busy week. As for me, the AI proposes, the system remembers, and it is the human who decides.