Personal notes
We have to keep watching over generative AI
After "Just have the AI do it," unreadable reports and hollow documents started going around. If work exists for people, what AI produces has to work for people too.
Naming the friction that appears once we hand work to AI and only carry the results, and why people still have to keep observing and supervising.
AI output tends to lean toward what is easiest for AI to understand. Work exists for people, so the job of filtering whether it actually works for people has to stay with people.
Before passing on something an AI made, read it yourself and check that the person receiving it can actually read and use it.
Contents
It has been quite a while since we first started hearing the words “generative AI.” By now you could even call it a slightly old-fashioned term.
In the industry I work in, AI has become a completely ordinary tool.
“Just have the AI do it”
Not many people hesitate over how to use it anymore. And if someone does hesitate or worry, somebody else steps in with advice.
“Just have the AI do it. It’ll handle everything.”
That one line raises plenty of doubts. But hand it a task once, then twice, and the doubts get tossed aside somewhere. Because it makes things so well.
From then on, whether it is a report or an email, we leave it to the AI, and our only job becomes carrying the result from one place to another.
The friction of just carrying results
That is when the friction starts.
- Reports get handed over that nobody can actually read.
- Unrelated stories are stitched together as if one caused the other.
- It sounds plausible, but the substance is missing.
These problems started coming up more and more. Some people say, “AI isn’t perfect yet,” and that we shouldn’t work this way.
I want to put it a little more concretely.
I believe “work” exists for people, and is done by people.
The .md files nobody read
For a while there was an era of enthusiasm for “.md” files. (An era… well, a few months.)
Because they are easy for AI to read, they were treated as the essential format for working with AI. But a question mark followed right behind: people were not reading them.
(Even I read the first few lines of an md file, think “it’s probably fine,” and move on. Reading it asks too much effort.)
So discussion moved on to ways of producing output in a form people can actually read, instead of md files.
Output has to work for people
If what AI generates were read by AI and used by AI, there would probably be no problem at all.
But we work toward a world where people can live better. What gets made that way has to work for people.
Most AI output leans toward whatever is easiest for AI to understand. (Of course it does. It is not human, but it tries to think like one.)
So to keep problems from happening, we have to keep observing, and we have to follow through on supervising. We have to be able to filter whether this is something made for people.
And if you ask why it went so well the first time or two, I think it was thanks to that doubt we later threw away.