4 notes
Notes
Bounded observations and working models. Notes are useful before they deserve a Case; they have a smaller epistemic claim and an explicit status.
13 September 2026 · Research note · AI & learning · Language & judgment
When an AI answer feels wrong before you know why
Sometimes an AI answer misses something before I can say what is missing. That reaction can matter, but it does not prove that I know the right answer or that my first explanation is true. One way to use that reaction is to write down what bothered me before asking for more explanation. Then I can compare possible reasons and find a case where they would lead to different decisions. That gives me something to test without assuming I was right.
12 September 2026 · Research note · AI & learning · Language & judgment
The Experience AI Shouldn't Replace
AI can make work easier while making learning shallower. The decisive question is what the human still has to do during the interaction. Research on expertise and AI tutoring suggests that technology can accelerate learning when it concentrates useful cases, preserves prediction and struggle, and provides diagnostic feedback. The same technology can improve assisted performance while weakening what a person can do alone. If AI becomes part of the environment in which we think, language and judgment become ways of staying inside that learning loop: seeing the frame, testing the answer, and deciding what remains ours to own.
11 September 2026 · Method note · Language & judgment · Work & organisations
The draft was fixed. The preview still said something else.
A correction reached the article but missed its summary. A case about preserving what a writer means across the whole piece a reader encounters, with a check you can try on your own work.
10 September 2026 · Method note · AI & learning · Work & organisations
A good AI conversation is not an operating system
A useful AI conversation can still leave the next session starting over. This article follows an attempt to make a writing job resumable through files, explicit decisions and checks—and what failed when those checks existed only on paper.