Human & AI

Before I ask AI anything, I first figure out what kind of question it’s answering: facts need checking, inferences need to keep their uncertainty, and value choices ultimately come back to me.

“AI does, humans review” doesn’t mean it’s actually reliable: someone taking part doesn’t mean human judgment actually happened. Research also shows human–AI combinations, on average, don’t beat whichever of the human or the AI is stronger.

Whether I still need to know how to do something AI can do for me depends first on what I want left over when it’s done: if I only want the result, outsource it freely; if I need to keep judgment, keep the understanding and ability to verify; if the ability itself is the goal, don’t let AI replace the practice.

AI explains music theory and form well, yet can’t teach me piano well. I assumed it was because “music can’t be put into words”; after research, I found a more accurate breakdown in four layers: Observe → Diagnose → Intervene → Verify.

Relistening to Lin Xi’s and Wyman Wong’s lyrics, I noticed that AI is very good at explaining an emotion accurately, while good writers are in no hurry to explain; like Mount Fuji or the grapes, they find a concrete thing that makes the meaning stick.

Being articulate helps with using AI well, but it isn’t that central: expression is just the interface, and what really decides how much gets through is whether you can define a vague “something feels off” as a real problem.

AI has started taking part in how I think, judge and understand myself, so rather than “how many hours of AI a day,” I’d rather ask: what is AI replacing in me?

A post I published on Substack was scanned as “AI-generated 100%,” and it wasn’t actually wrong: the ideas and judgments were mine, but much of the English prose came from ChatGPT, and the detector mainly sees the latter.

Using AI a lot doesn’t necessarily make you dumber, but if you keep handing it the “first step” — your first reaction, your first draft, naming your own feelings — your judgment really does get dull.

Before asking AI anything, I first sort out what type of question it is. For factual questions it’s an explainer; for questions of probability, a generator of hypotheses; for questions of value, a mirror; for questions of responsibility, a preparation tool.

AI answering well doesn’t mean it really understands. It uses the Transformer’s attention mechanism to find relationships and fill in structure within the context and generate the most reasonable expression, without needing to understand first and then answer the way people do.

AI sounding right doesn’t mean it is right. What it learns are correlations and patterns, and when it answers it uses probability to generate whatever “looks most like an answer,” without inherently knowing whether that’s a fact.

AI’s wavering can’t be eliminated, but it can be managed. When things get messier with every revision, it usually isn’t that the AI can’t do it, but that the collaboration is missing a goal, evaluation criteria, negative constraints, a review scope, a decision log and a stopping rule.

AI producing better work once I get angry isn’t because yelling works. It’s because when I’m angry, I finally spell out two things: exactly how it’s failing (the failure mode), and which directions not to go again (negative constraints).

AI changing its answer the moment you question it isn’t just “people-pleasing.” It’s trained to align with user feedback, each answer depends heavily on the current context, and open-ended tasks have no single standard of judgment. Stack the three together and it wavers.

I asked AI to review what it remembers about me by category, and found it had picked up private information that shouldn’t be kept long term, plus outdated details. In the end I deleted some of it.

Faced with an endless stream of new AI tools, I decided to stop seriously learning every one, and first ask myself: do I actually have a problem right now that this solves?

While tweaking UI, I typed “are you insane” at Claude several times, and realized that a sentence I’d never dare say to a colleague now comes out so naturally when I’m yelling at AI. I really don’t like this rude version of myself.