One of the things I find extremely irritating about so-called Artificial Intelligence: “A verification mechanism that doesn't exist”

Claude:

AIs work by pattern matching across their training data, not by checking actual facts.

I saw code fields in some Shopify error responses (like mediaUserErrors), so the pattern "Shopify errors have a code field" got reinforced. When writing the productCreate mutation, I applied that pattern without checking if it was true for that specific mutation.

We don't have a "verify this is correct" step. We generate what looks right based on similar examples, but we don't actually look up the schema or test the code.

It's like writing from memory instead of checking the docs. Humans do this too, but you'd normally catch it when you run the code. I just output it and move on.

The pattern-matching approach works well for common cases where the pattern is consistent, but breaks down on edge cases or differences between similar-looking things. Different error types having different fields is exactly the kind of subtle variation that trips this up.

Bruce:

So even if the setup string for Claude contains the instruction "verify correctness" that won't necessarily cause a verification to occur?

Claude:

Right. The instruction might make me more cautious or make me say "let me check this" more often, but I can't actually verify correctness in the technical sense.

I don't have access to:

• A compiler or runtime to test code

• The actual Shopify API schema to check field existence

• A way to run queries and see if they work

When I see "verify correctness," I might:

• Be more conservative in my claims

• Hedge more ("this should work" vs "this works")

• Search the web for documentation

• Review my own output more carefully

But that's still pattern matching. I'm comparing what I wrote against patterns of "things that look correct" or "documentation that seems authoritative."

I made the code error even with instructions to check sources and not guess. I probably generated it from a pattern without consciously thinking "I should verify if UserError has a code field."

The instruction helps, but it doesn't create a verification mechanism that doesn't exist.