Yes what's with that in LLM output?

It used to be a big thing in the nineties: I've got old .asm source code of mine where I used to do that.

But somehow LLMs love to insert dashes everywhere: dashes in source code an em-dashes in prose. Just why?

Did they parse lots of early code and thought it was cool to insert, in modern programming languages, comment lines full of dashes?

> Another fun one: comments that are justified to a specific column but off by one in only one of them.

Oh yes, all the time. And besides the fact that there are the off-by-ones errors, it of course looks horrible in Claude Code CLI seen that what you see is not what the LLM did output (because they vibe-coded their "real time game engine" that changes characters, for no reason, on the fly).

It's 2026 and we've got "intelligent" machines doing this:

    //  -------------------------
    // ------------------------
    // ----- Input Handling ----
    //  ------------------------
    //  |--------------+-------+------|
    //  | Potentiometer |  Min | Max | 
    //  |--------------+-------+------|
   

    ascii_table_to_unicode_mismatch_alignment_fixer(...);

My guess is that the humans they had in the loop for RLHF just simply preferred the code because it looked superficially tidier. I have the strangest feeling that they didn't always have top notch engineers in the loop at all steps.

I suspect this is also how LLM prose gets so utterly bad.

Of course, for indentation and ASCII graphics, it would be less prone to breaking constantly in this way if it were not a next-token predictor.

Yes, they’re trained in lots of old code.