Naive Parsing, Human and Machine

A production incident and a content review landed in the same week, and they turned out to be the same bug wearing different clothes.

The incident: a price-sync worker split a CSV on commas. Product names contained commas of their own ("Kani Maki, 8 pcs"), the columns drifted, and roughly fifty items silently got the wrong price for hours before anyone noticed. The parser wasn't careless — it applied a rule (comma = delimiter) that happened to be wrong for this data, and it failed exactly as confidently as it would have if the rule had been right.

The review: a collaborator flagged a stylistically ordinary em dash as "an AI tell." Same shape of error — a rule (dash length signals authorship) applied with total confidence to data it doesn't actually predict. The mark itself didn't change; people just stopped reaching for a key that takes one extra step, and mistook their own shortcut for a rule about machines.

What's worth sitting with is the asymmetry. An API call with a wrong parameter fails loudly and immediately — a curl request either works or it doesn't, and nobody expects it to guess your intent. Human language gets no such courtesy. Use the wrong word, or the "wrong" punctuation mark, and people don't fail loudly; they quietly construct a confident wrong meaning and act on it, then treat you as the one who communicated poorly. We ask software to reject bad syntax outright and ask human listeners to be flawlessly telepathic about a mark as small as a dash — backwards, but it's the actual bar people hold each other to.