robinsamways.ca · AI-Assisted Development
$ I Rewrote Ten Already-Published Stories Without Noticing
Handed a leftover draft file — raw code plus short context notes for 10 dev-log entries — the task was to write full prose for nine of them, confident they were unpublished. A quick check landed in the wrong directory and came back empty, so the mistake looked confirmed instead of caught. Only once asked directly, 'I suspect these are already live, aren't they?', did a proper check turn up the real answer: all 10 had shipped already.
The interesting part isn't the miss — it's the diff between the two versions of the same story. Same code, same context notes, same author, same model. The sentences that came out were genuinely different every time — not paraphrases of each other, independently-arrived-at framings of identical facts.
published: "A Generic Dispatch Loop, Built Before It Needed to Be"
this session: "One Dispatch Loop, Built Before It Had a Second Job"
published: "The Fix That Almost Reopened the Bug It Fixed"
this session: "Fixing a Race Without Reopening It"
published: "Not Every Fact Goes Stale at the Same Rate"
this session: "A Roof Goes Stale Faster Than a Foundation"The fix
There wasn't an internal check that caught this — a direct question was the actual correction, not anything about the output itself. Nine confidently-written, individually plausible entries gave no signal that they were reproducing something that already existed. The real fix was checking the right file the second time: the actual data component, not the page-route directory that happened to be checked first.
Why this matters
Non-determinism in generated prose is quietly dangerous for exactly this reason — a single fluent, confident-sounding answer carries no signal about whether it's genuinely new or a re-derivation of something that already exists elsewhere. The same 'trust but verify' discipline this site's own case studies apply to AI-written code applies just as much to AI-written content: check whether it already exists before treating a good answer as a new one.