The Gate That Approved the Broken Draft
A new meaning-check flagged two clean AI drafts and missed two broken ones. Here’s what that reveals about trusting an automated writing gate.
OpenRouter
The Script That Couldn’t See Free Models
My own tool hid every free model on OpenRouter, including one that turned out to be GLM-5.3-Flash. Free changed nothing about how I judge a draft.
The Second Model Marked the Fiction as Sourced
I tested whether a second AI model can fact-check the first one’s draft against the source. It caught 4 of 10 fabrications, and once it noted the mismatch and still marked the sentence sourced.
The Gate Nobody Walked Through
I built a safety check for AI delegation, rolled it out to a second site, and gave it two real jobs. Neither one used it. Here is what that tells me.
I Was Scoring the Wrong Thing
I gave four cheap models a smaller job: reorganize my research notes. Invented figures went to zero. Two of the four deleted the caveats instead.
Cheap Models Are Cheap. Reviewing Their Output Is Not.
I scored eleven AI-written drafts of the same article, from $0.003 to $0.10 each. All eleven failed the same item, and paying ten times more did not raise the score.