How do you actually verify AI-written content is accurate before publishing?

Been thinking about this a lot lately. Everyone talks about "editing" AI output, but editing usually means fixing tone and flow not checking whether the text still says what the source material actually said.

The failure mode that gets me: a rewrite pass can turn a hedge into a certainty ("may improve" becomes "improves"), or quietly shift a number, without any of it looking like an error. It reads perfectly clean.

Spellcheck and grammar tools won't catch it because grammatically it's fine it's just no longer accurate to the source.

For those of you publishing AI-assisted content regularly:

• Do you manually re-check every rewritten claim against your source, or is there a shortcut you've found?

• Has anyone actually caught a real case of this drift before it went out?

• Is this a real problem for your workflow, or am I overthinking it?

Not looking for "just read it carefully" curious if anyone's found something that actually works at scale, versus per-article manual re-checking.