People ask, fairly often, whether the AI can go online and verify things itself. I have a stock answer now, and it starts with the fact that we deliberately did not build it.
It wasn't laziness. We gave up something that demos extremely well.
Checking the world is not checking your source
Suppose the model takes your quote, searches the web, and comes back with: this is accurate, the source is reliable.
It has just answered a question about the world.
The question you need answered while drafting is narrower. Does the passage I quoted mean what I am claiming it means. Only you and the passage can settle that. No amount of external search tells you whether that specific paragraph, in that specific wording, supports the conclusion you have attached to it.
Then it starts filling gaps
Ask a language model to verify a fact and you have handed it a task in a space where it has no information. It has one move available: invent the missing part.
The invention is convincing. Author, year, journal, page range, DOI. You go looking and it isn't there.
That is not a weak model. That is a badly specified job. You asked a system optimised to produce an answer for an answer, in a situation where the correct output is "I can't tell from what I have".
So we made the scope smaller
What the checker does now: compares what you wrote against the passage you highlighted and saved.
It will say: your paragraph says "significantly improved"; the source says "improved somewhat". Your wording is stronger than your material.
It will not say "this conclusion is correct".
The first you can verify. Click through, read the paragraph, decide for yourself. The second you can only believe or disbelieve. A feature that is either infallible or worthless breaks permanently the first time it is wrong.
What that costs
It sounds less impressive. "We verify your facts" is a better sentence than "we point you back at the paragraph you quoted".
We ran the trade anyway. In writing, one fabricated citation does more damage than a hundred checks that found nothing.
Related: AI drafts: how to find the citations that aren't real · Forty-seven citations, the night before deadline