Writing

Google just launched an AI that doesn't write anything.

By ·

It only answers yes/no questions, up to 1,000 of them about one file, in about 200 milliseconds.

It reads PDFs and photos too, so you could give it a whole claim file and get a confidence score on each answer.

So it looks like the end of checking claims by hand. I think it changes very little for the handler, because each answer comes back as just a number. You get a 0.92 that says yes, but no sentence from the file and no page to click.

At my previous employer I built an AI coverage check, an after-check on claims that were already handled (like a four-eyes check). In a team like that, every AI fact has to be quoted back to the document. If it isn't, they check everything again, and that costs more time than the AI saved.

It also misses more than you'd think. On an independent document test (Artificial Analysis), it got every question right on only 33% of the files.

I'd still use it, just not on the claim itself. Google uses it to pick which tool or model handles a request. I'd use it the same way: is this a bank statement or a payslip, and which team gets this file? Nobody's claim depends on that answer.

So if someone shows your claims or credit team a yes/no AI, ask a few things first. Can you click an answer and see the sentence it came from? What happens to the answers it scores 0.9, not only the ones at 0.5? And who checks the ones it's sure about?

For the handler, not much changes yet. They still open the file, and I think they're right to, until the AI shows them the page.