It helps when you can see the confidence of each token, which downloadable weights usually gives you. Then whenever you (your software) detects a low confidence token, run over that section multiple times to generate alternatives, and either go with the highest confidence one, or manually review the suggestions. Easier than having to manually transcribe those parts at least.
Do they actually have access to that info "in-band"? I would guess not. OTOH it should be straightforward for the LLM program to report this -- someone else commented that you can do this when running your own LLM locally, but I guess commercial providers have incentives not to make this info available.
Naturally, their "confidence" is represented as activations in layers close to output, so they might be able to use it. Research ([0], [1], [2], [3]) shows that results of prompting LLMs to express their confidence correlate with their accuracy. The models tend to be overconfident, but in my anecdotal experience the latest models are passably good at judging their own confidence.
