At work we've tried AI summaries for meetings, but we spent so much time fixing those summaries that we started writing our own again.

Is there some training you applied or something specific to your use case that makes it work for you?

We stopped after it kept transcribing a particular phrase of domain jargon as “child p*rn”, again and again.

Unless a case goes down the legal road, nobody is ever bothering to read old call summaries in a giant call center.

When was the last time you called a large company and the person answering was already across all the past history without you giving them a case number first?

My guess is that the summaries are never actually read, so accuracy doesn't actually matter and the AI could equally be replaced with /dev/null

We tried Otter.ai, someone complained and asked: "Could you f-ing not? I don't trust them" and now Otter is accused of training their models on recorded meetings without permission. Yeah, I don't even care if it works, I don't trust any of these companies.