New AI Predicts Where Your Meeting Is Heading Before It Gets There
Could warn you a meeting is about to go sideways — while you can still fix it.
A team of researchers has built an AI that can read the first part of a meeting transcript, hear a question like "are we close to a decision?", and then write out the next several back-and-forths on its own. To train and test it, they assembled 2,207 real meetings along with 24,794 questions about where each discussion was going. It's the difference between a note-taker that summarizes what already happened and an assistant that senses what's about to happen.
The clever part is how they trained it, in a method called GLARE. Think of two AIs playing cat and mouse: one writes fake continuations of a meeting, the other tries to spot which version is the real human one. Every time the writer fools the judge, it gets a better score. It's like a forger and an art detective improving each other's skills. Human raters picked its output 66% of the time for usefulness and 70% for sounding like real people talking.
Here's the honest catch. The AI never beat actual human conversation — real people still scored higher. It also isn't trying to predict the exact words your colleagues will use; it produces a believable version of where things could go. And it only works if meetings are recorded and transcribed, which many workplaces and employees aren't comfortable with. So far this is a research paper, not something you can download and switch on.
Still, the direction matters for anyone who sits in meetings all day. Tools like Zoom, Teams and Otter already summarize what was said; the next step is nudging you mid-conversation — "this topic has circled for ten minutes," or "two people here disagree." The researchers also pitched their test as a way to compare big-name AI models on social smarts, the kind of thing assistants will need before they can sit in a room with you.
- The AI learned from 2,207 real meetings and 24,794 questions about where each discussion was headed, so it predicts the next few minutes of conversation rather than just summarizing the past.
- It was trained by pitting one AI against another — a writer and a fake-detector — and human judges preferred its results 66% to 70% of the time over older methods.
- Real humans still score higher than the AI, and the system only works on recorded, transcribed meetings — so privacy and accuracy remain open questions.
Why It Matters
Could help you spot a stalled meeting or a brewing conflict early — and cut pointless meetings short.