Do AI meeting notes make things up?

Short answer is yes.

Since AI note takers became popular, more and more people have reported the same thing: AI-generated notes are incredibly useful, but they often struggle with summarising key action items as well as getting precise numerical figures correct.

This is known as hallucination: the AI produces something that reads like a note but was never said. It isn't a fault in one product. It is how this generation of AI works, and meeting notes are one of the places it shows most, because the input is long, spoken and messy.

What is AI hallucination?

A hallucination is output that is fluent and confident but not grounded in the input or in fact. The well-known cases are invented citations and made-up figures. The everyday case is a summary that says more than its source did.

Why does AI hallucinate?

Large language models, the engine behind modern-day AI, work by predicting the most plausible next words based on a set of training data. They don't check a claim against a source unless there is a harness built around them to do so. So in cases where the input is incomplete, unclear, or spans many different topics, the model still produces a complete-looking answer, and the gap gets filled with whatever the model thinks is most likely. What makes hallucination hard to spot is that the models often sound confident and truthful even when hallucinating.

Why is it worse in meeting notes?

Most AI note takers are a recorder wrapped with an LLM. They transcribe the audio of the call, hand the whole transcript to a model and ask for a finished document. An hour of conversation is a very long, noisy input to digest in one go. The transcript itself is the first weak link: speech-to-text mishears names, figures and negatives, and the model has no way to know a word was misheard. Additionally, in a real conversation there is a lot of small talk and tangents, such as someone reading numbers off a different document or a proposal that was raised and dropped. All of it sits in the transcript with the same weight as the decisions, and it pulls the model's attention to the wrong places.

Compression does the rest. "We might revisit pricing next quarter" is a hedge; "Pricing to be revisited in Q4" is a decision. The summary template has slots for decisions and action items and none for open questions, so open questions become one or the other.

How Sorinai catches it

A better, more advanced model makes hallucination rarer. It doesn't make it impossible, and it doesn't make it any easier to spot. So Sorinai is built around three things that do.

Sorinai is a desktop notepad for Mac and Windows. It listens to the call on your own computer, no bot joins, and it writes into the notepad you're already typing in.

Every line points to what was said

Anything Sorinai adds to your notes after the call has to cite the exact sentence it came from, and with just one click, Sorinai takes you directly to the moment in the transcript where that sentence was said. The server checks every quote against the transcript and drops any that isn't there word for word, so a quote you can see is real.

The same rule holds when it fills a template: a field can't be answered without a direct quote supporting it, and ambiguous evidence is marked for your review. When you ask it to edit your notes, every change has to cite the transcript, unsupported quotes are dropped, and what changed lights up for you to keep or undo.

It builds around what you wrote

If you wrote anything down during the call, Sorinai sharpens and fills in around your points. It doesn't write a summary from scratch about what it thinks mattered. If Sorinai isn't sure whether something was said, it leaves it out; and for names, your spelling beats a garbled transcript.

It writes in the moment, not after

Real-time note taking is the other half. When you ask Sorinai to write something down, the line lands in your notepad while the person who said it is still on the line, so a wrong number gets fixed in the same minute instead of found in a long summary afterwards or in your inbox tomorrow. And while the call is on, it works from the latest stretch of the conversation rather than an hour of transcript at once. This allows us to greatly improve the accuracy of your notes as well as the quality of the model's output.

Hallucinations and poor meeting note quality come from AI note takers attempting to completely remove the human from note taking, whereas we at Sorinai believe that's the most crucial part: your notes staying personal to you.