AI note taker hallucinations in interview notes: what the research shows and how to check

AI note takers make two kinds of mistakes in interview notes. The speech-to-text layer mishears numbers, names, negatives and speakers, and occasionally inserts whole phrases nobody said: a study of OpenAI's Whisper presented at ACM FAccT 2024 found roughly 1% of the transcriptions it tested contained such fabricated phrases. The summary layer, a large language model, can then state things the transcript does not support or smooth over gaps, which vendors including Otter and Microsoft acknowledge in their own documentation. Neither is rare enough to skip: check every quote, number and negative before it reaches a scorecard or a client.

Below: what counts as a hallucination in interview notes, what the primary research and the vendors say (checked October 2026), where errors cluster, a ten-minute check, and how to test a tool before you buy it.

What counts as a hallucination in interview notes?

"Hallucination" gets used for five different failures. They need different checks, so it helps to separate them.

Strictly, only the second and fourth are hallucinations: text or claims with no source. But all five put words in a candidate's mouth, and all five end up in hiring records if nobody checks.

What does the research say?

• Fabricated phrases in speech recognition. Koenecke and colleagues tested OpenAI's Whisper, which they describe as outperforming industry competitors as of 2023, and found "roughly 1% of audio transcriptions contained entire hallucinated phrases or sentences", that 38% of those hallucinations included explicit harms, and that they occurred more often for speakers with longer non-vocal stretches, such as people with aphasia ("Careless Whisper: Speech-to-Text Hallucination Harms", ACM FAccT 2024). Interviews are full of thinking pauses, which is why a phrase that appears in a gap deserves suspicion.
• Accuracy differs by speaker. A study of five commercial speech recognition systems from Amazon, Apple, Google, IBM and Microsoft found an average word error rate of 0.35 for Black speakers against 0.19 for white speakers (Koenecke et al., "Racial disparities in automated speech recognition", PNAS, 2020). The systems tested were 2019 versions, but the lesson stands: if some candidates' transcripts are worse, notes built from them are worse too, and that is a fairness problem, not only a quality one.
• The model makers say so. OpenAI's Whisper model card states that predictions "may include texts that are not actually spoken in the audio input", reports lower accuracy on languages with less training data, and notes disparate performance across accents and dialects.
• Summaries drift from their source. In a study of abstractive summarization, human annotators found "substantial amounts of hallucinated content in all model generated summaries" (Maynez et al., "On Faithfulness and Factuality in Abstractive Summarization", ACL 2020). That work used news articles and 2020-era models; today's models are better, but the failure mode, a fluent sentence the source does not support, is the same one to check for.

We are not aware of an independent, published study that measures hallucination rates for interview note takers specifically. Vendor accuracy claims are not that study, so test your own tool (method below).

What do the vendors themselves say?

• Otter, in Understanding hallucinations across Otter AI Chat: "no vendor can guarantee no hallucinations". Its accuracy FAQ adds that accuracy varies with background noise, accents and conversation complexity, and recommends reviewing transcripts of important conversations.
• Microsoft, in Recap in Microsoft Teams: "Some AI-generated content might be inaccurate, incomplete, or inappropriate."
• Google, in its Take notes for me help page: "Sometimes, the meeting summary can be incomplete, inaccurate, or not generated."

All three were checked in October 2026. A vendor that claims its notes never contain errors is making a claim the rest of the industry does not.

Where do errors show up in interview notes?

The interview transcript example shows several of these in one raw machine transcript next to the corrected version.

How do you check AI interview notes in ten minutes?

1. Decide what you will use. List the quotes and facts going into the scorecard, submittal or ATS. Check only those; you do not need to proofread the whole transcript.
2. Find each quote in the transcript and confirm the speaker label. If you cannot find it, delete it.
3. Listen to the numbers. Jump to each pay, notice, date and size figure in the recording, or confirm it with the candidate in writing.
4. Read the negatives aloud in context. If the sentence makes sense either way, check the audio.
5. Trace each summary claim to an answer. "Led", "owned", "managed" and "built" need the candidate's own words behind them.
6. Restore gaps. Where the summary says "some experience", write what was actually said or "no evidence".
7. Strike anything from a pause or crosstalk that you do not remember hearing.

If you use AI to draft the summary as well, the prompts and checks in how to use AI to summarize interviews make this step shorter, because they ask the model to quote and to say "no evidence" instead of inferring.

How do you test a note taker's accuracy before buying?

A one-hour test with a colleague tells you more than any accuracy percentage on a pricing page.

1. Write a mock interview script with planted facts: five numbers (including a "fifteen" and a "fifty"), three negative answers, four uncommon names, one ten-second pause, and one stretch of deliberate crosstalk.
2. Run it on the setup you will really use: the same platform, headset or laptop microphone, and a phone screen if you do them.
3. Count errors in the transcript against the script: wrong numbers, flipped negatives, wrong names, wrong speakers, inserted text.
4. Count errors in the summary separately: unsupported claims, smoothed gaps, wrong attribution.
5. Repeat with a second speaker whose accent differs from yours; a tool that works for one voice may not work for your candidates.

Record the counts on the score sheet in the AI note taker evaluation checklist so tools are compared on the same test.

How do you reduce errors in the first place?

• Better audio: a headset for you and a quiet room. Otter's accuracy FAQ notes that recording through its meeting bot, which takes the meeting's internal audio stream, is generally more accurate than recording through a device's speaker and microphone.
• The right language setting before the call; a transcript in the wrong language is mostly invented. See transcribing an interview in another language.
• Custom vocabulary for client, product and technical names, where your tool offers it.
• Prompts that forbid inference: "quote word for word", "say 'no evidence' if there is none", "do not rate".
• Quotes tied to the transcript. Some tools check generated quotes against the transcript automatically; in Interview Signal (we make it), a quote that cannot be found in the transcript is removed from the scorecard. Whatever the tool, a quote with a timestamp is one you can verify in seconds.

Who is responsible when AI notes are wrong?

The person who signs the scorecard or sends the submittal. AI notes are a draft of the record, not the record. Keep ratings and decisions with the interviewers, write corrections down when you make them, and remember that candidates may be able to request their interview records under privacy laws where they live. Notes you would be uncomfortable defending line by line should not be in the file; what not to write in interview notes covers the rest of that standard.