Can an AI Notetaker's Summary Go in the Hiring File?

Why doesn't "nervous and rehearsed" belong in the file?

Because both halves are conclusions about a person rather than records of an interview. Nerves track the room: a high-stakes hour with strangers who decide the outcome. They do not track the work. "Rehearsed" penalizes preparation, which is the thing the invitation asked for. Neither phrase can be checked against the transcript, and neither tells the next reader one thing the candidate said.

Underneath the sentence is a credibility judgment taken off demeanor, and demeanor is poor at it. A meta-analysis synthesizing 206 documents and 24,483 judges found people average 54% correct lie-truth judgments when discriminating in real time with no special aids or training, correctly classifying 47% of lies as deceptive and 61% of truths as nondeceptive ³. The same synthesis reports that people "appear deceptive when motivated to be believed." The candidate who most wants the job is the one most likely to read as trying too hard.

The second problem is who the word describes. Nervousness is also what a stutter looks like from the outside, what a flat delivery looks like on an autistic candidate, and what a fourth language sounds like under time pressure. The EEOC's guidance says conditions like major depression, PTSD, bipolar disorder, schizophrenia and OCD "should easily qualify" as ADA disabilities, and that discriminating against someone simply for having one is illegal. "This includes firing you, rejecting you for a job or promotion" ². A demeanor note leaves a written trace of the thing that is not allowed to drive the decision.

This is the same error as reading a candidate's gaze. Eye movement off camera is not evidence either, and for the same reason: the observation is real, the inference is not. "Rehearsed" carries an extra wrinkle, because everyone now arrives coached and every STAR answer sounds the same whether or not a model wrote it. Polish separates who prepared from who did not. It does not separate who can do the job.

Fix the summary you already have, this week

Edit it before the file closes, and record who edited it and when. A notetaker summary is a draft until a person reads it, and it becomes the hiring record the moment nobody does. Delete the demeanor sentence, keep every line of what the candidate said, and add what the summary left out: what each answer settled, and what it left open.

One limit applies before you open the file. Under the EEOC's recordkeeping rule, personnel records are preserved "for a period of one year from the date of the making of the record or the personnel action involved, whichever occurs later," and where a charge or an action has been filed, "the respondent employer shall preserve all personnel records relevant to the charge or action until final disposition of the charge or the action" ¹. So this is a same-week act rather than a same-quarter one. If a complaint has already arrived, stop: preserve what exists, correct it by dated addendum rather than by deletion, and take the question to counsel.

Timing matters for a second reason that has nothing to do with the rule. A line removed the afternoon it was written is housekeeping. The same line removed after a candidate asks why they were rejected is a different document with a story attached to it, and whoever has to explain the rejection to that candidate inherits both.

While you are in there, read the rest of the summary for the same class of sentence. Notetaker output editorializes in a consistent direction: engagement, confidence, energy, pace, "seemed unsure," "took a while to warm up." One line is rarely alone, and the next one is usually sitting in somebody else's file.

Which part of the notetaker's output is the actual record?

The transcript. It logs what was said, line by line, against a timestamp, and anyone can check it against the recording. A summary compresses that, and compression is where a sentence nobody spoke gets attached to somebody's name. Keep the transcript as the artifact, treat the summary as a draft that needs a reader, and never let the draft be the only thing that survives.

That is not a complaint about one vendor. It is what abstractive summarization does. Work on faithful dialogue summarization reports that even with pre-trained models, "substantial amounts of hallucinated content are found during the human evaluation" ⁴. A meeting summary is dialogue summarization with speaker attribution on top, so the error has somewhere specific to land: a hedge becomes a position, a question becomes a claim, and an interviewer's own aside becomes the candidate's answer.

So the summary is a second document, machine-authored, unsigned, filed beside notes that all carry a human name. Six months later nobody reading the file knows which sentences a person stood behind. That is the whole risk, and four minutes with the transcript open beside the summary retires it.

Whether you may run the tool at all is a different question with a different answer, and it is what most of the published guidance is actually about: notice, consent and recording law by jurisdiction. The mirror image, where candidates record your interviews and run them through AI, turns on the same statutes. Neither settles what the resulting summary is allowed to say.

Write this instead: the swap, role by role

Replace the impression with the exchange that produced it. The test is whether a second reader could disagree with you by opening the transcript. "Seemed nervous" fails it, because there is nothing in the record to check. "Asked what the 12% was driven by, then restated it without the driver" passes, and it still means something in March when nobody remembers the conversation.

What the swap looks like in five different rooms:

• Financial analysis. Out: "shaky when pushed on the model." In: "Could not say what the 12% growth was driven by; found it in the segment note when asked where the figure came from."
• Software engineering. Out: "low confidence, hedged constantly." In: "Walked the cache invalidation path; raised concurrent writes only after a prompt, then named the test that would catch it."
• Marketing. Out: "over-rehearsed on positioning." In: "Described the campaign listed on the resume; asked what it moved, answered reach and impressions rather than pipeline."
• Healthcare revenue cycle. Out: "hesitant, unsure of herself." In: "Conceded the claim instead of appealing it, and said why: the chart note did not support medical necessity."
• Legal operations. Out: "nervous under pressure." In: "Flagged that the playbook and the signed precedent disagreed, and asked which one governs before taking a position."

Every line on the right is checkable, job-related, and readable back to the candidate without incident. Every line on the left is a mood with a name attached. The right-hand column is also the only one two interviewers can produce the same way, which is why the panel gets briefed on one standard before the loop rather than after it. An unbriefed panel writes demeanor notes because nobody told it what else to write.

Decide what the interview record is evidence of

Put one line in the interview kit: the record holds what was asked, what was answered, and what the interviewer concluded from the answer, never how the candidate seemed. Then name who edits the notetaker's summary, and by when. A standard applied after the sentence is already in the file is not a standard. It is one person's second thoughts about one candidate.

The working test is parity: write only what you would stand behind if the candidate read it, because in a growing number of places they can ask to. In California, the exemption that kept job-applicant information outside the CCPA became inoperative on January 1, 2023 by its own terms ⁵, and the right to know reaches "the categories and/or specific pieces of personal information they have collected about you" ⁶. In Illinois, an employer that uses AI to analyze a video interview must notify the applicant beforehand, explain "how the artificial intelligence works and what general types of characteristics it uses to evaluate applicants," obtain consent, and on request delete the interviews within 30 days ⁷. Whether a notetaker riding along on your video calls falls inside that act is a question for counsel, and the answer moves.

That is not only a compliance argument. It is why candidate-facing reports change what gets written at all: a note authored for one reader and a note authored for two are different notes, and the second one is better evidence. The same discipline is what stops a panel from rejecting people for sounding like AI. In both cases the fix is to write the observation down and leave the inference off the page. See how Olive measures this.

A perfectly written interview note holds what a candidate said about their work, under time pressure, in front of strangers. It does not hold the work. A transcript can show someone describing how they would check a confident claim. It cannot show them checking one. Every fix in this section improves the record. None of them closes that gap.

Read the evidence >