Can an AI note taker handle a group interview with several candidates? Consent, speaker labels and records

An AI note taker can capture a group interview, but it cannot be trusted to keep several candidates apart or to score them. On a video call where each candidate joins from their own device, the platform knows who is talking and speaker labels usually hold; in a room with one microphone, labels slip as soon as people talk over each other. Get consent from every candidate, split the transcript so each candidate's record holds only their own contributions, and let trained observers do the scoring on a grid. Use the transcript to check what was said, not to decide.

How to design the session itself is in how to run a group interview. The opposite case, several interviewers and one candidate, is covered in AI note takers for panel interviews. This page is about many candidates in one recording.

Can an AI note taker handle a group interview with several candidates?

It can help with some jobs in a group interview and should be kept away from others.

How do you get consent from every candidate?

Everyone in the room or on the call is a party to the conversation. Federal law allows recording with one party's consent, but California, for example, prohibits recording a confidential communication without "the consent of all parties" (California Penal Code 632), and a group may include candidates from several states. Our two-party consent states page lists the stricter states.

1. Say it in the invitation. "The group session will be transcribed so our observers can check what was said. The notes are used only to assess this role. If you would prefer it was not, reply and tell us; it will not affect your application."
2. Ask again at the start, to the whole group, and look for a clear yes from each person, not silence. On video, ask each candidate by name.
3. Plan for a "no." If one candidate declines, do not record the session with a note taker running around them; switch it off and rely on observers' notes. Alternatively, offer that candidate a place in another group.
4. Note the consent on each candidate's record: date, session and "agreed" or "declined."
5. Ask candidates not to record each other. They hear one another's answers, so add a line about keeping the session confidential.

This is general information, not legal advice.

Why do speaker labels break down with many voices?

Most tools separate voices with speaker diarization: they cut the audio into stretches, describe each voice as a numerical fingerprint, and group similar fingerprints under one label. It works well for two people taking turns. A group exercise is close to the worst case:

• Unknown speaker count. The system has to guess how many people there are. Amazon Transcribe, for example, separates up to 30 speakers and lets you set how many to expect (AWS, as of October 2026). A wrong guess merges two candidates or splits one.
• Short turns. "Yes," "good point" and "can I add" are too short to fingerprint and get attached to whoever spoke nearby, which is exactly the evidence of listening you want to score.
• Crosstalk. Group discussions overlap. Overlapping speech is the hardest part for diarization and often ends up under one name.
• Similar voices. Candidates of similar age and accent can collapse into one speaker.
• Distance. A laptop microphone at one end of a table hears near voices clearly and far voices poorly.

Video is easier because the platform knows whose microphone is active and can label lines by participant name, as long as each candidate joins from their own device. A room with one microphone has no such help. Our explainer on what speaker diarization is covers how it is measured and how to test a tool.

Ways to make labels hold up:

• On video, ask every candidate to join from their own device with their full name displayed.
• In a room, start with a round of names ("I'm Dana, I'm applying for the shift lead role") so each voice is introduced, and tell the tool the number of speakers if it lets you.
• Have the facilitator name people when handing over: "Sam, what do you think?"
• Keep the observer's written note of who said what for the moments that matter.

How do you keep one candidate's record free of other candidates' data?

A group transcript is personal data about every candidate in it. Pasting the whole thing into each candidate's file copies everyone's answers into everyone's record, which creates two problems.

• Data minimization. Under the GDPR, Article 5(1)(c), personal data must be "limited to what is necessary." Candidate A's file does not need Candidate B's answers.
• Access requests. Article 15(4) of the GDPR says a person's right to a copy of their data "shall not adversely affect the rights and freedoms of others." If a candidate asks for their interview data, you will have to redact the others anyway. See candidate data access requests.

A workable split after each session:

1. Correct the speaker labels in the tool before anything is copied.
2. For each candidate, extract their own contributions with timestamps, plus just enough of the surrounding question or prompt to make sense of them.
3. Put the extract, the observers' scores and the evidence notes on that candidate's record.
4. Keep the full transcript, if you keep it at all, in one restricted location for your retention period, then delete it.

Do not delete the full transcript too early. The EEOC's recordkeeping rule says that when a discrimination charge is filed, the relevant records include "application forms or test papers completed by an unsuccessful applicant and by all other candidates for the same position" (29 CFR 1602.14), and covered employers must keep hiring records for at least one year in any case. A shared session record may be exactly what shows that every candidate was treated alike.

Why should the observer grid still do the scoring?

Group interviews are scored on behaviors: building on others' ideas, bringing quiet people in, keeping the group on task. Those show up as who spoke to whom, when, and how others reacted, much of which never reaches a transcript. An AI summary also tends to credit the last or loudest speaker with an idea the group reached together.

• Observers score each assigned candidate alone, on the grid, after each exercise.
• Only then open the transcript, to check a specific claim ("Did Priya propose the rota first?").
• Never let an AI-generated ranking of the group enter the wash-up. A tool that scores or ranks candidates can bring automated decision rules into play; see NYC Local Law 144 and AI note takers.
• Score what each person did. A low score needs the same written evidence whether or not a transcript exists.

We make Interview Signal, and it is the wrong tool for scoring a group: it is built for one candidate at a time and labels everything from the call's sound as the candidate, so several candidates on one call share one label. In a room, both sides go on one microphone, so its quotes cannot be checked against a specific candidate. Use it, if at all, for each candidate's one-to-one segment.

A setup checklist for a transcribed group interview

1. Decide the purpose: evidence lookup for observers, not scoring.
2. Choose video with one device per candidate where you can; for in-person, a microphone placed centrally and a name round at the start.
3. Send the notice in the invitation and ask each candidate again on the day.
4. Turn off automatic summary sharing so no candidate receives notes about the others.
5. Brief observers that the transcript does not replace their notes.
6. After the session, fix labels, split the transcript and file each extract with that candidate's scores.
7. Set a retention period for the master transcript and record where it is kept; the assessment center guide has more on running larger multi-exercise days.