An AI notetaker in a coding interview records the call. It turns your speech into a transcript and AI notes. Interviewers can use those notes to fill in their scorecards. A hiring manager may read the summary and never watch the call. The notetaker hears your words. It doesn’t read your editor. Your code lives in a separate record. CoderPad’s keystroke playback is one example. A decision you make silently in the editor shows up in the code. It doesn’t show up in the notes. Fix that with one habit. Say each decision in one sentence before you type it, with the reason and the cost. Spell out any library or API name that a transcript could garble. In the last two minutes, recap what you built and what you left out. You can usually opt out of the recording. Ask your recruiter what that changes before you decide.
What the notetaker produces and who reads it
Start with what the tools say they do. Greenhouse’s Notetaker setup guide was updated on November 4, 2026. Its notetaker joins Zoom, Google Meet and Microsoft Teams interviews. It makes a transcript and AI-generated notes. Interviewers can use those notes to complete scorecards. Phone and in-person interviews aren’t recorded. Admins set how long recordings are kept, up to 60 days.
Greenhouse also describes how it runs the tool in its own hiring. Greenhouse sells the product, so read this undated page as a vendor’s account. The interviewer, the recruiting team and the hiring manager can see the recording. Managers may rewatch it for calibration. Recordings are deleted after 60 days. Transcripts are kept for two to three years, depending on location. Recordings may also be used to detect impersonation or deepfakes.
The summary can matter more than the recording. HR Dive reported on September 18, 2026 on an analysis by the outplacement firm Challenger, Gray & Christmas. It found a hiring manager may read an AI summary without watching the interview. That summary can become the first evaluation. The same report notes that summaries can drop context and nuance.
So the notes feed the scorecard. The scorecard then feeds the debrief. Our post on why you can be rejected after a good interview covers that last step. The debrief works from what was written down.
The transcript holds your words, and the pad holds your code
Coding rounds have two records. The notetaker records the video call. The coding tool keeps its own log. CoderPad’s playback mode tracks every keystroke of every participant in a pad. Interviewers can replay it after the interview and share it with teammates. Its docs describe code edits, runs and pastes. They don’t describe any audio.
None of the notetaker pages above mention live coding or screen sharing. Greenhouse also offers an audio-only recording option, added in October 2026. Its transcripts and AI notes still generate. Either way, the notes come from what people said. Typing that nobody explains has no words to transcribe.
Here’s the running example. You’re interviewing for an Android role. The prompt is an in-memory cache for image thumbnails with a fixed size. When it’s full, it should drop the thumbnail viewed least recently. You start with a plain HashMap. Ten minutes in, you see it has no order. So you switch to a LinkedHashMap in access order and override one method.
class ThumbnailCache(private val maxEntries: Int) {
private val map = object : LinkedHashMap<String, ByteArray>(16, 0.75f, true) {
override fun removeEldestEntry(eldest: MutableMap.MutableEntry<String, ByteArray>) =
size > maxEntries
}
fun get(url: String): ByteArray? = map[url]
fun put(url: String, bytes: ByteArray) { map[url] = bytes }
}
The true in the constructor is the whole decision. Per the Java docs for LinkedHashMap, it orders entries by access instead of insertion. That makes the map a natural fit for an LRU cache. Run it with a limit of 2. Put a and b, read a, then put c. The cache keeps [a, c] and drops b.
Say you made that switch in silence. The playback shows the edit. The transcript has a ten-second gap. The AI notes might say “candidate implemented a cache.” They won’t say you caught the ordering problem yourself, or why access order fixes it. That catch is the best evidence in the round. The notes miss it.
Technical words are the ones a transcript gets wrong
Speech-to-text has known weak spots. A February 2026 alert from the employment law firm Littler lists them. These tools can misread industry terms and acronyms. They can struggle with accents. Littler advises employers to correct records before relying on them. Littler represents employers. The alert is written for them.
A coding round is dense with exactly those words. “LinkedHashMap” can land as “linked hash map” or worse. “LRU” can become a different word. “removeEldestEntry” has little chance. You can’t control the model. You can say the term once in plain words next to the name. Saying “LinkedHashMap, the map that keeps order” helps. The summary keeps the plain words even if the class name breaks.
Opting out is allowed, but it changes the record
The vendors above give candidates a way out. BrightHire’s candidate FAQ, from September 2026, sends an email with an opt-out link before the first interview. It says choosing not to be recorded has no impact on your candidacy. Ashby’s consent docs let each employer pick opt-in or opt-out as the default. The page shows no date. Its default text says opting out doesn’t affect candidacy. The employer can edit that text.
Metaview sells a notetaker. It also adds a caution. Its May 2026 post on recording interviews warns that candidates who decline get judged on thinner evidence. It tells interviewers to write fuller notes to make up for it. So if you opt out, the interviewer’s handwritten notes become the whole record. Your narration matters even more.
Where you live can change the rules. HR Dive’s September 14, 2026 report names California, Florida and Illinois. Those states require every party’s consent to a recording. The Littler alert says about a dozen states have that rule. That’s why some employers ask you to opt in instead. None of this is legal advice. It explains why the consent email looks different from company to company.
A sample script for an AI notetaker in a coding interview
Here’s the thumbnail cache round, said out loud. Adapt the details to your own problem.
First, state the plan before any code.
“I’ll build a fixed-size cache keyed by image URL. When it’s full, it evicts the thumbnail viewed least recently. That’s an LRU cache. I’ll start with a HashMap to get get and put working. Then I’ll handle eviction.”
Then, when you catch the problem, name it before you touch the code.
“Wait, a HashMap has no order, so I can’t tell which entry is oldest. I’m switching to LinkedHashMap, the map that keeps order. I’ll pass true as the third argument. That orders it by access, so a read counts as use. Then I override removeEldestEntry to drop one entry when we go over the limit. Get and put stay constant time. The cost is two extra links per entry.”
Run a test and say what it proved.
“Limit of two. I put a and b, read a, then put c. It kept a and c and dropped b. So the read moved a to the front. That’s the behavior we want.”
Close with a recap in the last two minutes.
“So, to sum up. I used a LinkedHashMap in access order, with eviction past the limit. Constant time for get and put. One thing I didn’t handle is thread safety. LinkedHashMap isn’t synchronized. If two threads load images, I’d guard it with a lock first.”
Why each line of the script works
The plan line gives the summary a topic sentence. If the notes say one thing about your round, it should be what you set out to build. It also names the pattern, LRU, in plain words before any code exists.
The catch line carries the most weight. It puts your bug, your fix and its cost into words in the order you thought them. A reader of the notes can see you found the problem yourself. Without that line, the switch exists only in the keystroke log. A hiring manager reading a summary won’t replay that.
The test line turns a run into evidence. “It kept a and c” is a result anyone can check later. The recap puts your decisions in one place, near the end of the transcript. It also names what you left out. The Java docs say LinkedHashMap isn’t synchronized, so the point holds. Naming the gap yourself puts it in the notes as your call. Otherwise the interviewer may log it as a miss.
Narration has a limit. Don’t read every line of code aloud. That buries the decisions in noise. It also slows you down. Say one sentence per decision and stay quiet while you type the obvious parts. Our post on taking a hint in a coding interview covers the other time to speak up. That’s when you’re stuck.
Three questions to ask your recruiter first
BrightHire’s consent email comes before the first interview. Some companies put the notice in the calendar invite. Reply to the recruiter with three questions before the round.
“Thanks for the invite. I see the interview may be recorded with a notetaker. Three quick questions. Will the AI notes or transcript be shared with the hiring manager or the debrief? How long is the recording kept? And if I opt out, will my interviewer take notes by hand instead?”
The answers tell you how much weight the notes will carry. Whatever they say, the habit is the same. Say each decision before you type it, so both records tell the same story.