You find out which language model an AI interview tool runs on by asking the vendor in writing, because most product pages do not say, and you find out whether candidate data trains the model by reading the data processing terms rather than the privacy policy summary. The two questions are separate: a tool can use a well-known model through an API with training switched off, or a lesser-known model while retaining transcripts for its own improvement. This page gives the five contract questions, explains what the model provider typically sees, and states what CoPilot Interview does with audio, transcripts and the model you pick.
The five questions to put in writing
- Which model or models process candidate answers, and can we choose or restrict them?
- Is any candidate data (audio, transcript, scores) used to train or fine-tune a model, by you or by the model provider?
- Where is data processed and stored, and for how long by default?
- What does the model provider receive: audio, the full transcript, or only the text of answers?
- Can we delete a candidate's data on request, and how long does deletion take?
Vendors who answer these in a sentence each are usually fine. Vendors who answer with a link to a general privacy policy are telling you they have not thought about it. The models behind the main candidate-side tools are listed in which AI models power AI interview assistants.
What the model provider sees
In most tools, speech-to-text and the language model are separate services. The transcription service receives audio and returns text; the language model receives text, usually the question, the answer and some context such as the job description, and returns a score or suggestion. Whether either provider keeps that data depends on the vendor's contract with them; business API terms from the major providers typically exclude training by default, but that is the vendor's job to confirm. Our explainer on what the model provider sees walks through the data path.
Reading the data terms
| Phrase in the terms | What it usually means | Follow-up |
|---|---|---|
| "to improve our services" | May include training on your data | Ask for the opt-out and get it in writing |
| "de-identified data" | Transcripts with names removed can still identify people | Ask whether candidates can object |
| "sub-processors" | The transcription and model providers | Ask for the list |
| "retained as long as necessary" | No fixed period | Set one in the contract |
| "zero data retention" | The provider deletes inputs after processing | Ask whether it covers every model offered |
The privacy policy comparison shows how differently candidate-side tools answer the same questions; employer-side tools vary just as much.
What CoPilot Interview does
Audio is streamed in short segments for transcription and is not stored on our servers. A text transcript is saved only in the Documents folder of the computer running the app. The transcribed text is sent to the AI model the user picks: on the free plan Mercury 2.5 by Inception is the default; Pro adds GPT-6 Luna; Pro+ adds Claude Haiku 5.5 and Gemini 3 Flash. Sign-in is an email link, and an account is needed only for paid plans. For anything beyond those facts, read the privacy policy, and for team deployments ask us about data terms before a pilot.
Telling candidates
Whatever the answers, candidates deserve a short, true statement: which kind of tool is in use, that answers are transcribed and scored, that a model provider processes the text, and how long records are kept. A sentence at the start of the interview and a line in the application privacy notice cover it in most jurisdictions; some require more, and your counsel should map them.
Where CoPilot Interview fits
CoPilot Interview is the AI interview assistant candidates use to prepare and, in Interviewer Mode, the tool interviewers use to score; both run on the data path above, with the model chosen by the person running the app. If you are evaluating us for a team, see Interviewer Mode and ask for the data terms. With any vendor, get the five answers in writing, and follow your employer's rules on candidate notice.
Disclosure: CoPilot Interview is our product, so weigh what we say about it accordingly.
Frequently asked questions
Which LLM do AI interview tools use?
Most do not say on their product pages. Ask in writing. Many run on well-known models through business APIs; CoPilot Interview offers Mercury 2.5 by default, with GPT-6 Luna, Claude Haiku 5.5 and Gemini 3 Flash on paid plans.
Is candidate data used to train the model?
It depends on the vendor's contracts with its model providers and on its own terms. Business API terms often exclude training by default, but get the vendor's confirmation in writing.
What does CoPilot Interview store?
Audio is streamed in short segments for transcription and not stored on our servers; a text transcript is saved only in the user's Documents folder; transcribed text goes to the chosen model.
What should we tell candidates?
That an AI tool transcribes and scores answers, that a model provider processes the text, and how long records are kept, in one sentence at the start and a line in the privacy notice.
Can a candidate ask for their data to be deleted?
Ask the vendor how deletion requests are handled and how long they take, and make sure your own retention schedule covers reports and transcripts.
Try Interviewer Mode on your next screen
Role-specific questions, live rubric scoring with evidence notes and a structured report at the end of the call. Free in the desktop app for Windows and macOS, no signup.
Get the desktop app