What the AI Model Provider Sees When You Use an Interview Assistant

The AI model provider sees text: the transcribed question and whatever context you pasted into the assistant, sent to the model you picked and processed under that provider's terms rather than ours. In CoPilot Interview that provider is Inception (Mercury 2.5) on the free plan, OpenAI (GPT-6 Luna) on Pro, and Anthropic (Claude Haiku 5.5) or Google (Gemini 3 Flash) on Pro+. The model receives text, not audio, and the assistant never joins the meeting, so the provider has no view of the call itself. Below: the data flow in plain words, what stays on your machine, and how to pick a model with privacy in mind.

The data flow, step by step

1. Audio in. The assistant listens to the computer's audio output for the interviewer and to your microphone for you. That audio is streamed in short segments for transcription and is not stored on our servers.
2. Transcript saved. The resulting text transcript is saved only in your Documents folder on your computer.
3. Question detected. A new question is detected after a 20-second pause, or when you press Ctrl+Shift+A.
4. Text to the model. The transcribed question plus the context you pasted, such as your resume and the job description, are sent as text to the AI model you chose.
5. Answer back. The model returns a structured answer, shown in the assistant window in about 2 seconds.

The sentence to remember: the model provider receives text, not audio. What it receives is whatever was said in the segment that formed the question, plus the context you supplied.

Which provider, by plan

Standard ($24.99 a month) adds Dual AI Mode; see the pricing page for its model list. Dual AI Mode shows answers from two models side by side, which means the same text goes to two providers. Our page on which models power interview assistants explains how this works across the market.

What stays on your machine, and what does not

Stays: the transcript file in your Documents folder, which you can delete like any other document. The assistant is a separate window that never joins the meeting, so it has no access to the participant list, the chat or anyone's video, and neither does the provider.

Leaves: audio segments for transcription (not stored on our servers), the transcribed text and your pasted context to the model you picked, and, only when you use it, the screen. Ctrl+Shift+S reads a problem from the screen, and on Pro, GPT-6 Luna can read a screenshot of an on-screen problem, so whatever is visible at that moment is part of what is sent. Close unrelated windows before using it.

What the provider does with the text

Each provider processes the text under its own terms of service and privacy policy. Those terms set how long requests are retained, whether they can be reviewed, whether they are used to improve models, and where they are processed, and they differ between Inception, OpenAI, Anthropic and Google and change over time. We do not restate them here because we cannot keep a restatement accurate; see the terms of the provider behind the model you pick. For our own handling, see the privacy policy and our privacy overview. This is a plain-language explainer, not legal advice.

Picking a model with privacy in mind

• Paste only what the interview needs. A resume with your street address, salary history or a current employer's confidential project names sends all of that with every question. Our resume storage page covers trimming it.
• Fewer models, fewer parties. Dual AI Mode sends the same text to two providers. If one provider's terms are the ones you have read and accept, use one model.
• Read the terms for the model you choose. Which model to pick covers speed and quality; add the provider's terms to that decision.
• Use a device you own. On a managed work laptop, the Documents folder and network traffic are also within your employer's view, so follow your employer's rules and keep job-search tools on a personal machine.
• Mind the screenshot feature. It sends what is on screen; use it only with the problem window in front.

Where CoPilot Interview fits

CoPilot Interview is an AI interview assistant built around preparation first: AI mock interviews from your resume and the job description are free, and live help is for interviews where it is allowed. The data path is short by design: audio streamed in segments for transcription and not stored on our servers, one transcript file in a folder you control, text to the model you chose, an email-link sign-in with no password, and an account only on paid plans. Disclosure: CoPilot Interview is our product, so weigh what we say about it accordingly.

Frequently asked questions

Does an AI interview assistant send my data to OpenAI?

In CoPilot Interview, only if you pick an OpenAI model: GPT-6 Luna on the Pro plan. The free plan uses Mercury 2.5 by Inception, and Pro+ adds Claude Haiku 5.5 from Anthropic and Gemini 3 Flash from Google. What is sent is text: the transcribed question and your pasted context, under that provider's terms.

Does the model provider hear the interview audio?

No. The model receives text. Audio is streamed in short segments for transcription and is not stored on our servers; the transcribed text is what goes to the model you chose, together with the resume and job description you pasted.

Does the provider know who I am interviewing with?

Only if it is in the text. The assistant never joins the meeting, so there is no participant list or company account attached to the request. If you pasted the job description with the company name, or the interviewer said it aloud, that text is included.

How long does the provider keep what was sent?

That is set by each provider's terms, which differ between Inception, OpenAI, Anthropic and Google and change over time. We do not restate them; check the terms for the model you pick. The transcript on your own computer is yours to delete at any time. This is not legal advice.

Can I use the assistant without sending anything to a model?

No. Generating an answer requires sending the question text to a model. You can reduce what is sent by pasting less context, choosing a single model rather than Dual AI Mode, and leaving the screenshot feature unused.