Before the live interview, the prep tools are already recording

What is the real question now?

What matters most in online interviews and exams is no longer just whether AI is present, but what each system says it does with recordings, transcripts, and detection. That distinction matters because the same technology can be framed as a practice aid in one context and a scored or governed assessment tool in another.

The newest product pages from OpenAI, Greenhouse, and Turnitin all point in that direction. Each describes useful automation, but each also draws a boundary around when the tool is assistive, when it is evaluative, and when a human or a separate process still has the final say.

ChatGPT is built for preparation, not a substitute for the assessment

According to OpenAI, ChatGPT can accept documents, spreadsheets, presentations, images, and audio files, and users can ask it to summarize, compare, analyze, transcribe, and discuss them. That makes it a broad preparation tool for interview practice, lecture review, and other pre-assessment work.

OpenAI also limits audio uploads to paid ChatGPT subscriptions and workspaces, including Enterprise, and says they are not available on the Free plan. Availability can also vary by workspace settings, region, client version, and selected model. In other words, even the preparation layer is not uniform; access depends on the account and environment.

The help center also makes clear that uploads have technical boundaries. Failed attempts can count toward an upload-rate limit, saved-file storage is separate from that limit, and an audio file must use a supported format with a valid, decodable stream. WebM and MP4 files identified as video are not supported as audio uploads. That is a reminder that even before any interview or exam begins, the tool is designed around ingestion and processing rules, not around replacing the actual assessment setting.

Greenhouse draws a line between structured interviews and human decisions

Greenhouse describes its Voice AI interviews as conversational, role-specific, and scheduled on the candidate’s own time. It presents them as a structured two-way voice interview rather than a one-way recording or chatbot. Before the interview, the hiring team defines the questions and evaluation criteria, and every candidate is asked the same role-specific questions and evaluated against the same rubric.

That model is still tied to human oversight. Greenhouse says a recruiter reviews the Voice AI evaluation alongside application materials, humans make every hiring decision, and Voice AI does not make hiring decisions on its own. It also says candidates can opt out of Voice AI at any time and still continue through the standard process, with no impact on decision making.

The company’s AI-interview guidelines go further by saying that, with candidate and interviewer consent, it uses AI to transcribe and summarize interviews to reduce note-taking burden, speed up hiring, and mitigate bias from incomplete feedback. Its Notetaker product is described as an optional automated virtual meeting participant that can join supported interviews, record and transcribe audio and video, and generate AI-powered notes, summaries, and suggested responses mapped to scorecard questions or other interview materials. But the terms still say it is assistive, not a replacement for human interviewer judgment, and not intended as an automated decision-making system under the cited EU rule or as an automated employment decision tool under applicable law.

Transcripts, retention, and review are now part of the process

Greenhouse also shows that the operational life of a recording does not end when the interview ends. Interview recordings and transcripts may be accessed by the Talent Acquisition Manager, hiring manager, and others involved in evaluating the role, and sometimes by Security or Legal for fraud detection, prevention, or compliance review. Its Notetaker terms add that recordings are retained according to the customer’s Notetaker retention settings and then automatically deleted or made unavailable, while transcripts and AI-generated outputs follow the customer’s Greenhouse retention settings.

This matters because practice and scored sessions now differ not only in intent, but in how data is handled afterward. A prep recording might help a candidate rehearse a response or help a team summarize feedback. A governed interview transcript, by contrast, can become part of a review record with retention rules, access controls, and human evaluation attached to it.

The practical lesson is straightforward: candidates should treat tools like ChatGPT and Greenhouse Notetaker as prep aids unless the employer’s process explicitly says otherwise. If the company offers an opt-out or a standard process, that is the path to use when you want a human-led interview rather than an automated layer.

Turnitin shows that detection is still evolving, not settling

Turnitin’s AI writing detection model page adds a different version of the same tension. The October 2026 update improved recall while maintaining a low false-positive rate, with no setting changes required and no visual changes to the report. That means the detector can change under the hood without changing the surface experience.

Turnitin also says the update does not retroactively change previously generated reports, and existing submissions must be resubmitted to see the new score. It further says the detector can identify likely use of AI bypasser tools, which are meant to modify AI-generated text to appear more human-like and evade detection. On top of that, it now supports Spanish, Japanese, Arabic, Turkish, Dutch, and German submissions.

For students, that means a detector is not a permanent verdict, and for schools it means the output of the system needs context, not blind trust. A score is part of a process, not the whole process. If a result seems wrong, the useful move is not to guess how the detector works around it, but to ask for the institution’s review path and supply the material the school asks for.