Viva

Viva by NYU Stern

The oral exam is no longer a small-class luxury.

Assess what students can explain, not only what they submit.

Students discuss their projects, cases, or course topics with an AI examiner. Viva proposes rubric-based grades with the evidence behind them. You review, adjust, and decide what students see.

Create your first exam Hear a sample exam

Open to instructors at any institution: sign in with a Google account, a Microsoft account, or an email address. No approval step, and no charge during the alpha. Access, costs, and data

Developed at NYU Stern. More than 1,000 students have taken a Viva exam. The research behind it

Taking a Viva? Use the personal exam link your instructor shared with you. You do not need to sign in, and you do not need an account.

A student's grade on the instructor's review page in Viva: 14 of 20, with a score and a justification on each of five rubric dimensions, awaiting instructor confirmation.
The review page for Maya's exam. Hear the exam it grades

Sample exam

Hear an exam, then check the grade against it.

A 75-second excerpt from Maya's exam on her team project, and the grade Viva proposed for it.

1. The conversation
  1. ExaminerMaya, your team's model flags patients who are likely to be readmitted within thirty days. Suppose it flags twice as many patients as the nurses can call. What do you change first?
  2. StudentUm... I'd probably, uh... raise the threshold? So it only flags the, the highest-risk patients.
  3. ExaminerThat shortens the list. Is the highest-risk patient always the one a call helps most?
  4. StudentHm. Not necessarily, I guess? Like, someone who's really sick might come back anyway... no matter what. So you'd want the patients where the call actually... changes something.
  5. ExaminerHow would you find those patients?
  6. StudentUh... I'm not sure. Maybe compare the patients who got a call with, um... similar ones who didn't?
  7. ExaminerWhat could make that comparison misleading?
  8. StudentOh. Wait... the nurses chose who to call. So the two groups probably weren't, like, the same to begin with.

Recorded with a synthetic student voice. The examiner follows up one question at a time rather than giving the answer away.

2. The proposed grade 2 of 5 rubric dimensions

Problem framing

3 of 4, Proficient
“So you'd want the patients where the call actually... changes something.”

Moved from ranking by risk to ranking by the effect of the call, after one follow-up.

Claude 3, Gemini 3, GPT 3Agreed

Experimentation

2 of 4, Developing
“Wait... the nurses chose who to call. So the two groups probably weren't, like, the same to begin with.”

Named the selection problem once asked, but not how to get a fair comparison.

First scores 2, 3, 2; after deliberation 2, 2, 2Agreed after deliberation

Illustrative grade for this excerpt. It stays a proposal until you release it.

How it works

Three steps, from course material to released grades.

1

Set up the exam

Add topics and a rubric, or let the assistant draft them from your materials. Take the exam yourself, then generate one link per student.

The exam's Content view in Viva: four topics, with Generate with AI open, and the rubric's dimensions.
2

Students talk it through

Each student opens their link, with no account, and talks with the examiner about their own topics for the length you set.

A student's exam page in Viva: what to expect, how to prepare, and the button to begin.
3

Review and release

Check the proposed scores against the quotes and the recording, adjust any mark, and decide when students see it.

The exam's Results view in Viva: counts of graded and flagged students, and the control that releases grades to students.

Screens from Viva with a sample exam.

Create your first exam Hear the sample exam

Grading

Grades you can check before any student sees them.

  • Evidence for every mark

    Each score quotes the transcript. You can play the moment a quote comes from and read the whole conversation.

  • Disagreement is flagged, not hidden

    Three models grade independently, then compare their evidence. When they still disagree widely, the grade is flagged for your review and is not released until a person has looked at it.

  • You make the final call

    Change any mark with a note to the student. Marks wait for your release by default, and you can sample unflagged grades too.

The grade page's council table in Viva: each model's first and revised score on five dimensions, the chair's score, and whether the models agree.
On the grade page: each model's first and revised score beside the final mark.
How the grading models and the examiner work

Three models, then you

Claude, Gemini and GPT each score the transcript against your rubric on their own, then read each other's scores and evidence and revise or defend their own. A chair model writes the grade and the feedback. A session is flagged when the models are still more than two points apart on a dimension, or more than three on the total.

Transcript

The full conversation, with the student's name and IDs removed

Round 1
ClaudeGeminiGPT

Each scores every dimension alone, quoting the transcript

Round 2

Each reads the others' evidence and revises or defends its scores

Chair

Writes the grade, strengths and next steps

You

Review and release

Adjust any mark, then release to students

An examiner kept on track

Each exam is a sequence of phases, one per topic. Each phase carries only its own topic, and the conversation moves to the next phase when the examiner closes the current one.

  • One question at a time. The examiner is set up to ask a single question and wait for the answer, and students can ask for any question to be repeated.
  • Topics fixed before the call. Each student's topics come from their roster row or a reproducible draw from your pool, set when their link is generated. The model never picks at random.
  • Recorded and transcribed. Every session keeps its transcript, and spoken sessions keep their audio, so you can hear the answers a grade relies on.

Research

Developed and used in NYU Stern courses.

In the first deployments, described in a field report, most students agreed the exam tested their actual understanding, deliberation brought the grading models closer together, and grading cost under one dollar per exam.

Panos Ipeirotis and Konstantinos Rizakos, “Scalable and Personalized Oral Assessments Using Voice AI,” Communications of the ACM, November 2026 (forthcoming).

FAQ

Questions to settle before you try it.

Can I use it?

Who can sign up?

Any instructor, at any institution. Sign in with a Google account, a Microsoft account, or an email address, and your account is created the first time. There is no approval step.

What does it cost?

Nothing during the alpha: we cover the voice and model costs of alpha users' exams. Your profile page shows what your exams would have cost.

Are there usage limits?

New accounts can own up to 3 exams, generate links for up to 50 students in total, and allow up to 3 attempts per student. Each day, a new account can use up to 100 AI-assisted actions, 300 student emails, and 50 MB of uploads. Students who already have links never count twice. To raise the limits, use "Request higher limits" on your profile page.

Can I use it responsibly with my students?

What is recorded?

The transcript of each conversation, and the audio of each spoken one. Before they begin, students are told the conversation is recorded and that you can listen to it and read the transcript. Opening or downloading a recording is logged.

Do the grading models see who the student is?

No. The student's name, NetID, email and university ID are removed from the transcript before Claude, Gemini and GPT grade it.

Can I adjust the exam for a student?

You set the exam's length and how many attempts students get, and you can allow an individual student another attempt.

How does it support academic integrity?

Students explain and defend their reasoning in real time and answer follow-up questions on their own work, which is much harder to hand off than a written answer.

Can I run it in my course?

What do students need for a spoken exam?

A computer with a microphone and a current browser; we recommend Chrome. A quiet room and headphones help. The exam page checks microphone access before the exam starts, and a failed check uses nothing up.

What happens if the connection drops?

The page reconnects on its own. If the call still ends early, the student comes back to it and the examiner starts again at the first topic they had not finished, without using another attempt.

How do students know what to expect?

Point them to the sample on this page. You can also take the exam yourself as a preview before they do.

How do grades get into my gradebook?

Download a CSV with each student's total, per-dimension scores, flags and feedback, and import it into your gradebook.

Run your first Viva.

Set up an exam, take it yourself as a student would, then share the links.

Create your first exam Hear the sample exam