Wharton Generative AI Labs’ November 12, 2026 preprint describes Primer, an experimental system with a classroom role reversal: students explain concepts to an AI character that needs help. Its Teachback mode offers teachers a design for explanation practice, not evidence of improved learning.
The student explains; the AI plays the learner. Instead of waiting for the chatbot to produce a polished answer, the human has to make an idea understandable to someone else.
As a design proposition, Primer suggests a useful way to judge educational AI: look at the thinking it requires from the student, not just the answers the model can supply.
What students actually do in Teachback
Primer begins with the teacher. In Wharton’s account of the workflow, an educator supplies the learning goals, prior knowledge, likely misconceptions and lesson materials, then reviews a proposed design before it becomes a simulation.
The activity has three stages: a mentor prepares the student, a practice character receives their explanation, and a coach discusses the encounter afterward using session records. The teacher defines what the exercise is for; the characters are meant to keep the conversation serving that purpose.
The published Teachback prompt gives the practice character ways to make the student work: request an example, ask how an idea applies, or try applying it incorrectly. The student then has something specific to clarify or correct. These are instructions for intended behaviour, not observations from a successful classroom lesson.
That makes it a different activity from teacher-assigned Guided Learning in Google Classroom. Here, the defining question is not simply whether the AI guides the student towards an answer, but whether the student can explain the answer to a counterpart.
Why the AI pupil must resist being too helpful
A practice pupil who immediately supplies the perfect explanation has rather missed the assignment. The authors report that development testing exposed agents helping too quickly. In this setting, extra assistance can remove the very work the student was supposed to attempt.
The design challenge is therefore more precise than making a chatbot friendly or knowledgeable. A character needs enough confusion to give the student something to resolve, without becoming so baffled that the exercise stalls. An entertaining exchange alone would not settle whether that balance is right.
The pupil’s confusion is a role it plays, not a claim that the underlying model is being retrained. Its apparent understanding cannot independently establish the student’s mastery.
Feedback needs similar scrutiny. The feedback prompt calls for coaching based on available transcripts and acknowledgement of missing evidence. The authors also warn that failed handoffs can leave fluent dialogue without the records needed to support feedback. A sensible review would match a coaching claim to the actual exchange, not just its reassuring tone.
What the research establishes about learning
According to Wharton’s development summary, K–12 educators built and tried simulations in two-hour sessions between February and June 2026. Their feedback informed changes. Those sessions were not a classroom learning-outcomes trial, and the authors say they do not establish improved learning.
Nor did Primer invent the idea of teaching a computer. The October 2025 Chrysalis preprint describes an exploratory study with 36 participants comparing AI tutoring and learning by teaching. It provides a relevant precedent, not evidence that Primer works.
Wharton says current adaptation stays within one encounter; practice across different cases is future work. A conversation that adjusts its next question should not be mistaken for a validated personalised curriculum.
Voice adds another qualification. The Primer paper cautions that hesitant speech is not direct evidence of weak understanding, and audio misunderstandings can distort the exchange. For educators, that means separating the concept being assessed from how smoothly a student speaks.
A useful future evaluation would ask whether students can later explain or apply the concept without help. The relevant outcome is the learner’s independent performance, not the AI character’s satisfaction.
What educators can inspect before considering a pilot
The public repository contains code and a Teachback template, but no preinstalled experiences. Running the documented setup requires hosting, sign-in configuration and funded model access. Open-source availability makes the design inspectable; it does not supply a free, ready-to-assign classroom service. The documentation does not establish a fixed cost per student.
The model-configuration guide also draws an important boundary: backend or text agents can be adapted to other providers or local models, while the documented voice experiences require OpenAI. Hosting the application yourself would not, by itself, make its voice interactions fully local.
Wharton describes Primer as experimental and says student-privacy compliance and cross-language performance remain unestablished. Schools would need their own review before involving students. Rules on permitted AI assistance and independent homework remain a separate question from whether an activity is technically possible.
For an educator, a useful first decision is whether the design creates an explanation worth practising. Choose a concept, inspect the follow-up the character is supposed to ask, and identify what would count as a convincing answer. Then check whether the proposed feedback could actually be supported by that exchange.
Primer’s most interesting proposition is not an AI that knows more than the class. It is an AI willing to play the pupil—and leave the explaining to the person who needs the practice.