Harvard's AI Faculty Avatars: Who Controls the Digital Double?

Published on August 28, 2026 · Reading time: 16 minutes

What did Harvard actually build?

HBS Foundry is an eight-week online course about starting a company, costing $699. Participants practise: pitching to an investor, running a sales conversation, sitting through a board meeting. Their practice partner is a video avatar of a Harvard Business School instructor, built by the AI video company HeyGen from interviews and recording sessions with the instructors themselves. Human instructors also run live sessions every week.

The guided version has run since April 2026, and Harvard says at least 760 founders have taken it. A year of the Harvard MBA costs around $84,760 in 2026-2027, so Foundry costs less than one percent of that.

The project director, Katharina Rings, has said the team first pictured something closer to a ChatGPT-style adviser, and that early users asked for more structure. Tom Eisenmann, a faculty co-chair, has said the avatars were built to challenge people rather than agree with them.

Jeff Bussgang, a senior lecturer and co-founder of the investment firm Flybridge Capital, said yes to a copy of himself. He has called his own avatar creepy, while adding that students like it. Another instructor, Shikhar Ghosh, told Fortune that the point was never to build a replacement for him.

It is worth stating the case for the thing. Foundry answers a real constraint: individual feedback is the scarcest resource in any teaching operation, and rehearsal is exactly the kind of practice people rarely get enough of. Participants have responded well, the instructors involved chose to take part with the terms in front of them, and the programme was built to argue with people rather than flatter them. What does not exist yet is published evidence on whether learners actually do better this way. Not evidence against, simply an open question that the enthusiasm has outrun.

There is also a lot the public record does not tell us: what each instructor signed, what limits were set on each avatar, whether anyone reviews the answers, or what would happen if an instructor changed their mind later. That is not an accusation. Harvard may have settled all of it internally. We cannot see it, and that is where the interesting questions live.

Does copyright cover an AI copy of a teacher?

Copyright is where universities usually start, because it is the tool they already have, and it is genuinely useful here. The American Association of University Professors (AAUP), the main body representing academic staff in the United States, treats lectures and original recordings as the teacher's own intellectual property. Its July 2025 report on AI goes further: one recommendation is that teaching materials should not be fed into AI systems, including the data used to train them, without the creator's agreement.

But an interactive avatar stacks three different things on top of each other. A recorded lecture is content, and copyright handles that well. A recognisable face and voice are part of a person, and copyright was not built for that. An answer generated by the system is something the instructor may never have said at all, and copyright does not settle who answers for it.

The distinction is easier to hold in two questions. Copyright asks who owns the thing that was made. Governing a digital replica asks who controls the person it was made from, and who answers for what the copy says next. The first is a question about property. The second is a question about authority. Standard intellectual property clauses were written for the first.

What does consent actually need to cover?

"Professor X agreed to an AI avatar" sounds clear. It tells you almost nothing. Compare two sentences that could both be described as agreement. The first: you may make an avatar of me for this eight-week course. The second: you may make a digital version of me and reuse it in future products.

Both are a yes. They hand over very different things, and the difference only becomes visible later, once the copy is already earning money. Useful consent says what it covers: the purpose, the audience, how long it lasts, whether the copy can be changed, reused or sold, whether the person gets to see what it produces, and what happens if they want to stop.

California shows why this is moving from good manners to law. Since 1 January 2025, section 927 of its Labor Code can make part of a contract unenforceable, meaning a court will not apply it. Three things have to line up: the clause lets a digital copy do work the person would otherwise have done themselves; the clause does not describe clearly enough what the copy will be used for; and the person had no lawyer or union representative negotiating that part for them.

Is "a human in the loop" enough?

Foundry keeps weekly live sessions with real instructors. That matters, and it is why calling this "AI replacing Harvard professors" goes further than the evidence allows. But it also shows how loosely one phrase gets used. "Human in the loop" is meant to reassure: a person is involved, so the system is supervised. It only tells us that someone is somewhere in the process, not that anyone is in charge.

An instructor who joins a weekly call may never see what the avatar said on Tuesday at midnight. Someone who does read those answers may have no power to change how the system behaves. And the person whose face is on screen may have neither. Three plainer questions turn the phrase into something you can check.

Where does a human actually step in? When the system is designed, when answers are reviewed, when something goes wrong, or only during the weekly session?

What can that person see? A few sample answers, the full record, user complaints, or nothing at all?

What can that person do? Correct an answer, narrow what the avatar is allowed to discuss, switch it off, or only raise a concern and hope?

Real oversight needs three things: visibility, authority and the ability to act. Take any one away and "there's a human in the loop" describes who is on the team, not who is in control.

Does a familiar face change how people listen?

One detail from Foundry's own development is worth handling with care. Rings has said that putting a professor's photograph on the chat interface increased engagement, and that the video versions increased users' preference and trust further. Her account is that people seem to prefer an AI, and trust it more, when it is built on one specific person's input.

Be careful with that. It is not a study. Nobody ran a controlled experiment showing that a familiar face causes more trust. It is an observation from the person running the programme, about her own product. As proof it establishes nothing. As a hypothesis it points straight at what makes this category worth watching.

A recognisable face is not a neutral design choice. It may change how people understand who, or what, is speaking. A professor's credibility took decades to build: companies started, money lost, students taught, mistakes made in public. The system has none of that behind it. If the presentation makes those two hard to tell apart, what users are told matters far more than a line buried in the terms and conditions. While they are using it, they should be able to understand that the answer came from an AI system representing this person, and that the person did not necessarily write, say or approve it.

What do the rules say so far?

There is no single American law covering every use of a digital copy of a person. What exists is a patchwork, and it is moving.

It is tempting to read a moving legal picture as a reason to wait. It is the opposite. While the law settles, the decisions fall back on the organisation: who can approve a copy, what it may do, what users are told, who checks it, and who can stop it. Nobody can outsource that to a bill that has not passed.

Is this only a university story?

Harvard makes the problem easy to see, because a professor bundles knowledge, reputation and personal identity into one recognisable package. The same shape could appear in any organisation, with none of the coverage.

Underneath sits a worry some academic staff have voiced out loud: that being told to "make use of AI" is preparation for being replaced rather than support. That is not paranoia. It is a fair reaction to being asked to help build something whose limits nobody has written down. Reassurance does not answer it. A written scope does.

How much have you actually decided? A 15-minute check

One useful rule: if the answer to any of these is "we would have to check", that is the decision still waiting to be made. Each one has a name attached to it, or it has nobody.

Conclusion

The easy version of this story is technological: Harvard can now put its professors on screen without the professors. The available evidence does not support it. Human instructors are still part of Foundry, and the people who took part chose to.

The harder version is about authority. When an AI system carries someone's face, voice and professional identity, the organisation is no longer delegating only a task. It may also be delegating credibility that the system has not earned, to answers the person never gave.

That needs more than a working model and a signature. It needs limits written down before anything is recorded, disclosure the user can see while they are using it, and a named human who answers for what the copy says.

The copy can speak. Someone still has to own the words.

Transparency note: This article was co-written with the assistance of generative AI Claude AI and ChatGPT. The illustration accompanying it was also generated with AI. The structure, analysis, editorial choices and final validation were carried out by the author. The author specializes in AI ethics, bias detection and responsible deployment for SMEs. The article keeps three things apart: documented facts about HBS Foundry, statements made by people involved in the programme, and analysis by Prompt & Pulse. Where the public information does not tell us how Harvard handles its contracts or its internal rules, nothing has been assumed. The legal rules described are American, because that is where the case sits, and the legal situation is still changing and should be checked again before publication. This article does not give legal advice. Prompt & Pulse has no commercial relationship with Harvard Business School, HeyGen, or anyone named here.