April 10, 2024

We introduced our first AI avatar professor. Here’s why.

hero image for blog post

Everyone in edtech is grappling with how to teach AI. At the rate AI is evolving, content is out of date very quickly, making it difficult to deliver fresh insights every time.

So far, we’ve approached this by putting a lot of manual time into running our live workshops monthly, updating decks every time, and prepping professors. But we’ve also asked ourselves a question we tell our students to ask:

What do we do today that can get more efficient with AI, and what are the things we don’t do today because we don’t have the tools for it?

Using AI avatars to teach courses could check both boxes. But there are obvious risks – mainly, avatars are simply not as good as humans right now. So here’s a peek at our calculus to figure out whether to pursue it.

The benefit to us: Cost + time

We have two sources of friction in our course creation process: Cost and time.

First, time. Currently, 80% of our instructors’ time is spent filming  – flying to the studio, sitting in the studio, recording, and re-recording. It’s about 8 hours total (out of a 10-12 hour commitment from them to create the course). Reducing this time means we (Section) can make more courses, satisfying student demand. It also means we can ask instructors to spend this time elsewhere (i.e. running the course more frequently).

Next, cost. Travel expenses, studio space, equipment, and tons of in-house editing adds up to about $15-20K per course.

In contrast, generating an avatar requires 30 minutes of filming, which can be done at a studio near the professor. Refining the avatar costs less than $2K per course. Reducing these costs means we can invest to build more courses (and therefore serve more student demand and increase margins).

The benefit to students: Relevance + accessibility

Reducing the burden of cost and time helps us do two things central to our mission:

One, refresh our course content constantly. Currently, we’re able to re-record with a professor every other year (if we’re lucky). With AI avatars, we can refresh course content 4 times a year by writing a new script and prompting the avatar with it.

Two, scale translations into many different languages. We’re also currently limited by the number of languages our professors can speak (in most cases, one). Our AI avatars can share  the same content in hundreds of different languages, making it much more globally accessible.

The risks: Lower quality + reliance on third parties

We refuse to be Pollyannas about AI, so here’s a clear-eyed look at the risks involved in AI avatars teaching courses.

1. Lower quality

Right now, avatar professors are not as good as human professors. We’re seeing that in student feedback: The NPS for the version of the course taught by an avatar is lower, and some students are turned off by the avatar’s speaking style or find it less engaging than a human.

AI could catch up to human delivery in the next 12 months – We used Synthesia to create our first avatar, and have already moved on to another vendor, HeyGen, whose avatar generation is much stronger.

But this could also remain an area where human delivery for the strategic content we teach will always feel more authentic. We won’t know for at least 12 months, as we and the world experiment with this evolving technology and get more used to these types of experiences.

2. Ethical concerns

We know people have concerns about AI avatars. As they become more realistic, people wonder: Will avatars be created without the source’s permission? Will avatars be introduced without making it clear they’re not human?

Today, we’re being very explicit with our users about when a video features a human vs. an avatar, and how our instructor created the course content and scripts that the avatar reads. But as these capabilities evolve, we’ll need to continue to evaluate how we handle concerns that students, instructors, or others might have.

3. Reliance on a third party

Outsourcing avatar generation to a third-party (like I mentioned, we’ve already tried two) creates risk to our model because we don’t control the technology or the roadmap. If this works, we’ll need to think about ways to mitigate this risk.

How we’re thinking about avatars right now

We’re using those risks to help inform our approach to using avatars in courses going forward:

Avatars won’t replace professors

Avatars are a delivery method – everything they share was created by the instructor they resemble. That’s why our instructors are irreplaceable. We not only need their expertise on the subject, but their ability to turn that knowledge into a framework that can be taught. So even when you don’t see them, you are still learning from a real person.

Plus, you’ll still interact with the human instructor during live lectures and office hour sessions. The point is to optimize human contributions, not do away with them.

We’re being as transparent as possible

We don’t want students to feel tricked, so we make it clear when an avatar is speaking and when a course includes an AI instructor. At the beginning of every avatar-led video, we have the avatar clearly introduce itself.

We aren’t waiting for perfect

Our first avatar was not the best quality (we’ve since found better technology and are building our next AI avatars with it) and wasn’t received well by 20% of our small cohort, and that was a valuable benchmark for us to get. We chose to optimize for speed over perfect quality so that we could get these kinds of learnings faster.

Our switch to HeyGen brings us much closer to the experience we want, but we probably won’t really know for another 12 months if our students like it as much as we’re hoping. The point is, we don’t want to wait for the tech to be polished to find out.

Our learnings so far

Using avatars as instructors will be an adjustment for students and involve some risk for us, but it’s the most viable solution for teaching a subject that’s evolving real-time.

And the benefit of AI’s rapid change of pace is that the quality of avatar generation tools is improving just as rapidly. If we were limited to the capabilities of the first tool we tried, Synthesia, this test would have told us not to use avatars at all. But the quality of HeyGen’s avatars makes it worth pursuing.

Keep an eye on the course calendar for upcoming avatar-instructed courses – and make sure to take the survey at the end of each course so we can learn along with you.

Greg Shove
Taylor Malmsheimer, Head of Strategy