What Does TreadEd Actually Do?

Generative AI is not a silver bullet for education – it's potential. Three problems stand between that potential and genuinely better learning. And here's how TreadEd tackles them.

Nothing helps me learn better than Claude. I hardly read any book “solo” anymore, and the same goes for blog posts or academic papers – though the foundation for real understanding is, of course, already laid with a NotebookLM podcast.

So I would never dispute that AI assistants can be enough to learn better than ever. At the same time, I also think it's a major challenge to put this better learning into practice systematically. Or, as Claude would put it: generative AI is not a silver bullet for education – it's potential.

But what exactly is the bridge we need to build to unlock the potential of generative AI for education? And how do we try to address it at TreadEd?

I see three problems that could prevent generative AI from becoming the breakthrough for education it could be:

An assistant wants to take resistance away from you; a tutor has to give you resistance.

The simplest form of this: learners who aren't necessarily motivated to understand content can now remove the resistance of a task more easily than ever. My favorite example was a student recently in a clip from the German satire show Heute-Show who, when asked whether he used ChatGPT for homework, said with almost disarming honesty that his time was too valuable for anything else.

More surprising, however, is that even learners who don't necessarily want to cheat learn worse with AI assistants.

So the problem isn't (only) that it's easier to cheat. The problem is more that it's easier to get to answers – whether you read them and try to understand them, or simply copy them blindly.

And reading an answer can feel so much like your own understanding. We all know it: we listen to a podcast and genuinely stay on the ball. Only afterwards, when we want to tell a friend what it was about, do we notice that it felt more like understanding than it actually was. Herbert Simon, father of cognitive science and (economics) Nobel laureate, articulated the problem decades ago.

It has become so wonderfully easy to get a very good answer to almost any question … unfortunately, that's not enough to learn – even if you read it. In other words: the activity, the effortful part, is (unfortunately) the point. An analogy to fitness goes a long way.

Just as muscles have to burn to grow, the mind has to smoke to fill up. And just as an e-bike makes you less fit because it bypasses the burn, an AI assistant fills the mind less because it bypasses the smoke.

Because in my experience, too, it often isn't the motivation to learn that fails, but the fact that effective learning is more complicated than one might think. How do I actually know whether I've understood something? And that is exactly what a good tutor makes accessible: they don't give the right answers, they ask the right questions.

We are, of course, not the first and only ones to recognize this. Solutions like Khanmigo from the (inspiring) Khan Academy also rely on polarizing technological potential toward resistance.

What is new and unique about TreadEd, however: we don't spread this potential of generative AI through our own learning content. Instead, we let others – content experts who want to use AI for their educational offering – prepare interactions with it.

But for content experts to put their name behind it, two decisive ingredients are needed that few AI products deliver: reliability and transparency.

It could work! But also not … and above all not always.

The big problem is that the whole thing is an untenable combination of unreliable and opaque. The larger the training assignment given to the model, the more likely the model loses sight of the goal. And how am I supposed to know what the learner really accomplished and therefore learned?

If education creators are paid to take responsibility for learning progress – how are they supposed to stand behind such a black box?

At the same time, you only really use the potential of artificial intelligence if you let it make intelligent judgment calls. No one would be helped if AI now reliably quizzed multiple-choice questions with a clear answer key. There is more potential in evaluating learners dynamically within the interaction – that is, automating validation (e.g. “expectation horizons”) rather than just specification (e.g. “answer keys”), as AI celebrity Andrej Karpathy put it last year.

But how do you reconcile such judgment calls with reliability?

The answer is a clear focus on content and pedagogy, framed by a dynamically adjusted agenda and division of roles.

And what sounds so complicated is actually quite easy to grasp from a human perspective.

The first time you teach, you still have to build the analogies for the first time. A lot of effort, a lot that can go wrong, and many reactions from learners that are even harder to respond to under that load. “Why wasn't the parallel between ChatGPT and a linear equation understood?”, “Which insight has to be secured beforehand?” – those are experiences that come to mind. You learn working sequences of insights and become more flexible at bringing them about.

But over time, these points solidify. I become more reliable because I know which story I want to take learners through, with which insights and which possible missteps.

Put simply, this is exactly what happens in an interaction at TreadEd.

The interacting model has a delineated framework of content and also a pedagogical agenda in mind – but always only as much as it needs for the current part of the interaction. It has no binding protocol, but flexible response intuitions that a) give the teacher a degree of control over the interaction, with the option to contribute their own expertise as added value, and b) support learners flexibly as needed.

The interacting model does know which larger interaction framework it is in, but it is focused on the current sub-goal of the interaction. It can concentrate on this role of optimal challenging because the evaluation and the resulting shift in focus (the dynamic development of the system prompt) are handled by a complementary model.

Both models have a challenging and important, but clearly bounded focus. Anyone who wants to establish reliability in AI-based processes has to economize the models' attention: What matters right now? How can attention be sensibly divided?

Sometimes I get the feeling that surprisingly much – both in generative AI and in learning – can be explained by cognitive load management. But that's perhaps a post for another day.

Generative AI shines in many situations because you can get something impressive off the ground from scratch very quickly. But “starting from scratch” would also mean that everything already in place has to be left behind – or that our new AI interaction would have to be shared in parallel with existing e-learning.

As an education creator: a separate product. As a company: additions outside the LMS.

We can do better.

TreadEd is therefore LTI-ready and can be embedded in SCORM content. Or more simply: built to meaningfully augment existing content in an LMS – with what generative AI in education should be.

Under these conditions, generative AI can certainly help cover ground. It just takes people who make sure it goes in the right direction.

About the author

Dr. Christoph Kenntemich is Co-Founder of TreadEd.

TreadEd is a creator tool that makes it possible to design dialogue-based learning paths in which learners are always optimally supported through one-on-one guidance by an AI tutor. Experienced employees fully shape the tutor didactically, passing on their knowledge and experience at scale.