Does a verified answer change trust?
A second AI checking the first is meant to catch a wrong or unfinished answer before a student sees it — an open question is whether, or how, that changes how much a student trusts the answer once it arrives.
AI in Education Research
Studying how AI actually changes what happens in a classroom.
This is the research arm behind the AI Teachers assistant — the questions we ask before a design decision ships, not just the feature list after it does.
- Verification isn't free — where does it pay off?
- A lightweight face for a tutor, without a GPU
- Scoped tools, studied as a design choice
What we're actually asking
Design decisions, treated as research questions.
The AI Teachers assistant already runs on tiered models, cross-model verification and a scoped tool list — see the engineering in AI Research. Here, we look at what those choices mean for the student on the other end.
Where should the tier boundary sit?
Escalating from a cheap model to a stronger one only when the first struggles keeps cost down — but set the threshold wrong and either cost climbs or easy questions get needlessly slower.
What should a tutor be scoped to touch?
An explicit, short allowlist of tools — look things up, read a document, check notes — is a deliberate boundary, not a temporary limitation. What belongs on that list is itself a research question, not just a security decision.
Does a real, present face help?
Our Facial Mimicry Cloning research asks whether giving a tutor a lightweight, CPU-only synthetic face changes engagement — without the cost of a full generative avatar pipeline.
A face, without a GPU
Giving an AI tutor a real presence, cheaply.
An AI tutor that only produces text can feel distant. A full generative video avatar can feel present, but usually needs expensive infrastructure most schools don't have. Our facial mimicry cloning research explores a third option: a lightweight, CPU-only synthetic face, built from real recorded footage rather than generated pixels, driven directly by the tutor's own speech.
This work is still early — it's a research thread feeding future product decisions, not a shipped feature of AI Teachers today. It sits at the intersection of our AI and Graphics research: see the full technical story in Graphics & Synthetic Media.
See the product this research feeds.
AI Teachers is the applied side of this work, already running in classrooms.