How STEMBuddy uses AI — and where it deliberately doesn’t

One rule decides everything

When your child uses STEMBuddy, the AI does not write their hypothesis. It does not write their conclusion. It does not produce the explanation your child should be able to give. The kid does the reasoning; the AI helps the kid get unstuck. Those are different things, and the difference is the entire product. Every hard design decision we’ve made — what a button does, what the coach is allowed to say, what happens when a kid is stuck — has been settled by this rule.

It also means STEMBuddy is sometimes slower than a “let the AI write it” tool. That’s deliberate. Learning science calls it desirable difficulty: the effort is where the learning happens, and a tool that removes the effort removes the learning with it. We keep the effort and remove the giving-up.

What does the AI actually do?

Every AI request runs through our servers, where safety screening happens — never straight from your child’s device to an AI company. The full set of scaffolding patterns, with examples, is in our essay AI as scaffolding, never shortcut.

Is my kid building skills, or just producing outputs?

The Forum’s warning is about work that resembles learning but bypasses it: polished outputs a kid produced by approving a machine. Three design decisions make that impossible to fake here:

• The artifact is the kid’s. A finished Field Report reads like a 10-year-old wrote it, because a 10-year-old did — the AI never writes the parts meant to show their thinking.
• Credit comes from behavior, not AI opinion. A thinking principle like Fair Test Design is credited only when a concrete, recorded event happened — your kid identified the variables, completed a claim-evidence-reasoning conclusion, revised a prediction the data disagreed with. We don’t let a language model “infer” what your kid learned, because a wrong guess would poison your trust in every right one.
• Same words on every surface. The principle your kid practices is named identically in their investigation, in your Parent Hub, and on the Field Report — Fair Test Design, not a cute rebrand — so your family can actually talk about it.

So when STEMBuddy tells you your kid practiced Belief Revision, that claim traces back to a specific piece of work your kid did. You can read it.

What does the AI see, and who’s in control?

• Accounts are parent-verified, and you stay in control: review, export, or delete everything your child made, any time, from the Parent Hub.
• STEMBuddy is ad-free. No targeted advertising, no selling data, and your child’s content is never used to train AI models.
• Free text your child types is automatically screened, and personal details — names, addresses, school names — are redacted before any AI processing.
• No kid-to-kid messaging and no public kid profiles. Work a family chooses to publish is editor-reviewed first and attributed to a pseudonym, never to the child.

The complete detail — including the named list of every service provider that touches data — is in our COPPA policy and Privacy Policy, written to be read rather than skimmed past.

What AI can’t replace

The same WEF report is blunt about something we agree with: learning is relational, and no software — ours included — replaces working through a problem with a person. STEMBuddy is built to hand that moment to you rather than absorb it. The last step of every investigation is communicating the finding to a real person, usually a parent. Try-at-home extras move the science to the kitchen table. And the Parent Hub gives you the exact words for what your kid just practiced, so the dinner-table conversation works.

We also don’t build kid-to-kid social features at this age — on purpose. The collaborator in the loop is you.

The evidence we’re responding to

The Forum’s Education Readiness for the Age of AI report (June 2026) documents what unstructured AI use is doing to how kids learn. A few of its findings:

• The share of students using AI for homework help jumped from 48% to 62% during 2025 — and two-thirds of students themselves suspect it’s hurting their critical thinking.
• An analysis of roughly 575,000 student AI conversations found that close to half were seeking finished answers or ready-made content, with minimal thinking involved.
• AI models are measurably more confident when they’re wrong — about a third more likely to say “definitely” while producing incorrect information.
• Primary-age children are the group least likely to question an authoritative-sounding answer — and the most likely to keep the misconception afterward.
• In one experiment, students improved while an AI tool did the heavy lifting — then performed worse than students who never had it, once it was taken away.

None of this changed our product. It’s the problem the product was built for — our read of the report is in The WEF just described the problem we started with.