AI Transparency
AI Transparency
AI transparency at Ed.ai
Four commitments, honestly held: no training on student work, mandatory teacher oversight, de-identification before every model call, and readiness for the state AI laws now coming into force.
How our AI works
The four products
Math Grading & Feedback — Follows step-by-step reasoning on handwritten and typed math work, awards partial credit, generates feedback
Standards-Aligned Analytics — Maps student errors to CCSS / TEKS / NCSCOS standards; detects misconception patterns
Targeted Practice Generation — Generates personalized practice based on identified misconceptions
AI Teaching Assistant — Conversational interface connected to teacher's class data
Across these products, Ed.ai automatically selects the best-performing model for each task from a set of US-hosted enterprise models — Azure OpenAI, Azure Mistral, and Google Gemini. There is no fixed model-to-product mapping. All inference happens on US-hosted infrastructure under no-retention, no-training terms. See the sub-processors page for the full list and contractual posture.
The de-identification step
Before any student work reaches a model:
Identifying zones (names, roster numbers, class labels) are detected.
Those zones are masked with opaque white pixels — irreversibly.
The masked image is transcribed via OCR.
The transcription is sent to the model.
Full pipeline and validation metrics on the de-identification page.
Training data
Our 100,000+ math problems used to develop and evaluate Ed.ai come from curated, licensed sources — textbooks, released exam banks, partner content, and mathematics databases under proper license.
Our four AI commitments
No training on student work
Neither Ed.ai nor any of our model vendors (Azure OpenAI, Azure Mistral, and Google Gemini) train on student work processed through Ed.ai. This is:
Embedded in our contracts with each vendor (no-retention + no-training clauses)
Published on this page and on the Trust Pledge
Verifiable on request — district DPOs can receive contract excerpts under NDA
Human-in-the-loop — the teacher always has the final say
No final grade leaves the system without teacher validation.
The AI proposes: a grade, a feedback sentence, a correction, a next exercise. The teacher confirms, edits, or rejects.
This is not an opt-in — it's a structural property of the product.
No grade and no graded work reaches a student until a teacher has reviewed and released it.
De-identification before any LLM call
We don't rely on vendor promises alone — we strip identity from student work before it is sent for inference. See §2.2 above and the de-identification page.
No autonomous high-stakes decisions
The AI does not:
Assign a final grade without teacher approval
Place a student in a remediation track without teacher choice
Flag a student to administrators without teacher awareness
Determine who receives a specific intervention outside of teacher discretion
Where the product surfaces recommendations, the teacher sees them clearly labelled as AI-proposed and chooses whether to act on them.
Regulatory alignment
NIST AI Risk Management Framework
We map our AI practices to the NIST AI RMF categories:
Govern — Designated AI owner; internal policy for model selection, evaluation, and deployment
Map — Each product use case documented: purpose, data inputs, model, evaluation method, risk assessment
Measure — Accuracy benchmarks (99% on handwritten math) · bias testing on math content · hallucination rate tracking
Manage — Model changes go through internal review; high-risk changes trigger extended testing; material AI behavior changes are reflected in release notes
Emerging state AI laws
Several states have enacted or proposed legislation affecting ed-tech AI vendors — including Colorado, California, Texas, Illinois, and Utah. We monitor these developments and design our AI to align with their core requirements: human-in-the-loop oversight, no autonomous high-stakes decisions, documented and licensed training data, de-identification before inference, and clear labelling of AI-generated content. As specific obligations take effect, we adapt our practices accordingly.
COPPA
Ed.ai does not draw inferences from children's information to build profiles.
Student data is retained only as long as necessary, per our retention schedule (the Privacy Policy §8).
FERPA
AI processing does not change our FERPA posture: Ed.ai remains a "school official with legitimate educational interest." AI analysis of student work is a service performed at the direction of the school, not an independent use.
FAQ
Is student work used to train AI models?
No. Not by us, not by our model vendors. This is contractual and verifiable.
Who sees my students' data?
Only the teacher, the school/district admins they authorize, and our infrastructure (encrypted, access-audited). The AI models receive de-identified transcriptions — they never see names or roster identifiers. Our model vendors process this data under enterprise no-retention and no-training terms.
Can I opt out of AI?
Ed.ai is an AI-based product — there is no non-AI version of it, and there is no per-teacher or per-student AI opt-out. A school that does not want a particular student to use Ed.ai handles this at the access level: it can choose not to use Ed.ai for specific students or classes. Contact your school to arrange an exclusion.
What models power Ed.ai?
Azure OpenAI (GPT family) — Azure US
Azure Mistral — Azure US
Google Gemini — Google Cloud US
The choice of model for each task is made automatically based on performance; we can provide the current mapping to district DPOs on request.
How do you prevent AI hallucinations in grading?
Three layers:
Structured prompts with explicit grading rubrics tied to standards (CCSS/TEKS/NCSCOS)
Confidence scoring — low-confidence outputs are surfaced to the teacher with a flag
Teacher validation — every grade passes through a teacher (see §3.2)
We also track hallucination rate as an internal metric and publish aggregate accuracy (99% on handwritten math).
Will AI ever give a final grade without my approval?
No. That's a structural commitment built into how the product works.
How often do the models change?
Model versions are updated as vendors release improvements. Material changes to AI behavior are reflected in our product release notes.
know
The AI itself holds no memory. Context for the AI Teaching Assistant is retrieved from the school's own account data at query time, de-identified where possible, and not retained by the model.
Request more detail
District DPOs and researchers can request:
Our internal AI governance policy (redacted)
Evaluation methodology and bias testing protocol
Sub-processor contract excerpts covering no-training clauses (under NDA)
Write to privacy@ed.ai.