AI Ethics & Safety Standards
How we build responsible, accurate, and unbiased Artificial Intelligence for the classroom.
Building AI for education carries a higher responsibility than other fields. A wrong answer in a creative writing prompt is an annoyance; a wrong answer in math is an educational failure. Here is how we ensure safety and accuracy.
Combating Hallucinations
Generative AI is prone to “making things up.” To solve this, MathGPT uses a Neuro-Symbolic Architecture.
- Step 1: The LLM parses the natural language.
- Step 2: A deterministic math engine calculates the result.
- Step 3: The AI compares its output against the engine before responding.
Student Data Privacy
You are not the product. We strictly adhere to student data protection principles.
- We do not use user inputs to train 3rd-party advertising models.
- We do not build shadow profiles of students.
- Chat history is stored locally or anonymized for quality assurance.
Bias Mitigation
Math is universal, but word problems can carry cultural bias. We rigorously audit our training datasets to ensure:
- Diverse representation in word problem names and scenarios.
- Culturally neutral context in financial and social examples.
- Accessibility-first output formatting (screen-reader friendly).
Human-in-the-Loop
Automation has limits. We maintain a review board of mathematics educators who:
- Regularly audit random samples of AI conversations.
- Manually correct recurring errors in the model.
- Set pedagogical guidelines for how the AI explains concepts.
System Safety Status
Our models undergo nightly automated red-teaming tests to ensure they refuse inappropriate non-math requests.