Skip to main content
Transparency

How AI is Used (and Not Used)

Complete transparency about AI in SayVeritas: when it runs, what it outputs, and its limitations.

When AI Runs

AI is used in the following specific contexts within SayVeritas:

🎙️

Speech-to-Text Transcription

Student voice recordings are transcribed to text using OpenAI speech recognition. This creates a reviewable record of what students said.

📊

Rubric-Aligned Scoring (Dual-Model Consensus)

Transcripts are analysed by two AI models (OpenAI and Google Gemini) for consensus scoring against teacher-defined rubric criteria. This dual-model approach reduces single-model bias. Scores are recommendations only.

🖼️

Assessment Image Generation

AI can generate images for assessment prompts (e.g., diagrams, scenes for analysis). Teachers preview and approve all generated images before student use.

🔍

Socratic Follow-Up Questions

In StudyLab mode, AI generates follow-up questions to push student thinking deeper. Designed to guide, not evaluate.

📈

Class-Level Insights

AI aggregates responses to surface common misconceptions and patterns. All insights are evidence-linked to source transcripts.

What AI Does NOT Do

  • No autonomous final grades — AI never assigns grades that go directly to students without teacher review
  • No certification or credentialing — SayVeritas does not issue certificates, diplomas, or formal credentials
  • No high-stakes proctoring — We do not claim exam-level security or identity verification
  • No emotion or sentiment analysis — We deliberately exclude tone, mood, or emotional inference
  • No training on your data — Student recordings are never used to train public AI models
  • No advertising — Student data is never used for marketing or ad targeting

Known Limitations

We believe in honest disclosure. AI in SayVeritas has the following limitations:

Speech Recognition Errors

Transcription accuracy varies. Uncommon names, technical vocabulary, and fast speech may be misrecognised. Teachers can view original audio.

Accent & Dialect Variation

Speech models perform differently across accents. We provide teacher override to correct for any systematic bias.

Audio Quality Dependency

Background noise, poor microphones, or connectivity issues affect transcription quality. We flag low-confidence transcripts for review.

Rubric Interpretation

AI scoring is probabilistic. Edge cases and nuanced responses may not align with teacher expectations. Human review is essential.

Teacher Review & Override

Every AI output is reviewable and overrideable.

  • Teachers can listen to original audio alongside transcripts
  • Any AI-suggested score can be overridden with a documented reason
  • Feedback can be edited or replaced entirely before release
  • Override reasons are logged for accountability and audit

Our Data Promise

No Training on Student Data

Student voice recordings and transcripts are never used to train AI models. Your data serves only your educational purposes.

No Third-Party Sharing

Student data is not sold, shared, or commercialised. Our subprocessors access data only to deliver the service.