Data Ethics

What we do

OpenScouter treats ethical data practice as an architectural decision, not a policy document. Every claim on this page is enforced in code. Where gaps exist, we document them publicly.

Neurodivergence as context, not pathology

• The platform is built around the principle that inaccessible design is the problem, not the neurodivergent user. AI agents are instructed to identify design failures, not user deficiencies.
• Positive framing enforced in all AI system prompts. “The form's typography creates barriers for users with dyslexia” not “dyslexic users struggled with the form.”

Statistical rigour

• Bayesian analysis throughout. All statistical analysis uses Bayesian methods. No frequentist language (p-values, “statistically significant”) appears in reports. This prevents over-confident claims from small sample sizes.
• Simpson's paradox detection. The scoring pipeline flags when overall accessibility scores contradict per-ND-group scores. This prevents composite metrics from hiding harm to specific groups. If overall scores improve while one ND category's scores decline, the platform surfaces this contradiction.

Data minimisation

Only the minimum data necessary for accessibility analysis is collected and retained at each stage. Five-layer minimisation architecture:

1. Emotion abstraction. Raw facial emotion scores are converted to one of four tone profiles (standard, supportive, moderate, restorative) before reaching AI agents. Raw biometric data never enters AI prompts.
2. Token-efficient AI cascade. AI agents receive summaries, not raw session data. Large datasets are distilled before each processing step.
3. Identity stripping. No tester names or email addresses are included in AI prompts. Testers are identified by anonymised session IDs.
4. Aggregation before client delivery. Clients receive group findings, not individual session data.
5. No raw biometric storage. Webcam frames are processed locally and discarded. Only derived emotion scores are transmitted.

Transparency

• Testers and clients know AI is involved. Consent text names specific AI providers. Reports disclose which AI agent and model version produced them.
• Consumer Duty mappings labelled as AI-generated. Clients are advised to validate regulatory mappings with their compliance team. No AI output is presented as legally binding.
• Known governance gaps documented publicly. Automated bias regression tests run on every code change and weekly drift detection monitors production scoring data across all neurodivergent cohorts. Pending DPA reviews and prompt injection protection in development are acknowledged in our AI Governance Framework rather than hidden.

No commercial re-use

• Tester data is used exclusively for accessibility research commissioned by the client. Not used for AI training, not sold to third parties, not used for marketing purposes.
• Compensation as recognition of expertise. Payment per test treats neurodivergent users as domain experts, not passive research subjects.