1. Our AI Architecture
SafetyMeter uses a two-layer architecture designed to ensure that AI is never a single point of failure in any assessment:
Layer 1 — Deterministic Scoring Engine
All risk scores, harm ratings, compliance flags, and priority classifications are produced by rule-based scoring engines. These engines map your product profile to known risk patterns, regulatory criteria, and framework requirements using documented, auditable logic. The same inputs always produce the same structured output. No AI is involved in this layer.
Layer 2 — AI Narrative Enrichment
After the scoring engine produces structured results, a large language model (LLM) reads those results and writes plain-language narrative: explanations of why risks matter, scenario descriptions, tailored recommendations, and debrief coaching. The AI reads scores — it does not create them.
This separation means that even if AI narrative generation is unavailable or produces suboptimal output, the underlying assessment data remains valid and reproducible.
2. AI Provider
SafetyMeter uses Anthropic's Claude model family for all AI narrative generation. Anthropic is a safety-focused AI company. Their API usage is governed by Anthropic's Privacy Policy and Acceptable Use Policy.
3. What Data Is Sent to the AI
When AI narrative generation is triggered, we send Anthropic's API a structured prompt containing:
- Your product name and organisation name
- A summary of your product's features and characteristics (as you described them in the form)
- The structured scores, risk categories, and priorities produced by the deterministic engine
- The policy type and tone preference (for the Policy Generator)
- The scenario title and decision summary (for the Incident Simulator debrief)
We do not send to Anthropic:
- Personal data about your end users, customers, or employees
- Sensitive personal data of any kind
- Financial data, health records, or other regulated categories of data
- Any data that could directly identify a natural person (other than the organisation name you provide)
4. AI Training and Data Retention
Anthropic's API usage policy governs how submitted data is handled on their side. As of the date of this policy, Anthropic does not use API inputs and outputs to train their models by default. We recommend reviewing Anthropic's current privacy documentation for the most up-to-date information.
SafetyMeter does not store AI-generated outputs on our servers after they are delivered to your browser session.
5. AI Output Quality and Limitations
AI-generated content produced by SafetyMeter may:
- Contain factual inaccuracies or outdated regulatory references
- Reflect general patterns rather than jurisdiction-specific legal requirements
- Produce different narrative output for similar inputs across sessions (non-deterministic)
- Miss context that is not captured in the structured form inputs
We mitigate these limitations by:
- Grounding AI prompts in the deterministic scores, so narrative is always anchored to verified data
- Using structured output formats (JSON) to reduce hallucination risk in critical fields
- Providing template-based fallbacks when AI generation fails, so you always receive a valid result
- Including prominent disclaimers on all AI-generated content
6. Responsible AI Commitments
SafetyMeter is itself a responsible AI platform. We hold ourselves to the same standards we ask of our users:
Transparency
We are transparent about when and how AI is used in our platform. Every tool clearly distinguishes between deterministic outputs and AI-generated narrative. We do not obscure the AI's role.
Non-Deception
We do not use AI to create false impressions of authority, certainty, or certification. AI outputs are clearly labelled as AI-generated and are always accompanied by appropriate disclaimers.
Fairness
We regularly review AI-generated outputs for systematic bias, particularly in risk assessments, harm modeling, and recommendations. We update our prompts and templates when bias is identified.
Human Oversight
AI is a tool in our platform, not a decision-maker. All AI-generated assessments are designed to support human decision-making, not replace it. We encourage users to review outputs critically and consult professionals.
Safety by Design
Our AI prompts include explicit instructions to avoid harmful, discriminatory, or misleading content. Responses that violate our content policies are filtered before delivery.
7. Model Versions
We periodically update the AI models used in SafetyMeter to benefit from improved capabilities and safety improvements. When we update models, we test outputs to ensure quality and consistency are maintained. We will update this policy to reflect significant model changes.
8. Opting Out of AI Narrative
If AI narrative generation is unavailable (e.g., due to API outage), SafetyMeter automatically falls back to template-based outputs derived from deterministic scoring. The structured assessment data is always valid regardless of AI availability. Currently we do not offer a user-facing option to disable AI enrichment, but template fallbacks ensure the Platform is always functional.
9. Contact
For questions about our AI usage or to report concerns about AI-generated content, contact info@trustedtechafrica.com.