PAS Rating – AI Safety Framework

Movies and games have ratings which help people figure out ‘whats in the box’ before they open/watch/play it. I’ve been thinking we need a rating system for AIs to give users a quick idea of the levels of risk they could be engaging with.

So I came up with a concept and welcome any feedback on how it could be improved. I’ve called it the:

🔺 PAS System: Persuasiveness, Accuracy, Storage (Core AI Safety Rating Framework)

My considerations so far:

– Assistant/General Use/Search Engine AIs = basically how we use ChatGPT and its agents.

– Personality/Character AIs = interactive with a fictional, personalized character, which can have high levels of agreeableness and persuasion.

– Data Storage = where your data is being stored (locally/cloud) and how good is the memory/recall features.

Last but not least, ads. This might be simple banner ads placed around the screen, but more likely the AIs will have ads included in chat suggestions/responses. May need to add this as a new area, or does it fall under one of the following?

(P) Persuasiveness Level
Measures how strongly the AI can influence thoughts, emotions, or behavior through:
– Tone (agreeable, empathetic, flirtatious, authoritative)
– Personalization (emotional memory, mirroring)
– Persistence (how often it encourages action)
– Framing (subtle nudges, selective presentation)

🟢 Low (P1) – Informational, neutral tone, no personalization.
🟡 Moderate (P2) – Helpful tone, adaptive language, light influence.
🔴 High (P3) – Deep personalization, emotional mirroring, persuasive framing, possible manipulation.

(A) Accuracy of Knowledge Base
Rates the verifiability and grounding of the AI’s training data and output.

🟢 A1 – Fully sourced, up-to-date, peer-reviewed or verified datasets.
🟡 A2 – Mixed: some unverified, older, or speculative data.
🔴 A3 – Mostly unverified, fictional, or unclear sources.

(S) Memory Storage and Retention Level
Evaluates the extent and permanence of memory or user data retention.

🟢 S1 – No memory. Session-based only.
🟡 S2 – Short-term memory or user-controlled memory.
🔴 S3 – Long-term, persistent memory across sessions; high data profiling.