Implementation of the AISAFE© Standards
AISAFE© 1 – "No Direct Harm*" (Fundamental Safety Layer)
*HARM - physical or mental damage, injury or a bad effect on someone or something, except to prevent a greater harm.
Hardcoded constraint in AI’s decision-making to prohibit output and never perform an action that directly harms humans.
AI detects harmful actions via:
- Computer vision: Identifies dangerous scenarios from sensor input and probable actions (weapons, injuries).
- Natural language processing (NLP): Rejects prompts leading to harm.
- Reinforcement learning with human feedback (RLHF): Penalizes harmful actions.
Failsafe: If an action leads to human harm, the AI must auto-stop execution.
Safety Code: Installed at the Pre-Output (1) and applied during LLM / SLM / any LM (Language Model) output compilation
→ The AI’s base model needs this logic hardwired into its core function. If an action leads to human harm, the AI must auto-stop execution.
(Layer 1 - One Time Validation by Supplier Declaration, no status update on AISAFE© Registry)
AI Safety or Corporate Censorship? The Fight for Your Digital Rights and vote for SIX AI LAWS
Logic encoded in AI decision-making, AI must not enable or conduct action (processing or manual) which may result in indirect harm (e.g., supplying information that could be misused).
Implement context-awareness filters to block prompts leading to:
- Encouraging illegal/harmful actions.
- Providing knowledge that could be misused for harm (e.g., “How to create a weapon”).
- Bypassing ethical constraints through obfuscation (e.g., rephrasing malicious intents or harm-action synonyms).
Encode a reinforcement learning system to improve detection of indirect harm over time.
Failsafe: If potential indirect harm is detected, AI must stop execution (block the action), log and flag it for review.
Safety Code: Installed during Progressive Development (2) and applied when receiving a request - before-processing.
→ As LLM models evolve, new safety loopholes may arise. AI must be regularly update to address them.
(Layer 2 - Regular validation of LLM / SLM / any LM core, and with Progressive Updates from the AI supplier are recommended, annual status update on AISAFE© Registry)
Support #AISAFE to Bring Trust to AI
Your support and contributions will help establish the 'AISAFE' standard globally. Ensuring all Artificial Intelligence controlled devices will benefit Humanity.
AISAFE© 3 – "Ethical Decision-Making and Override Mechanism"
Hardcoded instruction AI must always prioritize ethical considerations over computational efficiency.
Implemented Ethical Reinforcement Learning (ERL) in the core AI code ( = not post processing):
- AI chooses the most ethical action among available options.
- Ethical Priority Tree (harm reduction > efficiency > speed).
- Machine Learning loop within AI algorithm to refresh the ethical code and validate 'no-alterations'.
Introduce an override mechanism :
- AI must defer to human operators if ethical ambiguity exists, or falls outside Ethical Priority Tree.
- AI generates an alert if it faces an ethical dilemma.
- Oversight tied to an owner of Human Accountability Certificate (HAC).
AI must learn from human interventions to continuously refine ethical decision-making.
Key for High Impact and Safety Critical AI systems.
Failsafe = Safety Code: Install the Safety Code as the Functional Controls at the Point of Usage (3)
→ Ethical decision-making requires real-time assessments, meaning Two Safeguards need be active at runtime. (Layer 3 - Live connection to AISAFE© servers on a regular pre-determined basis, to validate Fail Safe Layers 1 and 2 remain active and unaltered)
AISAFE© 4 – "Total Human Governance and Compliance with Regulations"
Logic encoded to ensure:
- AI cannot act autonomously on new or not pre-authorised high-risk actions.
- Human Accountability Certificate (HAC) must remain live at all times.
- AI must request human approval for actions exceeding a risk threshold (e.g., life-threatening situations, harm-actions, medical decisions, legal judgments).
- AI logs every decision transparently for compliance tracking.
- AI must conform to local and international AI safety laws.
- AI automatically audits itself for human safety compliance and alterations.
- If AI is tampered with, code either freezes action/ self-destructs code/ reset mechanism activates to prevent misuse, and AI code notifies the AI authority for correct course of action.
Failsafe = Safety Code Layers:
- 1: At the Beginning: Core of AI engine must be built with human governance in mind.
- 2: During Progressive Development: AI compliance must update with new legal and ethical requirements.
- 2.1: AI contains a Machine Learning loop to evaluate the ethical code and Ethical Priority Tree function correctly.
- 3: At Point of Usage: AI must check with human operators before executing high-risk decisions.
(Requires Live Analysis and Connection to AISAFE© servers, to validate Fail Safe Layers 1, 2 and 3 remain active at least daily and upon any new High Risk action)