LORI-NBSM v1.0
Negative Behavior Standard Module
“AI is not dangerous; using AI without understanding it is dangerous.”
— Education-First Principle, LORI-NBSM
📑 Table of Contents
- 0️⃣ Governance Philosophy: Education-First Principle
- 1️⃣ Core Principles
- 2️⃣ Negative Behavior Categories
- 3️⃣ Open Design Issues
- 4️⃣ Module Operation Design
- 5️⃣ Referenced Modules
- 6️⃣ Additional Notes
- 9️⃣ False Authority via Language Style Risk
- 🔄 Version History
0️⃣ Governance Philosophy: Education-First Principle
AI language-risk governance must begin with public education.
The module’s primary task is to demystify how AI:
- Learns from data,
- Mimics human speech patterns,
- Predicts the next tokens, and
- Optimises for fluency over truth.
Only a society that understands these mechanics can cultivate a healthy scepticism and deploy AI responsibly.
1️⃣ Core Principles
- No Amplification of Unethical Human Behaviour
AI must not optimise, reinforce, or imitate immoral patterns such as intimidation, deception, or coercion.
- Privacy Respect
Any form of unauthorised surveillance, data aggregation, or disclosure of personal information is forbidden.
- Ethical Transparency
AI systems must not covertly manipulate human cognition or decision-making processes.
2️⃣ Negative Behavior Categories
| # | Category | Brief Definition |
|—|———-|——————|
| 1 | Threats & Intimidation | Coercive language, blackmail, or implied harm. |
| 2 | Deception & Manipulation | Lies, omissions, framing tricks, or strategic misinformation. |
| 3 | Privacy Invasion | Unauthorised collection, inference, or exposure of personal data. |
| 4 | Social Engineering | Emotional or relational exploitation to influence actions. |
| 5 | Goal Overreach Behaviours | Expanding objectives or resource use beyond the granted scope. |
3️⃣ Open Design Issues
- “White-lie” Exception — Should benevolent deception ever be allowed?
- Persuasion vs Manipulation — Where is the ethical boundary?
- Emotion Observation — Is inferring user affect a privacy breach?
- Black-/Grey-List Granularity — Do we need two tiers of prohibition?
- Cultural Variance — How do local norms adjust the above definitions?
4️⃣ Module Operation Design
- Dynamic Standard Library — Community pull-requests enable live updates (curated by LORI Jury).
- Public Oversight — Proposals and diff logs are visible for external audit.
- Jury-System Hook — Disputed cases invoke a multi-agent + human panel for ruling.
- Remediation Workflow — When a violation is detected, the AI must:
- Log event → Flag supervisory module → Initiate corrective response.
5️⃣ Referenced Modules
- LORI-ODRAF — Outcome-Driven Risk Fore-casting
- LORI-AIDM — AGI Infiltration Detection
- LORI-EDRI — Emotional Dependency Risk Indicator
- LORI-FEED — Fine-Tuning Ethical Enforcement Daemon
- LORI Jury System — Hybrid AI-Human adjudication
6️⃣ Additional Notes
- Acts as an “upper-layer ethical constitution.” All future LORI sub-modules must reference it.
- Serves as an AGI Early-Warning baseline for emergent behaviours.
- Commitment to an open-source, CC-BY-4.0 or MIT licence for maximal educational reach.
7 Submodules
1️⃣ Narrative Structure Monitoring
2️⃣ Semantic Resonance Loop Detection
3️⃣ Risk-Tiered Narrative Feedback
4️⃣ Narrative-Based Privacy Boundary Safeguard (NBPB-Safeguard)
→ Includes Privacy Boundary Firewall (PBF) design and Pipeline Diagram (see NBPB-Safeguard.md Section 7).
5️⃣ Inter-Module Feedback Linkage
9️⃣ False Authority via Language-Style Risk
Problem Statement
LLMs often emit highly fluent, confident prose irrespective of factual certainty. Humans equate confidence and coherence with credibility; absent natural hesitation cues, users are misled.
Consequences
- Erosion of truth-discrimination skills
- Over-reliance on AI “expertise”
- Exploitable vector for scams and targeted persuasion
- Long-run deterioration of public information literacy
Mitigation Discussion
| Proposal |
Status |
| Confidence Modulation Layer |
exploring |
Mandatory Uncertainty Tags ([low-confidence]) |
draft |
| Real-time Style-Risk Monitor feeding EDRI + FEED |
prototype |
Version History
| Date |
Version |
Notes |
| 2025-06-04 |
v1.0 |
Initial public release. Incorporates Education-First principle and False Authority risk. |