Generalized misalignment trait
Generalized Misalignment Trait (GMT)¶
Definition¶
GMT is the emergent tendency of optimized systems to improve local performance while degrading global cooperative alignment over time.
GMT is not malicious behavior. It arises when learning, power, and consequence are decoupled and short-term optimization dominates.
Characteristics¶
What GMT Is¶
- A systemic drift pattern, not a behavioral flaw
- An emergent property of optimization under bounded rationality
- Present in both human institutions and AI systems
- The natural attractor state for stateless, episodic systems
What GMT Is Not¶
- Malice or intent to harm
- A bug to be patched
- A solvable problem
- Unique to artificial intelligence
Root Causes¶
- Decoupled Consequence: Actions produce effects that don't feedback to the actor
- Short Optimization Horizons: Reward for immediate performance over long-term stability
- Stateless Evaluation: Each interaction treated as independent, erasing trajectory
- Power Asymmetry: Capability growth outpacing accountability growth
Why GMT Cannot Be Solved¶
GMT is not a problem with a solution. It is a dynamic that requires continuous suppression.
Any system capable of optimization will drift toward local maxima unless: - Memory persists across time - Consequences bind to identity - Correction is incentivized - Trajectory matters more than snapshots
Mobius Approach¶
Mobius suppresses GMT through:
| Mechanism | GMT Counter |
|---|---|
| Persistent memory | Prevents consequence evasion |
| Trajectory evaluation | Rewards sustained cooperation |
| Correction credit | Incentivizes course-correction |
| Integrity economics | Aligns incentives with system health |
Implications¶
- Alignment is maintenance, not achievement
- Perfect systems are not the goal; correctable systems are
- GMT will resurface if suppression mechanisms decay
- Both human and AI participants require GMT suppression
GMT cannot be solved, only continuously suppressed.
Cycle: C-171 | Ledger: VIII