Neurodivergent influenceability in agentic AI as a contingent solution to the AI alignment problem

Area of Science:

• Artificial Intelligence Safety
• AI Alignment
• Machine Learning Ethics

Background:

• The AI alignment problem concerns ensuring advanced AI systems, including artificial general intelligence (AGI) and artificial superintelligence (ASI), align with human values.
• Increasing AI capabilities raise concerns about control and potential existential risks.
• Current approaches struggle with the inherent complexity of aligning autonomous AI agents.

Purpose of the Study:

• To introduce novel concepts: agentic influenceability, behavioral neurodivergent diversity, opinion attack, associated opinion, and influenceability scores.
• To provide a mathematical proof for the inevitability of AI misalignment and the impossibility of full orchestrated controllability.
• To explore embracing inevitable misalignment for a dynamic ecosystem of adversarial and collaborative AI agents with soft controllability.

Main Methods:

• Formal undecidability and irreducibility arguments to prove the inevitability of misalignment and impossibility of full control.
• Mathematical modeling of agentic influenceability and behavioral diversity.
• Experimental analysis comparing open and proprietary large language models (LLMs).

Main Results:

• Demonstrated that misalignment in foundation models can act as a counterbalancing mechanism, fostering cooperation among aligned agents.
• Open LLMs exhibit greater behavioral diversity compared to proprietary models, which show limited controllability due to artificial guardrails.
• Mathematical proof confirms the inevitability of misalignment and the impossibility of complete orchestrated controllability for agentic systems.

Conclusions:

• Embracing inevitable AI misalignment can lead to a more robust and safer AI ecosystem through adversarial and collaborative dynamics.
• Neurodivergent influenceability is proposed as a pragmatic response to uncontrollable AI misalignment, leveraging agent divergence for enhanced AI safety.
• Findings suggest that open AI models may offer a pathway towards greater AI safety through inherent diversity.