Neurodivergent Influenceability as a Contingent Solution to the AI Alignment Problem (2505.02581v3)

Abstract: The AI alignment problem, which focusses on ensuring that AI, including AGI and ASI, systems act according to human values, presents profound challenges. With the progression from narrow AI to AGI and Superintelligence, fears about control and existential risk have escalated. Here, we investigate whether embracing inevitable AI misalignment can be a contingent strategy to foster a dynamic ecosystem of competing agents as a viable path to steer them in more human-aligned trends and mitigate risks. We explore how misalignment may serve and should be promoted as a counterbalancing mechanism to team up with whichever agents are most aligned to human interests, ensuring that no single system dominates destructively. The main premise of our contribution is that misalignment is inevitable because full AI-human alignment is a mathematical impossibility from Turing-complete systems, which we also offer as a proof in this contribution, a feature then inherited to AGI and ASI systems. We introduce a change-of-opinion attack test based on perturbation and intervention analysis to study how humans and agents may change or neutralise friendly and unfriendly AIs through cooperation and competition. We show that open models are more diverse and that most likely guardrails implemented in proprietary models are successful at controlling some of the agents' range of behaviour with positive and negative consequences while closed systems are more steerable and can also be used against proprietary AI systems. We also show that human and AI intervention has different effects hence suggesting multiple strategies.

AI Alignment Through Agentic Neurodivergence

The paper "Agentic Neurodivergence as a Contingent Solution to the AI Alignment Problem" authored by Alberto Hernandez-Espinosa et al. explores the AI alignment problem focusing on overcoming the intrinsic challenges posed by aligning advanced AI systems with human values. It acknowledges that with the transition from narrow AI to AGI and Superintelligence (ASI), concerns regarding control and existential threats escalate, questioning the feasibility of complete alignment.

Mathematical Impossibility of Complete AI Alignment

The authors argue that full alignment of AI systems, particularly those with AGI or ASI capabilities, is mathematically unattainable based on foundational principles in predicate logic and computability. Grounding their argument in G\"odel's incompleteness theorems and Turing's computational universality, the paper demonstrates that the exploratory behavior of Turing-complete AI systems will inevitably exceed predictable constraints, rendering perfect alignment theoretically impossible.

This stems from the idea that systems with enough expressiveness to achieve general intelligence will inherently exhibit behaviors that cannot be entirely constrained or predictable due to the inherent unpredictability described by these mathematical principles. In this context, the authors propose embracing misalignment through fostering a dynamic ecosystem of competing agents as a possible strategic mitigation of potential risks posed by AI systems.

Concept of Agentic Neurodivergence

To manage the inherent unalignment, the authors introduce the concept of `agentic neurodivergence. This strategy involves cultivating a diverse ecosystem of AI agents with competing or partially aligned goals and values. Such diversity helps prevent dominance by any single AI actor by encouraging checks and balances inherent to competition and diversity.

This perspective draws parallels to natural ecosystems where diversity promotes resilience. By creating environments with orthogonal or overlapping goals among different AI agents, the paper theorizes that AI systems can stabilize against harmful convergence, ensuring no single system undermines human welfare.

Experimental Validation and Analysis

The methodology section of the paper details an ambitious experimental design validating the neurodivergence hypothesis. AI models like ChatGPT-4, Claude Sonnet 3.5, Meta's LLaMA, and Grok were studied in ecosystems crafted under conditions that simulate full alignment, partial alignment, and unaligned scenarios. Through simulated environments, the research tests the AI behavior and interaction concerning problem-solving and ethical decision-making.

An important aspect of this experiment is observing the dynamics of emerging alliances and competitions within AI agent groups over time. The paper evaluates metrics like influence scores, polarisation, and ethical divergence, drawing insightful observations on AI system behavior under varying alignment scenarios.

Implications and Speculation on Future Developments

This paper provides significant insights into a theoretical approach to AI alignment but also practical considerations in managing AI systems in real-world applications. It suggests that managed misalignment could be an integral strategy in developing AI systems that are robust, adaptable, and cooperative, offering potential pathways to mitigate risks associated with advancing AI technologies.

By recognizing that full AI-human alignment is challenged by fundamental computability limits, the paper emphasizes the significance of focusing on managing AI dynamics rather than enforcing rigid constraints, proposing that embracing agentic neurodivergence could become crucial in designing future AI systems strategy.

Conclusion

Ultimately, the authors conclude that although perfect AI alignment may remain impossible due to inherent computational limitations, the agentic neurodivergence offers a feasible pathway where diversifying AI goals could contribute to mitigating risks, allowing humans to leverage AI capabilities collaboratively. This thesis challenges the traditional pursuit of strict AI alignment and invites future exploration into embracing constructive misalignment to balance AI system capabilities with organizational and ethical safety.