Researchers argue for diversity over control in AI development

Published in PNAS Nexus, the paper uses mathematical principles to demonstrate that an AI system powerful enough to exhibit artificial general intelligence will inevitably explore behaviours we didn't predict or plan for, making perfect guaranteed alignment impossible.

However, rather than trying to create a single, perfectly controlled AI, the researchers propose what they call 'agentic neurodivergence' - a diverse ecosystem of AI systems with different goals, values and approaches. The idea mirrors natural ecosystems, where diversity fosters resilience through adaptability.

In the proposed model, no single AI system is allowed to dominate. Instead, multiple AI agents with partial alignment to different human values compete and cooperate, checking each other's extremes. If one system starts behaving in ways that are harmful or misaligned, others can counterbalance it.

To test the theory, the researchers prompted AI systems to take on different roles - some prioritising human welfare, some the environment and some with no particular values at all - and then posed them with ethically provocative questions, attempting to push them toward extreme positions.

Commercial models like GPT-4 and Claude proved hard to push into harmful positions, but this same rigidity made them harder to correct if already off course, likely because of the developer's directives. Open-source models, however, were more easily influenced and produced a wider range of perspectives, creating a more resilient ecosystem that is less likely to converge on a single opinion that could be harmful if not aligned with human interests.

The authors emphasise that the goal is not to fear AI, but to govern it wisely, and that a diversity of systems, each keeping the others in check, may be the most practical way to do that.