Michael Darius Eastwood · Independent AI alignment researcher
Published
On 14 April 2026, PNAS Nexus published "Neurodivergent influenceability in agentic AI as a contingent solution to the AI alignment problem" by Alberto Hernandez-Espinosa, Felipe S. Abrahao, Olaf Witkowski, and Hector Zenil (volume 5, issue 4, article pgag076; DOI 10.1093/pnasnexus/pgag076; preprint arXiv:2505.02581). The paper is short, technical, and precise about its scope. It is worth reading it for what it argues rather than for what it is sometimes made to say.
The authors argue that Godel's incompleteness results and Turing's halting problem have direct consequences for AI alignment: any Turing-complete system cannot be fully predicted or externally controlled with certainty. Perfect uniform alignment through outside verification is therefore mathematically ruled out. The authors then argue that a promising, and on their account contingent, route around the impossibility is cognitive diversity. Their empirical section reports that model diversity has different effects depending on how it is produced: aligned open-source and community-derived variation was associated with beneficial diversity, while heavily fine-tuned or guardrailed variation was not. They frame this as an argument, not a theorem. The word "argue" is doing real work in the title of the paper.
The paper is not a proof that neurodivergent people are more accurate, more ethical, or better at any specific task. It is not a claim that any particular clinical group is superior. It uses "neurodivergent" as a functional term for non-uniform cognition and as an analogy for the kind of variance the empirical section measured across model variants. The paper's practical policy suggestion, that open-source diversity may be safer than closed rigidity for the specific problem of alignment, is offered as an argument for further work, not as a settled scientific fact. This piece states that carefully because the difference matters.
The gap between "cognitive diversity may be a protective factor" and "our clinical group is better" is the whole game. The first is a structural argument that different filters see different things, and that a monoculture will miss the same things in the same direction at the same time. The second is a superiority claim, and it is not what the paper makes. The ARC/Eden programme's own working paper on the Polymathic Neurodivergent Profile is explicit on this point: PNP is descriptive and non-diagnostic, and it disavows superiority claims on the same page it describes the profile. The PNAS Nexus finding is cited in that working paper as an independent structural argument for cognitive diversity, not as a licence for grand narratives about the researcher's own cognition.
The paper names a real problem for alignment research. If external verification is bounded by undecidability results, and if uniform training tends to produce uniform blind spots, then teams staffed and trained the same way will systematically miss the same failure modes. The argument does not solve alignment. It re-frames one of its constraints. That re-framing is what the authors offer, and what a careful reader should take from it. Anything stronger belongs to the reader's own extrapolation, not to the paper.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.