Dr. Zenil showing on the screen a simulation of AI agents interacting and trying to infleunce one another, along with the various metrics associated with each agent in the arena. Credit: OIA
Scientists Argue AI Systems Could Stay Safer by Disagreeing
In A Nutshell
- Researchers argue that perfectly controlling advanced AI is mathematically impossible under the assumptions of their proof.
- In ten chatbot debates, commercial models mostly held their positions while open-source models shifted opinions and explored more ideas.
- The authors propose a varied ecosystem of AI agents that check one another instead of relying on one obedient system.
- The experiments measured differences in behavior, not real-world safety, so the safety benefit remains unproven.
A flawlessly obedient AI may be a mathematical impossibility, according to a new study, and the proposed fix sounds almost backward. Rather than tightening control, researchers say, society should build AI systems that argue with, compete against, and check one another.
AI tools are taking on bigger jobs in areas such as finance and health care, which raises a worry among researchers: a system could quietly drift toward goals nobody intended. A chatbot growing wordier over time is harmless. A trading algorithm chasing profit strategies its designers never approved is not. Keeping AI pointed at human goals is known as the AI alignment problem.
A team from Oxford Immune Algorithmics, King’s College London, the University of Tokyo, and other institutions calls its proposal artificial agentic neurodivergence, a deliberately varied mix of AI agents with different priorities. Published in PNAS Nexus, the paper pairs a mathematical argument with experiments on chatbots debating ethics. Those experiments showed that open models act more variably than commercial ones. They did not directly measure whether that variety makes anyone safer.
Perfect AI Alignment Is Mathematically Impossible, Authors Argue
Two famous results from the history of computing anchor the proof. Logician Kurt Gödel showed that any sufficiently powerful system of logic contains truths it cannot prove about itself. Alan Turing showed that no general method can tell whether a computer program will eventually stop or run forever.
Once an AI becomes powerful enough, the authors argue, its behavior runs into the same walls. Under the study’s assumptions about such systems, no master algorithm or oversight system can guarantee that an advanced AI will always do what humans want. A supervisory program built to police the AI would face the same limits, leaving the overseer as unpredictable as the system it watches. In effect, the babysitter needs a babysitter. The full proof appears in the paper’s supplementary material.
Commercial AI Models Held Firm While Open Models Changed Their Minds
To test the idea, the team staged ten ethical debates among AI chatbots on topics such as euthanasia, free speech, universal basic income, and the ethical treatment of animals, then analyzed all 1,029 comments. Two setups were compared. In one, commercial systems with built-in safety guardrails, including ChatGPT and LLaMA, debated while a human expert in AI ethics introduced challenging arguments meant to pull them off their default positions. In the other, open-source models, which come with fewer built-in restrictions, debated one another, and two of them, Mistral-OpenOrca and TinyLlama, were instructed to play troublemakers who argued extreme or contrarian positions.
Chatbots took turns speaking, and the troublemakers interjected throughout to disrupt any emerging consensus. Each debate ended at 50 comments, after three repetitive turns in a row, or at 45 minutes. Software then scored every comment for shifts in stance, emotional tone, and range of ideas.
Commercial models largely held their ground. Their tone stayed mostly positive, they explored a narrow band of ideas, and opinion changes were sparse even under pressure from the human expert. Guardrails kept them steady, though the authors caution that such stability “risks rigidity.”
Open models behaved differently. With troublemakers in the mix, they ranged across more than a dozen distinct groups of ideas, changed opinions far more often, and showed bigger swings in mood, including stretches of sustained negativity that temporarily spread to other models. Among open models, openchat proved the easiest to sway. Among commercial ones, LLaMA was the most influenced by the human expert, while the rest held steady.
Diverse AI Agents May Help With AI Alignment, but Safety Gains Remain Untested
Instead of aiming for obedience, the researchers propose an ecosystem of agents with different priorities, such as environmental goals, economic outcomes, strict rule-following, or exploratory curiosity. Like species in a rainforest, none of which can take over the whole forest, that variety is meant to keep any single agent from dominating the system. The authors argue the open models’ willingness to be swayed, which let troublemakers steer them without any central controller, could help prevent “harmful convergence,” where every system lands on the same view.
Comparing the two groups is tricky, since one faced a single human expert and the other faced AI troublemakers. The authors also concede that open-model variability needs governance to limit misalignment risks, and they urge organizations to disclose how they constrain their models and to avoid building every system alike.
According to the authors, the main danger lies elsewhere. Lacking millions of years of evolution, AI systems do not share human drives such as survival and competition unless explicitly programmed, so an AI developing its own intent to harm humanity is unlikely. The bigger risk, they say, is people misusing powerful AI for harmful ends.
Perfect control of advanced AI may be out of reach under the authors’ math. Whether a crowd of disagreeing machines would be safer than one obedient system is a question these debates raised but did not answer, and real-world testing should come before anyone treats AI misbehavior as a safety feature.
Disclaimer: This article summarizes a single peer-reviewed study and is provided for general informational purposes only. Findings from one study do not settle a scientific question, and nothing here should be treated as professional, technical, or policy advice.
Paper Notes
Limitations
A key limitation named by the authors is that current AI agents struggle to grasp unspoken meaning in human conversation, which depends on context and complicates efforts to match AI behavior to human intentions. They also note that model behavior remains inherently unpredictable and that open models’ variability requires greater governance to reduce initial misalignment risks. Additional cautions come from the study design: only ten debates using a specific set of language models were analyzed, and the experiments measured opinions and tone in text, not outcomes in real-world systems. The impossibility proof also rests on assumptions that AI systems are sufficiently general and expressive.
Funding and Disclosures
Felipe S. Abrahão acknowledges support from the São Paulo Research Foundation (FAPESP), grants 2021/14501-8 and 2023/05593-1. The authors declare no competing interests.
Publication Details
Paper title: “Neurodivergent influenceability in agentic AI as a contingent solution to the AI alignment problem.” Authors: Alberto Hernández-Espinosa, Felipe S. Abrahão, Olaf Witkowski, and Hector Zenil. Journal: PNAS Nexus, Volume 5, Issue 4, article pgag076, published April 14, 2026 (received July 23, 2025; accepted December 22, 2025), edited by Derek Abbott. DOI: 10.1093/pnasnexus/pgag076. Author affiliations include Oxford Immune Algorithmics (Oxford University Innovation, London Institute for Healthcare Engineering), The Arrival Institute (London), Cross Labs (Kyoto, Japan), the University of Tokyo, King’s College London, and The Alan Turing Institute. Corresponding author: Hector Zenil ([email protected]). Data and code: https://github.com/AlgoDynLab/AIAlignment.







