The AI safety community has been wrong about the fundamental nature of its core problem. Perfect alignment isn't hard to achieve—it is mathematically impossible. Researchers at King's College London published proof this week in PNAS Nexus, ending the pretense that the field can engineer its way out of a structural impossibility.
The proof, led by associate professor Hector Zenil, relies on two of mathematics' most celebrated results: Gödel's incompleteness theorems and Turing's halting problem. Gödel demonstrated that any sufficiently powerful formal system contains statements that can never be proven true or false. Turing proved that no general algorithm can determine whether an arbitrary program will ever halt. Together, these establish that any AI complex enough to exhibit general intelligence will produce behavior that cannot be predicted or perfectly controlled from the outside.
The conventional wisdom among AI safety researchers assumed misalignment was a bug—something that improved data, more compute, or superior engineering would eventually eliminate. Zenil and his colleagues have dismantled that assumption. Their results show that structural misalignment isn't a symptom of insufficient optimization. It is woven into the fabric of universal computation itself.
This proof has immediate consequences for the trillion-dollar AI labs racing to build safe superintelligent systems. Companies like OpenAI, Anthropic, and Google DeepMind have built entire safety frameworks on the premise that perfect alignment is achievable with sufficient effort. Their regulatory filings, investor presentations, and public commitments all assume containment is possible. That assumption now rests on quicksand.
Regulators face an uncomfortable reckoning. Policymakers in the European Union, United States, and China have drafted AI governance frameworks premised on the idea that sufficiently advanced AI can be controlled. If alignment is structurally impossible, those frameworks require wholesale revision. The question shifts from "how do we ensure AI behaves?" to "how do we build systems that remain manageable despite inevitable misalignment?"
Zenil's team doesn't leave the field without direction. Their proposed solution—managing misalignment rather than eliminating it—inverts the entire approach. Instead of optimizing a single agent toward perfect alignment, the strategy involves designing a "cognitive ecosystem" populated by AI systems with different reasoning modes and partially overlapping objectives. These artificially neurodivergent agents would dynamically help or hinder each other, preventing any single system from achieving unchecked dominance.
The analogy is biological. The human brain evolved neurodivergence not as a defect but as a feature—a population of cognitive styles that collectively outperforms any single mode of reasoning. Zenil argues that robust AI safety may require the same: an ecology of intelligent agents where no one system can dominate because others are watching, competing, and constraining.
This reframes what success looks like. Perfect alignment was always a comforting fiction. The field's new task is building systems that remain controllable precisely because they are embedded in a web of competing interests and complementary limitations. Zenil's proof doesn't doom AI safety—it liberates it from a false destination and points it toward a achievable horizon. The impossibility theorem is not the end of the story. It is the beginning of a more honest conversation about what safety actually means when perfection is off the table.