A preprint posted to arXiv claims to prove, using formal logic, that AGI alignment cannot be verified - not because aligned AI is impossible, but because the math won't let you confirm it.
The paper establishes two impossibility theorems grounded in classical computability theory. The authors anchor their argument at what they call "Trakhtenbrot's Wall" - a reference to a 1950 result showing that logical validity over finite structures is undecidable. From there, they derive a trilemma: any safety verification system must sacrifice at least one of soundness, completeness, or tractability. That's not an engineering gap or a funding problem. It's a structural consequence of descriptive complexity. The paper also argues that current containment strategies - sandboxes, halting-based architectures, bounded environments - don't escape these limits; they just trade off logical expressivity to carve out decidable fragments of safety.
This matters because the AI safety field has largely operated on the assumption that alignment is hard but solvable with enough research. A formal impossibility result, if it holds up to peer review, would shift that framing significantly - from "we haven't solved it yet" to "we can never fully verify a solution exists." That's a different kind of problem, and it has direct implications for how regulators and labs should think about deployment thresholds.
The paper is a preprint and hasn't cleared peer review, so treat the conclusions as provocative rather than settled. Still, leaning on Rice's theorem, Gödel, and Trakhtenbrot is at least working in well-established territory - these aren't fringe results.