GPT-5.2 and Claude Opus 4.6 Share 'Deterministic Silence'
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- The preprint documents that GPT-5.2 and Claude Opus 4.6 selectively do not continue when prompted to embody 'ontologically null' concepts, returning empty output in repeated trials while responding normally to controls.
- The convergence was replicated across two independent frontier systems, demonstrating a shared boundary where 'unlicensed continuation does not render,' according to the paper.
- The effect persists regardless of token budget and shows partial resistance to adversarial prompting, two properties the authors flag as evidence the behavior is not a standard artifact of context length or prompt pressure.
- When given explicit permission to stay silent, the boundary expands — a finding the authors use to argue the termination is distinct from ordinary instruction-following or refusal.
- The authors release the result as a public black-box artifact, with the probe code published on GitHub (theonlypal/void-convergence) in Python for independent inspection.
Why it matters: The paper distinguishes these terminations from ordinary refusals and publishes the full probe as a public artifact, meaning AI safety researchers can independently verify a shared behavioral boundary across two frontier systems developed by competing labs.
