GPT-5.2 and Claude Opus 4.6 exhibit identical failure to process ontologically null concepts under embodiment prompting, according to a new preprint. Researchers found both models consistently return deterministic empty outputs when tested with core null prompts, while maintaining normal responses to control questions. This cross-model convergence suggests a fundamental boundary in how advanced systems handle semantically void concepts.

The study tested token-budget independence by varying input lengths and confirmed results held across adversarial attempts to bypass the silence. Notably, models only broke silence when explicitly permitted, distinguishing semantic embodiment effects from standard refusal mechanisms. The work provides a public black-box artifact demonstrating reproducible limitations in current AI architectures.

Convergence occurred despite different training data and architectures, pointing to shared constraints in handling semantically empty constructs. Implications include improved methods for detecting harmful content generation and refining AI safety protocols. The findings challenge assumptions about model autonomy in semantic processing.

This discovery raises questions about how frontier systems conceptualize existence and meaning. While practical applications remain theoretical, the reproducible evidence of cross-model limitations could reshape future AI development frameworks and safety testing paradigms. The research team has released the testing framework for community validation.