Unlearning as Ablation: Toward a Falsifiable Benchmark for Generative Scientific Discovery

Bold claims about AI's role in science-from "AGI will cure all diseases" to promises of radically accelerated discovery-raise a central epistemic question: do large language models (LLMs) truly generate new knowledge, or do they merely remix memorized fragments? We propose unlearning-as-ablation as a falsifiable test of constructive scientific discovery. The method systematically removes a target result and its entire forget-closure (lemmas, paraphrases, and multi-hop entailments) and then evaluates whether the model can re-derive the result from only permitted axioms and tools. Success provides evidence for genuine generative capability; failure exposes current limits. Unlike prevailing motivations for unlearning-privacy, copyright, or safety-our framing repositions it as an epistemic probe for AI-for-Science. We argue that such tests could serve as the next generation of benchmarks, much as ImageNet catalyzed progress in vision: distinguishing models that can merely recall from those that can constructively generate new scientific knowledge. We outline a minimal pilot in mathematics and algorithms, and discuss extensions to physics, chemistry, and biology. Whether models succeed or fail, unlearning-as-ablation provides a principled framework to map the true reach and limits of AI scientific discovery. This is a position paper: we advance a conceptual and methodological argument rather than new empirical results.