Anthropic Unveils Identity‑Proof System for Large Language Models

In early November, Anthropic announced work on an identity‑proof system for large language models such as its own Claude.^{1 2}

The core idea is to attach a unique cryptographic signature to each token produced by the model during inference. When a user receives an output from Claude, they can check the accompanying proof against a public registry to confirm that the text indeed came from that specific instance of the model. This mechanism would allow developers, regulators, and end‑users to trace content back to its origin without revealing sensitive internal parameters or training data.

From a technical standpoint, scaling this approach presents significant challenges. Experts note that even before accounting for signing operations, storing per‑token signatures could require more than ten terabytes of space for a very large LLM. Moreover, ensuring efficient communication between distributed compute nodes while maintaining low latency remains nontrivial. Nevertheless, the team believes that with careful engineering—such as batching signatures and compressing key material—the overhead can be kept within acceptable bounds for production deployments.

Beyond the engineering hurdles, the broader impact on trust in AI systems is substantial. By making it possible to verify provenance at the granularity of individual outputs, Anthropic’s tool directly addresses concerns around deepfakes and misinformation campaigns amplified through large language models. The ability to prove authorship also supports regulatory frameworks that demand accountability for automated decision‑making processes.

Developers must integrate signature verification into their pipelines, which may involve changes to existing security protocols and user interfaces. Legal questions arise about liability when signed text is misused; clear standards will need to evolve alongside the technology. Nonetheless, early demonstrations suggest that users can already validate Claude outputs against an immutable ledger, giving them confidence that the content originates from the intended source.

In summary, Anthropic’s identity‑proof system represents a meaningful step toward transparent, accountable deployment of powerful language models. While technical and policy challenges remain, the approach offers a concrete method to distinguish genuine model output from potential impersonations or tampering—a capability increasingly vital as LLMs become integral to business, media, and public discourse.