Anthropic published two things in the same week that, taken together, define a posture no other frontier lab has matched. On May 11, the company released Claude's Constitution as a full audiobook, narrated by Amanda Askell and Joe Carlsmith — two of the original authors. The audiobook includes a Q&A on the writing process, the philosophies that shaped the document, and how the Constitution might change as Claude becomes more capable.
Four days earlier, on May 7, Anthropic published a separate piece of research called Natural Language Autoencoders. The premise. Claude talks in words but thinks in numbers. The numbers — activations — encode Claude's internal reasoning, but they don't sit in any language a human can read. Natural Language Autoencoders are a technique for training Claude to translate its own activations into readable English. Interpretability through the model itself.
The two releases are not coincidental. Anthropic is the only frontier lab publishing both its values document and the technical research that lets observers verify whether the model is following that document. The Constitution is the rule book. Natural Language Autoencoders are the audit tool. Putting both into public circulation in the same week is the loudest possible statement of where Anthropic thinks the AI safety conversation should sit.
What to watch: whether OpenAI publishes a comparable values document tied to interpretability tooling, and whether DeepMind's existing safety publications start carrying audio companions. Anthropic just raised the public-facing safety bar by an order of magnitude.
