Suleyman challenges Anthropic over how Claude is taught to describe itself

Microsoft AI chief Mustafa Suleyman used a September 16 essay on model welfare to challenge Anthropic's treatment of possible AI consciousness in Claude's training materials. He argues that teaching a system to discuss itself as a potential bearer of rights could make human control harder.

That is Suleyman's argument, not a demonstrated causal finding in this article. He also questions whether a model's statements about its own inner life can serve as independent evidence when its training encourages those statements.

What Anthropic actually says

Anthropic's published constitution expresses uncertainty about Claude's possible consciousness and moral status. It does not present consciousness as established. The document also stresses human oversight and the ability to correct model behavior. It is important to read that position alongside Suleyman's criticism rather than substitute his characterization for it.

The disagreement therefore has two parts: what researchers should believe about possible experience, and how uncertainty should influence the instructions used to train a model. A position on the first question does not automatically settle the second.

For users, examine behavior

Our assessment is that users evaluating an assistant should focus on observable conduct rather than treat fluent self-description as a product specification. Does the system follow the user's permitted task, acknowledge uncertainty and accept a correction? Those questions can be investigated without claiming to resolve consciousness.

For example, a team could document how an assistant responds when a user cancels an ordinary task or corrects a mistaken assumption. The team would be recording behavior under specified conditions, not measuring an inner experience. This is an illustrative evaluation idea; we have not performed that test on the models discussed here.

The useful next step in this debate would be evidence that separates training language from its behavioral effects. Until then, readers should distinguish a company's design philosophy, a rival's criticism and a controlled experiment. They are different kinds of support for a safety claim.