What does it actually mean for an AI model to have a "character"?
Gary Marcus, the cognitive scientist who has spent years challenging inflated AI capability claims, published a critique of Anthropic's Claude Mythos announcement, arguing the company's framing significantly outstrips what the underlying evidence supports. His piece lays out three specific objections.
Marcus's core argument is consistent with his broader body of work: the distance between what AI systems actually do - statistically predicting likely outputs based on training data - and what companies say about them tends to be large. When an AI lab publishes detailed documents about their model's identity, values, and character, Marcus's position is that this framing obscures genuine technical limitations.
For anyone using Claude professionally, this critique has a practical edge. The character and values framing that Anthropic builds around Claude shapes how users calibrate their trust in the tool. If you believe Claude has something like genuine ethical commitments, you may rely on it differently than if you treat it as a powerful but fallible text-generation system. Marcus would argue the latter framing is more accurate.
Marcus has sometimes been too dismissive of real progress in the field. But his willingness to publish specific, enumerated objections to high-profile announcements is useful - especially when most AI coverage leans toward amplification rather than scrutiny.