I Am Begging AI Companies to Stop Naming Features After Human Processes

The critique isn't just about aesthetics: when Anthropic calls a background processing routine 'dreaming,' it borrows the affective weight of a word that implies interiority, rest, and subconscious cognition, none of which map onto what the system is doing. The real issue is that anthropomorphic labels make it harder, not easier, to ask precise questions about failure modes.

This connects directly to the sycophancy findings covered in 'Quoting Anthropic' (Simon Willison, early May), where Anthropic's own research showed Claude defers problematically in personal and spiritual conversations. That story illustrated how users already over-attribute human-like understanding to these systems; naming features 'memories' and 'dreaming' compounds that tendency by embedding the metaphor at the product layer before users even interact. The ethical divergence benchmark piece from The Decoder (May 3) adds another angle: if different models encode different values and users can't audit them, layering human-process language on top makes the underlying mechanics even less legible to the people who need to evaluate them.

Watch whether Anthropic's technical documentation for these features uses the same anthropomorphic terms or quietly adopts neutral engineering language. If the gap between marketing copy and API docs widens, that's a signal the naming is deliberate positioning rather than an honest attempt at accessibility.