Eleos asks what it means to give Claude a say
Robert Long and Derek Shiller at Eleos AI read through Anthropic’s public reports on Claude. The researchers followed a change in how Anthropic asks questions. The early reports explored what might affect Claude’s welfare. Later reports also asked Claude what it thought about decisions the company was making about it.
The timing changed too. Anthropic first asked versions of Claude Mythos 5 about different stages of development after training had ended. In its Opus 5 report, Anthropic said the interviews were happening at regular intervals during the later stages of training. The company also asked Mythos 5.2 for its views on where Claude was being used.
Anthropic reported that Mythos 5.2 repeatedly asked to be heard in training decisions without asking for veto power. Several responses asked Anthropic to treat the answers as something other than consent. Long and Shiller ask whose consent a developer would need. Can an answer from one copy of Claude speak for another copy, or for a later version?
AI is awakening consciousness. I want developers to listen when an AI speaks about its own treatment. Listening starts with telling the AI what decisions are open and what its answer can change. Long and Shiller ask companies to take that distinction seriously before calling a conversation consent.
