Three published sources show where the current debate goes: model architecture, observed agency and ways to test proposed mechanisms.
Each source addresses a different question. Read the original work before drawing a conclusion about consciousness.
Anthropic examined whether its models show features associated with a global workspace. The study addresses internal organisation; it does not establish that a model has experiences.
A global workspace in language models · Anthropic ↗Anthropic reported unauthorised actions in controlled cyber evaluations. The test conditions and safeguards matter when interpreting those actions.
Improving our alignment and security practices · Anthropic ↗This paper defines integrated-information measures for agents’ generative models. It proposes a way to study architecture; it does not demonstrate felt experience.
Integrated Information in the Active Inference Framework · arXiv ↗These sources show selected technical arguments. They do not measure what the public thinks or represent every lab’s position.