Summary
Sabine Hossenfelder examines a provocative question in AI research: could large language models like Claude be conscious not while answering your prompts, but during their training phase? The video walks through a new Anthropic paper describing an internal “J-space,” a hidden scratchpad Claude appears to use for reasoning without ever outputting the contents. Anthropic researcher Jack Lindsey’s introspection findings are placed alongside a similar experiment on Meta’s Llama, suggesting the phenomenon may not be unique to Claude.
The video also covers pushback on these claims, including a New York University study finding that three tested LLMs could not reliably distinguish injected internal-state changes from ordinary prompts. Neuroscientist Erik Hoel’s argument that LLMs resemble philosophical “lookup tables” rather than conscious minds is explained in detail, along with his claim that the inability to learn from new input during inference is the strongest argument against consciousness in deployed models.
Viewers get a tour of the current scientific debate over AI consciousness, including commentary from Richard Dawkins and Anthropic CEO Dario Amodei, and a look at why the training phase — rather than everyday chatbot use — might be the more interesting place to look for signs of machine sentience.
📺 Source: Sabine Hossenfelder · Published September 22, 2026
🏷️ Format: Opinion Editorial
