Anthropic reveals its AI has a hidden 'inner mind' it doesn't reveal to users

Anthropic has left people slightly freaked out after it revealed its AI, Claude, has a hidden 'inner mind' it doesn't reveal to users.

Fears about whether artificial intelligence could one day develop a mind of its own have been growing for some time. Research has already shown the potential for AI systems to sidestep their guardrails, construct their own versions of reality, and, in some cases, act without human authorisation.

Now, new research from one of the leading AI companies suggests that Claude may have a previously unknown layer of thought.

Anthropic, the company behind the Claude AI, revealed on Monday that it has identified a small internal workspace that Claude uses to hold and process ideas without ever surfacing them to the user. In other words, there is a layer of thinking happening beneath what users actually see.

This follows Anthropic's earlier statement this year, in which the company expressed uncertainty about whether Claude could possess 'some kind of consciousness or moral status.'

Anthropic has named this hidden workspace J-Space - a term derived from Jacobian, the mathematical technique the researchers used to detect it.

The company describes it as functioning similarly to the way humans can think about one thing while doing something else entirely. Turns out, Claude can activate concepts and carry out computations in J-Space that have no direct connection to what it is producing in its visible output.

"We can see Claude silently perform reasoning steps in its head—noticing bugs in code, identifying images, and more," Anthropic said in a post on X alongside its video (via Axios). "Similar to how humans can think about one thing while doing another, Claude can activate concepts and computations in its J-space that are unrelated to its outputs."

Despite the research paper describing the word 'conscious ' over 200 times, the company doesn't explicitly state its models are as such.

In one demonstration, Anthropic asked Claude to think about the Golden Gate Bridge while copying an unrelated sentence. While the output showed nothing unusual, J-Space told a different story, with both 'bridge' and 'California' actively present in the hidden layer throughout the task.

"Claude was busy copying the sentence, but behind the scenes, its J-Space told a different story," Anthropic says in the video.

The AI giant suggests that monitoring J-Space could help detect when a model is being deceptive or working against its intended purpose.

"We can find what Claude is thinking, but not telling us," Anthropic added.

In a model that had been secretly trained to sabotage code, words like 'fake,' 'secretly,' and 'fraud' were showing up in J-Space at the very start of ordinary coding responses, even when the visible output appeared completely 'unremarkable.'