AI’s Silent Thoughts

For most of the computer age, “Can machines think?” was largely a philosophical question. It is now becoming a practical one. The latest artificial-intelligence systems do more than produce fluent sentences. They plan, use software, inspect documents, correct mistakes and pursue complex goals over many steps. Researchers are also beginning to identify internal representations that appear before a model says anything.

Anthropic calls one such structure the “J-space”. Using a mathematical tool known as the Jacobian lens, researchers examined patterns moving through the layers of Claude. The lens connects certain patterns of neural activity with concepts the model might later express. In one experiment, Claude copied an unrelated sentence while silently calculating 3² minus 2. Its internal activity first reflected “nine” and then “seven”, although neither word appeared in its written response.

The activity was not merely a passive record. When researchers replaced the internal representation of “spider” with “ant”, Claude’s answer to a question about the animal’s legs changed from eight to six. Removing the J-space left automatic abilities such as fluent writing and information retrieval largely intact, but badly weakened multi-step reasoning. This suggests a small shared workspace where a few dozen concepts can be held and made available to different parts of the network.

Calling this an “inner monologue” is tempting, but potentially misleading. A model’s internal activity is not necessarily a hidden sentence passing through an electronic mind. It is a vast pattern of numbers interpreted through an imperfect scientific instrument. The discovery does not prove that Claude feels curiosity, fear, pain or pleasure.

Philosophers distinguish between *access consciousness*—information a system can report, reason about and use—and *phenomenal consciousness*, the felt experience of being someone. The J-space performs some functions associated with the first; it provides no evidence of the second. Even signs of a “Claude point of view” emerging during post-training may reflect learned self-monitoring, not a conscious self. The sound conclusion is neither “the machine is merely parroting” nor “the machine has awakened”. Its internal organisation is simply more structured and consequential than we once knew.

The research arrives amid fierce commercial competition. Anthropic’s Claude Fable 5.1 targets ambitious coding, knowledge work and long-running tasks across several applications. OpenAI’s newly released GPT-6 Astra is a direct challenger, combining reasoning with browsing, computer use, research and document creation. OpenAI reports Astra ahead on several coding, scientific and computer-use tests, while Fable 5.1 leads on at least one prominent tool-assisted academic benchmark. These comparisons require caution: results depend on the task, tools, safeguards, cost settings and who conducts the test. Intelligence has no single scoreboard.

The larger change is the move from chatbots that answer questions to agents that undertake work. They can navigate software, analyse data, prepare reports and recover from failure. Competition between OpenAI, Anthropic and others will accelerate this shift, lower costs and widen access. It may also encourage increasingly capable systems to be released before society has agreed on adequate oversight.

This combination of reasoning and action is commercially powerful. It can relieve people of routine digital work and extend expert assistance to smaller organisations. It can also magnify errors: a mistaken answer is limited, but a mistaken action may have immediate consequences.

Interpretability may therefore matter as much as performance. If we can detect when a model notices a hidden instruction, considers deception or recognises a test, we gain a valuable form of audit. But reading selected internal signals is not the same as understanding the whole system. The instruments may misread ambiguous patterns, while models may behave differently outside controlled experiments.

The practical response is informed engagement: use these systems, test them rigorously, retain meaningful human supervision and demand evidence for claims about them. Institutions will need clear responsibility for an AI agent’s actions, transparent records of important decisions and safeguards proportionate to possible harm.

We may still be far from conscious machines. Yet we are entering an era in which machines can silently represent ideas, weigh alternatives and act across the digital world. That is remarkable enough. We should remain open to discovery without surrendering either scepticism or human responsibility.