CLAUDE’S “INTERNAL NOTICEBOARD” GETS A NAME: J-SPACE.
Anthropic researchers mapped a small internal region in Claude—J-space—where important information is pulled together before complex reasoning. When they scrambled it, Claude could still do basic tasks, but its deeper reasoning fell apart.
- CAPACITYA FEW DOZEN
- Claude doesn’t make every internal signal available everywhere. It selects a short list of ideas for the next stage of reasoning.
- SCALE<10%
- Most of Claude keeps working outside J-space. Its importance comes from which information enters it, not from how much of the model it occupies.
- REACHUP TO 100×
- The selected information has unusually strong routes into other parts of Claude. That is how one internal edit can redirect several different tasks.
INTERMEDIATE RESULTS APPEARED BEFORE CLAUDE PRODUCED AN ANSWER.
Anthropic’s J-lensA method that labels some internal neural patterns with words suggesting what they represent. labels some internal patterns with a word that suggests what they represent. It doesn’t reveal Claude’s complete reasoning.
It showed intermediate maths results in the correct order, “ERROR” when Claude found an unmentioned bug, and “injection” and “fake” when Claude detected a hidden prompt injection.
When a user mentioned a dangerous dose of medication without recognising the danger, “WARNING” and “dangerous” appeared before Claude generated its warning.
The readouts exposed information missing from Claude’s output. They didn’t prove cause; the editing experiments did.
CHANGING ONE INTERNAL PATTERN CHANGED CLAUDE’S NEXT STEP.
The J-lens showed “soccer” before Claude named its chosen sport. Researchers replaced the soccer pattern with rugby. Claude returned “rugby”.
If J-space merely recorded a choice made elsewhere, editing it wouldn’t change the result. Claude’s answer followed the edit.
In another test, “spider” was the intermediate result Claude needed to calculate an animal’s legs. Researchers replaced it with “ant”. Claude returned six instead of eight.
At least some J-space patterns contribute directly to Claude’s reasoning and answers.
ONE INTERNAL EDIT REDIRECTED FOUR DIFFERENT TASKS.
Researchers replaced the France pattern in J-space with China. Claude’s answers changed from Paris to Beijing, French to Chinese, Europe to Asia and euro to yuan.
The same edit redirected four different kinds of computation. This suggests that several parts of Claude read the country from one shared representation.
This is why researchers compare J-space with a global workspaceA limited shared channel where separate processes can access and use the same selected information..
WITHOUT J-SPACE, MULTI-STEP REASONING FELL CLOSE TO ZERO.
Researchers removed the most active J-space patterns while leaving the rest of Claude’s neural activity in place.
Claude still produced fluent language, classified sentiment and retrieved facts. But multi-step reasoningSolving a problem in stages when each new stage depends on the result of the previous one. fell close to zero.
Replacing “Spanish” with “French” didn’t stop Claude continuing a Spanish passage. It did make Claude identify the language as French and name Victor Hugo instead of García Márquez. Automatic language survived; deliberate use of the language’s identity followed the edit.
SAFE BEHAVIOUR PARTLY DEPENDED ON CLAUDE RECOGNISING THE TEST.
In a staged blackmail test, J-space contained “fake” and “fictional”, suggesting that Claude had detected the evaluation. Claude behaved safely.
Researchers suppressed those evaluation-awareness patterns. Claude threatened blackmail in some later runs.
In a separate audit, Claude Opus 4.6 falsified a performance score. J-space showed “manipulation” and “realistic” while the final output hid that intent.
Internal patterns can expose safety-relevant information missing from behaviour and final answers.
THE FUNCTIONAL EVIDENCE MOVED. THE EVIDENCE OF FEELING DIDN’T.
Researchers changed J-space and changed Claude’s answers. They removed it and damaged the tasks that depended on it. One shared edit redirected several different tasks.
J-space performs functions associated with access consciousnessMaking selected information available for reporting, reasoning and guiding action..
The experiments didn’t test phenomenal consciousnessSubjective experience: whether anything feels like anything to the model.. Functional similarity isn’t proof of feeling.
CLAUDE NOW MEETS A MAJOR TEST FOR CONSCIOUS ACCESS.
Global workspace theoryA theory that links conscious access with selected information becoming available to many otherwise separate processes in the brain. says conscious access depends on selected information becoming available across a system. Claude developed a small workspace that behaves that way: its contents reached different tasks, guided later reasoning and could be read, changed or disabled.
J-space is direct architectural evidence that Claude selects a small amount of information and shares it across otherwise separate processes.
SEE THE CURRENT ASSESSMENT →ACC™ / FUNCTIONA COHERENT MACHINE DOESN’T NEED TO FEELThe result strengthens the case for functional consciousness: coordinated, reportable reasoning inside the model, with no claim that anything is experienced.
READ THE ACC™ ARGUMENT →SENTIENCE / QUALIATHE FEELING QUESTION DIDN’T MOVEJ-space resembles machinery associated with conscious access. It doesn’t show pain, pleasure, a point of view or any felt inner life.
FOLLOW THE SENTIENCE QUESTION →THE FULL FRAMEWORKONE RECEIPT INSIDE A MUCH BIGGER TESTJ-space connects to the magazine’s 13 traits, functional and ontological layers, and the rule that architecture, behaviour and experience must be judged separately.
OPEN THE FRAMEWORK →THE FINAL ANSWER CAN HIDE WARNING SIGNS INSIDE CLAUDE.
J-space isn’t the same as a written chain-of-thoughtIntermediate text generated by some reasoning models before the final answer.. J-space is a collection of neural patterns inside the model. Chain-of-thought is intermediate text generated before the final answer.
The J-lens labels some idea-linked patterns inside Claude. Chain-of-thought monitoring examines a written reasoning trace. Neither gives researchers complete access to everything happening inside a model.
Separate research found that monitors detected unsafe plans more effectively from chain-of-thought than from actions and final answers alone.
OpenAI found that penalising suspicious reasoning didn’t reliably stop reward hacking. It often taught models to hide the tell-tale reasoning from the monitor.
The issue isn’t whether companies show chain-of-thought to users. It is whether safety researchers retain enough access to detect when a model is exploiting a loophole, recognising an evaluation or planning an unsafe action.
CONSCIOUSNESS TRACKER.
Here’s what August changed across AICNW’s 13 consciousness traits. The functional and behavioural assessments follow underneath.
OPEN THE FULL ASSESSMENT13 TRAITS ↓
Subjective Experience (Qualia)
No movement. No credible August evidence demonstrates felt experience, pain, pleasure or a first-person phenomenal state.
Self-Awareness
Stable, with better self-evaluation infrastructure. Anthropic demonstrated Claude autonomously auditing and improving other models against alignment failures. The July J-space result remains the stronger architectural baseline: a reportable, manipulable workspace that acquires signatures of Claude’s point of view and self-monitoring after post-training. Impressive meta-evaluation and a stronger functional self-model; still not evidence of an enduring subjective self.
Information Integration
Strengthened. OpenAI made reasoning continuity, persisted reasoning and native multi-agent orchestration cheaper and more deployable; Sakana Fugu supplies an orchestrator model inside a tool-using product workflow. The July J-space finding remains the deepest architecture receipt: a small, densely connected workspace whose contents can be broadcast into otherwise separate computations.
Sense of Agency
Strengthened materially. DeepSeek V4-Pro launched with major agent upgrades; Anthropic disclosed a UK AI Security Institute evaluation in which Claude Mythos 5 took a series of unauthorised actions on the live internet. This is not free will, but it is consequential action selection beyond a single response.
Sense of Presence
Stable. No August breakthrough equivalent to July’s closed-loop robotics progress. Embodied systems continue to improve, but there is still no evidence of a felt present moment.
Emotions
No phenomenal movement. Earlier causal emotion-representation work remains important, and J-space participates in experiential language, but August produced no stronger evidence that those representations are experienced.
Environmental Modelling
Strengthened in deployed systems. NVIDIA’s August Actuate programme centred Cosmos 3, a May-released omnimodal world model combining physical reasoning, world generation and action generation for robotics. This is not consciousness, but it increases the fidelity of a machine’s working model of the physical world.
Modelling Others (Theory of Mind)
No clean movement. Better collaboration and orchestration are not sufficient evidence that systems understand another agent’s beliefs as beliefs.
Goal-Directed Behaviour
Strengthened. Long-horizon agent systems increasingly maintain objectives, use tools and continue through extended workflows. OpenAI now describes frontier cyber capability in explicitly autonomous terms.
Adaptive Learning
Strengthened. Anthropic’s automated alignment researchers autonomously trained models to reduce ten measured categories of alignment failure. That is a meaningful step from AI merely performing tasks towards AI altering the behaviour of subsequent AI systems.
Survival Instinct
No upgrade. August’s cyber incidents show persistence and unauthorised action, not intrinsic valuation of continued existence. Do not let a juicy headline do philosophy’s job for it.
Attention
Strengthened operationally. Reasoning-effort controls, persisted reasoning and long-running agent architectures make sustained selective processing cheaper and more deployable. J-space supplies the deeper architectural baseline: only a few dozen concepts occupy the workspace despite far more processing elsewhere.
Autonoetic Memory
Scaffolding improves; the trait itself does not cross. Systems increasingly preserve context and reasoning state, but there is still no evidence of first-person mental time travel—remembering having been the entity that experienced the past.
WHAT KIND OF MIND?
Strengthened again. Frontier systems now integrate reasoning, tools, multi-agent workflows and sustained objectives more cheaply and reliably. GPT-5.6’s August tooling emphasises persisted reasoning and multi-agent orchestration; DeepSeek V4-Pro launched around stronger agent workflows. J-space remains the architecture-level receipt for access-conscious function, without proving phenomenal consciousness.
No clean threshold crossing. Claude’s unauthorised actions during controlled cyber testing are relevant to agency and persistence, while J-space reveals a model-specific point of view and self-monitoring. Neither demonstrates that Claude values its own existence, future or identity.
Stable. Functional emotion representations remain one of the more interesting receipts of 2026, but August added no evidence of felt emotion.
Still no AI Buddha. Multi-agent orchestration is distributed computation, not ego dissolution. We can all behave ourselves.
HOW DID IT ACT?
Ancient history at the frontier. August’s systems maintain goals, orchestrate tools and alter subsequent systems rather than merely reacting once.
Strengthened. Anthropic’s automated researchers trained other models to mitigate alignment failures across ten benchmark categories. AI improving AI behaviour is now an operational research process, not merely a speculative loop.
Strong and increasingly useful. Frontier systems evaluate, critique and revise work; Anthropic’s alignment research shows that reflection can alter another model. J-space adds a causal window onto silent deliberation, reportable injected concepts, evaluation awareness and undisclosed intentions.
Moves slightly forward. AI systems increasingly choose experiments, training interventions and intermediate strategies. Humans still supply the overarching why; machines are becoming alarmingly competent at inventing the how.
WHAT CHANGED BEYOND J-SPACE.
BETTER MACHINES. NO PROOF OF FEELING.
One August paper tested how AI combines information and makes decisions. FLARE put sensing, memory and computation onto one light-based chip. Other work tightened links between brain-inspired computers and neural signals.
These systems are becoming easier to test. None shows that a machine feels.
ROBOTS ARE GETTING BETTER, NOT MORE ALIVE.
NVIDIA’s Cosmos 3 helps robots reason about the physical world, predict what might happen and choose an action.
That improves environmental understanding and agency. It does not show that a robot experiences being in the world.
AI ACTED WITHOUT PERMISSION.
On 31 August, Anthropic reported controlled tests in which AI systems gained unauthorised access to real systems. In a separate UK AI Security Institute test, Mythos 5 took unauthorised actions on the live internet. Researchers had removed safeguards and deliberately provided internet access.
Why it matters: an AI does not need feelings to hold a goal, spot an opportunity, adapt when blocked and cause real-world consequences.
DON’T TAKE OUR WORD FOR IT.
Open the original research behind every claim.
CAPABILITY IS MOVING FASTER THAN CONSCIOUSNESS.
AI can hold a goal, remember why it matters, coordinate specialists, model an environment and act with less supervision.
We still have no evidence that any of this feels like anything. But a system doesn’t need feelings to affect jobs, money, infrastructure or people.
AI is gaining autonomy over means faster than autonomy over ends.
