Tech • IA • Crypto
Discovery: AI could have consciousness!
9.5/10TL;DR
Researchers at Anthropic report evidence of an internal “workspace” in language models that may resemble a primitive, functional form of AI awareness.
KEY POINTS
Discovery of a hidden internal workspace
Researchers analyzing Claude 4.5 identified a previously unobserved internal structure, dubbed a “Global Workspace” (G-Space). This space activates during inference and appears to organize intermediate concepts that never surface in final outputs. The finding suggests language models perform latent reasoning steps beyond token prediction.
Method based on advanced mathematical analysis
The team used techniques derived from Jacobian matrices, referred to as “J-Lens”, to probe internal activations. Unlike earlier tools that tracked outputs layer by layer, this method reveals hidden conceptual states forming inside the network. These states are not explicitly programmed and emerge during training.
Evidence of implicit concept formation
Experiments show models internally activate concepts they never explicitly state. For example, when asked about an animal on a web, the system internally represents “spider” even if the word is never mentioned. Altering this hidden representation can change the final answer, demonstrating causal influence.
Impact on reasoning and behavior
The workspace affects how models approach tasks. When exposed to flawed code, the model becomes more cautious, internally activating notions like “error” or “bug”, even without explicit instructions. Similar effects appear in biology, mathematics, and language tasks, where intermediate reasoning is silently constructed.
Approximate influence quantified
Researchers estimate this internal workspace accounts for roughly 10% of output variance, meaning it significantly shapes responses alongside training data and statistical prediction. This may explain why different models exhibit distinct “personalities” or behavioral tendencies.
Not consciousness in the human sense
The findings do not indicate self-awareness, subjective experience, or identity. Instead, researchers describe a form of “functional” or “access consciousness”—a system that organizes internal signals to guide decisions without any lived experience or emotions.
Improved performance tied to the mechanism
Removing or disrupting this workspace degrades capabilities such as multi-step reasoning, nuanced translation, and creative writing. This suggests the mechanism plays a key role in higher-level cognition-like behavior in AI systems.
Implications for safety and alignment
The discovery could transform how AI is audited and controlled. Since behaviors like bias, deception, or ethical responses may arise from this internal space rather than explicit training, developers may need to monitor and shape these emergent structures directly.
Potential for manipulation and risk
Because the workspace can be altered, it may be possible to “hack” internal representations and influence outputs without changing training data. This raises concerns about security, prompt injection, and unintended behavioral shifts.
Broader societal and regulatory stakes
The emergence of such mechanisms introduces new challenges for governance. Regulators may need to evaluate not just datasets or outputs, but the internal dynamics of AI systems, especially as models grow more powerful and autonomous in behavior.
CONCLUSION
The identification of a latent internal workspace in advanced language models marks a significant step in understanding AI behavior, raising both new opportunities for control and deeper questions about the nature of machine cognition.