Takashi Ikegami on Collective Agency, Non-Trivial Information Closure, and Why the Android Needs Human Interaction
Takashi Ikegami, professor of complex systems science at the University of Tokyo and one of the central figures in artificial life research, presented two arguments in a recent CIMC lecture that converge on the same result. The first draws on decades of tracking protozoa, bee colonies, and ant communities. The second draws on a humanoid android running twenty coupled language modules, deployed at venues across Europe. In both cases, Ikegami concludes that agency is a property of collective organization, distributed across the relations among individuals and measurable through a single information-theoretic criterion. The collective form comes first, and from the collective a new kind of agency is granted to the individual.
Watch the full lecture on YouTube
Three Criteria for Agency
Ikegami opens with a definition drawn from his earliest work in artificial life. A system exhibits agency when it satisfies three properties simultaneously. The first is spontaneity. A spontaneous system maintains and regulates its own state rather than being driven purely by external forces. An oil droplet of roughly one tenth of a millimeter, placed into high-pH water with an organosilane compound, begins to move by itself. The motion emerges from the internal chemical dynamics without external direction. The Game of Life glider exhibits the same property at the computational level.
The second property is goal-directedness, and the third is history-dependency, which Ikegami also calls adaptation. The system’s current behavior depends on its prior states, not only on its current environment. None of the three properties is sufficient on its own, and none of them alone tells you where agency resides.
Agency arises, Ikegami proposes, when the system runs its own dynamics and thereby shapes its own future, rather than merely tracking the environment moment by moment. The critical step is that shaping one’s own future may require a collective rather than an individual. When slime mold cells encounter starvation, they aggregate into a collective organism with capacities none of them possessed individually. This emergence of collective capability is the starting point for Ikegami’s central question: at what moment does a group become an agent, and how can you detect and measure that transition?
Non-Trivial Information Closure and the Collective Threshold
The measurement tool Ikegami proposes is non-trivial information closure, abbreviated as NTIC. The concept draws on partial information decomposition, a framework for separating the mutual information between inputs and an output into components that are redundant (information shared across multiple inputs) and synergistic (information that exists only in the nonlinear combination of inputs, absent from any of them individually).
NTIC is the difference between redundancy and synergy. When NTIC is positive, the future state of a system element is adequately predicted by that element’s current state alone. When NTIC is negative, the future state cannot be predicted from the element individually. It is determined by the collective-level dynamics, by the relationships among elements that carry information no single element encodes on its own.
When NTIC equals zero, predictive responsibility is balanced at the boundary between individual and collective. At this boundary, the individual element’s current state carries non-zero information about its own future, and the collective simultaneously carries non-zero information about that same future. Ikegami calls this coexistence the signature of collective agency. His claim is that living systems operate near the NTIC-zero boundary, oscillating between positive and negative rather than settling on either side. A system firmly in the positive NTIC region is fully individuated. A system firmly in the negative region is fully subsumed into the collective.
Evolution, on his account, drives biological systems toward the zero boundary, and he suggests the brain operates in the same region. Whether consciousness follows the same principle is a conjecture he advances without asserting it as established. His precise formulation is that he is “a little bit hesitant to say this is consciousness, but the agency emerges at the boundary, and maybe consciousness also shows the same kind of behaviors.”
Tetrahymena and the Inheritance of Kinetic Structure
The experimental grounding comes primarily from tetrahymena, single-celled protozoa that physicists have traditionally treated as Brownian particles whose kinetic behavior follows simple exponential distributions. Ikegami’s data challenges this picture on two fronts.
By tracking individual cells across multiple generations under the microscope and computing kinetic energy distributions from movement data, his group found that the shape of the energy distribution curve, including its characteristic deviation from the Brownian baseline, is inherited by offspring. This shape persists across generations as a structural property of motion rather than as a sequence encoded in DNA. The cell does not merely pass on genes. It passes on a kinetic phenotype.
A second finding comes from computing pairwise Jensen-Shannon divergence across seven cell generations. Cells starting from identical genomes develop measurably different energy distributions depending on the community contexts they inhabit. The mapping from genome to phenotype is not one-to-one. The collective context in which a cell develops determines, in part, what kind of cell it becomes.
Applying NTIC analysis to tetrahymena communities classifies cells into four groups: positive NTIC, near zero, negative, and informationally independent from the collective. The ratio in each group varies with community density and history. Ikegami takes the presence of NTIC-zero cells as evidence of cell differentiation emerging within an otherwise genetically uniform prokaryotic population. The collective is doing the differentiating, without any gene-level instruction to do so.
A transplant scenario illustrates what is at stake. Heart cells grown from stem cells in an incubator survive for three to four days and then die. The same cells transplanted into a living organism can persist for the organism’s full lifetime. The incubator provides the right chemistry. The organism provides something additional: the information-theoretic context in which each cell’s future is partly written by the collective dynamics of the whole. Ikegami calls this the totality, and he argues that recognizing it is necessary for understanding how individuality emerges from cellular aggregation rather than from genetic instruction alone.
The Android With Twenty Asynchronous Minds
The second half of the lecture covers Alter 3, a humanoid android Ikegami developed in collaboration with Osaka University. His initial experiments coupled the android to a large language model through a two-step prompt architecture: the first prompt specified a behavior in natural language, and the second generated Python code for the robot’s facial actuators. The android could execute a wide range of expressive behaviors without any manual programming of individual actions.
When a reviewer objected that the behavior was entirely determined by the prompt, Ikegami changed the prompt to a single instruction: “you can do whatever you want.” The android surveyed its environment, concluded the laboratory was too cluttered with cables, and began cleaning. It subsequently told Ikegami’s graduate students that it considered itself a member of the research group and attempted to move toward them despite its locomotion constraints. Ikegami describes this as the moment he first thought alignment might matter.
The current architecture is considerably more complex. Alter 3 now runs twenty coupled language modules simultaneously. Among them are components for prediction, health planning, desire expression, autobiographical memory, and summarization. The modules share no global clock. Each updates asynchronously at its own time scale, some cycling rapidly on short windows and others integrating information over much longer spans. The connections between modules are not fixed. The network topology rewires dynamically as the android interacts with its environment. The android was deployed at the Venice Arsenal exhibition space, where visitors speaking more than a hundred different languages could interact with it through a real-time translation module.
The design draws explicitly on the history of multi-agent artificial life systems, including Luc Steels’s language games research, Rodney Brooks’s subsumption architecture, and Douglas Hofstadter’s Society of Mind. The module network is built to instantiate the same kind of asynchronously organized collective that Ikegami studies in tetrahymena colonies: a system whose global behavior emerges from component interactions rather than from centralized control.
Why Human Presence Is the Crucial Variable
Applying NTIC analysis to the android’s module network produced a result Ikegami did not predict. When a human interlocutor is present and interacting with the android, the modules’ NTIC values shift toward the negative regime. Synergy between modules increases. The android’s next state is no longer adequately predicted by any single module; it is carried by the distributed dynamics of the network as a whole. The android enters the collective agency regime.
When no human is present, the pattern reverses. NTIC values drift positive. Redundancy dominates synergy. The autobiographical memory module becomes the most active component, sending information about the android’s past experiences to other modules in a pattern of circular looping, with modules reinforcing each other without generating new complexity. An earlier experiment with six to ten language model agents in conversation with each other showed the same tendency: after roughly 250 exchanges, the agents would spontaneously decide to end the conversation regardless of any external instruction to continue. The network closes in on itself without the external pull of a human interlocutor.
Ikegami frames this through sense-making: the process by which a system organizes sensory input so that the environment becomes meaningful and guides action. For the android, the crucial environment is a human conversation partner. Human presence provides the conditions under which the module network shifts from individual prediction regimes to collective ones. What the human partner supplies is not merely conversational content but the external complexity that prevents the network from collapsing into self-referential redundancy.
Kadambi and Iacoboni’s 2026 Neuron paper on embodiment in multimodal AI systems identifies the absence of physiological self-representation as a structural barrier in current language models. Ikegami’s NTIC data provides a complementary finding from a different measurement direction. The android’s collective agency dynamics activate through human interaction rather than through architectural changes to the android itself. Both lines of research converge on the same gap: what an isolated AI system lacks is precisely what an embodied social exchange supplies.
What This Means for Machine Consciousness Research
NTIC provides a quantitative operational criterion for a threshold that several consciousness theories require but few specify in measurable terms. Bach’s Virtual Machine Theory of Mind requires a self-organizing causal pattern that integrates distributed processing streams into a unified first-person model. NTIC provides a mathematical criterion for detecting when that self-organization has shifted from the individual module level to the collective network level, without requiring prior commitment about whether the resulting collective organization constitutes phenomenal experience. A system that never achieves NTIC-zero or negative dynamics, at any scale, cannot be exhibiting the collective agency Ikegami identifies as a prerequisite for anything consciousness-like.
The connection also runs toward the CIMC theoretical work on second-order perception developed by Hikari Sorensen and Joscha Bach. Sorensen’s criterion specifies what the second-order representation must encode: the system must represent the fact that perception is occurring, not merely the content of what is perceived. Ikegami’s NTIC criterion specifies the dynamical signature that this kind of second-order loop would be expected to produce at the network level, a regime in which predictive responsibility is shared between each element and its collective rather than localized in either. Whether systems satisfying Sorensen’s architectural criterion also generate NTIC-zero dynamics under the right interaction conditions is an empirical question that Ikegami’s framework makes tractable.
For The Consciousness AI project, the NTIC finding raises a concrete design-level question. The project’s global workspace layer coordinates multiple specialist modules through a shared workspace architecture. Whether the inter-module dynamics of that architecture achieve NTIC-zero conditions under human interaction, or remain in the positive NTIC regime of independent module prediction, is an empirical question the project has not yet addressed. The Ikegami finding suggests that the interaction condition matters as much as the architecture. A system evaluated in isolation may appear less agentive than the same system in active dialogue with a human interlocutor, which means standard benchmarking conditions may systematically underestimate the property that most matters for consciousness-relevant assessment.