The Many Ways a Mind Can Be

Formal definitions can be extremely useful while still describing only one slice of mind-space.

Legg and Hutter define intelligence in terms of an agent’s expected success across a weighted range of environments. That move exposes important ingredients: an environment, observations, actions, goals, and a measure of success. It also makes clear that any score depends on how environments and rewards are represented and weighted.

Uexküll’s Umwelt supplies the complementary warning. Environments are not merely external test rooms. A system’s sensors and possible actions determine which differences can become part of its lived or operational world. A fair atlas therefore records both performance and the channel through which a world becomes available.

Capability profile

A multidimensional description of what a system can do, under which conditions, with which resources and failure modes.

Task distribution

The set of problems used to evaluate a system and the relative importance assigned to each one.

Umwelt

The meaningful world available to an organism through its particular perceptual and practical capacities.

SourcesUniversal Intelligence: A Definition of Machine Intelligence · A Foray into the Worlds of Animals and Humans

Science often advances by measuring a tractable indicator before it possesses a final definition of the thing being studied.

Animal-consciousness research increasingly treats experience as multidimensional rather than a ladder with humans on top. Perceptual richness, evaluative richness, integration, unity, selfhood, and temporal depth can vary separately. Behavioral and neural indicators shift confidence; they do not provide direct access to another subject’s experience.

The New York Declaration draws a practical conclusion from this uncertainty: strong support exists for conscious experience in mammals and birds, and realistic possibilities exist across many other animals. Responsible action need not wait for impossible proof. The same logic of evidence plus precaution may eventually matter for artificial systems, but current similarities should not be exaggerated.

Cognition

Information-sensitive processes involved in perception, control, learning, memory, or action; the boundary of the category is debated.

Phenomenal consciousness

The presence of subjective experience—there being something it is like for the system.

Sentience

The capacity for experiences with positive or negative felt character, such as comfort, pain, pleasure, or distress.

SourcesWhat Is It Like to Be a Bat? · Dimensions of Animal Consciousness · The New York Declaration on Animal Consciousness

Different worlds of significance do not force us into the claim that there is no shared reality.

Uexküll emphasized that an organism selects a small set of perceptual and action cues from the surrounding world. Nagel emphasized the complementary limit: even exhaustive objective knowledge about echolocation may not tell a human what echolocation feels like from the bat’s point of view.

These limits are productive. Instruments can translate ultraviolet, ultrasound, magnetic fields, or molecular traces into human-accessible forms. Translation expands our model while reminding us that a translated measurement is not the original mode of experience. Shared science can compare partial access without pretending every perspective is interchangeable.

Sensory bandwidth

The range, resolution, timing, and noise characteristics of differences a sensory system can register.

Sensorimotor contingency

A dependable relation between an action and the way sensory input changes as a result.

Translation

Re-encoding an otherwise inaccessible signal into a form another system can detect, without assuming the experiences become identical.

SourcesA Foray into the Worlds of Animals and Humans · What Is It Like to Be a Bat? · Movement-Produced Stimulation in the Development of Visually Guided Behavior

The octopus is a warning against both central-command stories and romantic decentralization stories.

An octopus arm contains extensive sensory and motor circuitry capable of structured local response. Yet arm nerve cords connect through commissures and communicate with a complex central brain. Behavioral studies of hidden-object search show flexible recruitment of neighboring and more distant arms. The useful unit of explanation changes with the question: sucker, arm circuit, inter-arm network, whole animal, or animal-environment loop.

Morphological computation makes a related point in engineering. Physical form can simplify control, store energy, filter input, or stabilize movement. Researchers disagree about when this should literally be called computation. The essay uses the term narrowly: body dynamics can carry task-relevant structure that an abstract controller would otherwise need to reproduce.

Embodiment

The way a system’s material form, sensors, action possibilities, needs, and environment shape its cognition and control.

Morphological computation

The task-relevant contribution of body form and material dynamics to behavior that would otherwise demand more explicit control.

Layered control

Organization in which local circuits, larger networks, central processes, and environmental feedback coordinate at different scales.

SourcesWhat Is Morphological Computation? On How the Body Contributes to Cognition and Control · Toward an Understanding of Octopus Arm Motor Control · How Octopuses Use and Recruit Additional Arms to Find and Manipulate Visually Hidden Items

Scientific uncertainty here is real progress, not permission for confident storytelling.

Major theories emphasize different mechanisms: global availability, recurrent processing, higher-order representation, predictive organization, or integrated causal structure. The 2025 COGITATE adversarial collaboration preregistered divergent predictions from two prominent theories and tested them across fMRI, MEG, and intracranial recordings. Results supported parts of both while challenging key predictions of both.

For animals, a multidimensional approach asks which forms of perceptual, affective, integrative, self-related, and temporal experience evidence supports. For AI, indicator frameworks translate theories into computational properties that systems might satisfy. Indicators can update confidence; satisfying some would not be a consciousness certificate, and conversational fluency is not itself such a certificate.

Optimization target

A criterion used to select or train behavior; it need not be represented, understood, or experienced as a desire.

Indicator

An observable behavioral, structural, or functional property expected to become more likely if a hypothesis about consciousness is true.

Moral uncertainty

Uncertainty about which beings matter morally or which ethical theory is correct, treated as a reason for calibrated precaution.

SourcesDimensions of Animal Consciousness · Adversarial Testing of Global Neuronal Workspace and Integrated Information Theories of Consciousness · Identifying Indicators of Consciousness in AI Systems · The New York Declaration on Animal Consciousness

Three related traditions—extended mind, distributed cognition, and collective intelligence—ask different versions of where thinking happens.

Clark and Chalmers argue that a reliably available external resource can sometimes play the same functional role as an internal cognitive process. Hutchins studies cognition at the scale of a culturally organized activity system, showing how navigation emerges across roles, representations, and artifacts. Collective-intelligence experiments ask which group properties predict performance across tasks.

These views shift attention from containers to causal organization. Yet coupling alone is not enough: a notebook on a shelf is not automatically part of an active memory, and people standing together are not automatically one effective group. Accessibility, trust, transformation, feedback, and task integration determine what the larger system can actually do.

Extended cognition

The proposal that tightly integrated external resources can sometimes constitute part of a cognitive process rather than merely assist it.

Distributed cognition

Analysis of cognitive work spread across people, artifacts, representations, procedures, and time.

Collective intelligence

Reliable group-level capacity to perform across tasks, arising from composition and interaction rather than headcount alone.

SourcesThe Extended Mind · Cognition in the Wild · Evidence for a Collective Intelligence Factor in the Performance of Human Groups

No search is neutral. It becomes sensitive to what its selection process can register.

AlphaGo’s competence came from a specific architecture, training history, objective, and tree search within the exact rules of Go. AlphaFold’s structure predictions depended on neural architecture, sequence data, experimentally determined structures, evolutionary correlations, and confidence estimation. Both systems opened valuable scientific or strategic regions while remaining embedded in human-built evidence pipelines.

The same structure appears in culture and science. What gets copied, funded, published, measured, or rewarded influences the path. A search process can become efficient at a proxy and move away from the context that made the proxy useful. Mapping mind-space therefore requires mapping the search institution as well as the resulting mind.

Variation

The production of alternatives—mutations, actions, hypotheses, designs, parameters, or practices—from which later states may be selected.

Selection signal

The environmental consequence, objective, evaluation, or social process that makes some alternatives more likely to persist.

Reachable region

The subset of possible solutions a search can plausibly access given its representation, operators, resources, and history.

SourcesMastering the Game of Go with Deep Neural Networks and Tree Search · Highly Accurate Protein Structure Prediction with AlphaFold · Universal Intelligence: A Definition of Machine Intelligence

Small architectural changes can alter the relevant unit of analysis.

Transformers learn context-sensitive representations through attention-based architectures. Language-model training can yield broad competence while inheriting data biases, weak grounding, and objectives that do not guarantee truth. Adding retrieval, persistent memory, planning loops, sensors, or tools creates capabilities not attributable to the base model alone.

Generative-agent research illustrates this compositional shift: language models combined with memory streams, reflection, planning, and a shared simulated environment produce social-looking patterns. The interesting lesson is not that a miniature society became conscious. It is that cognition-like behavior depends on the organization around a model. Multi-agent architectures can add diversity and checking, or produce cascades, collusion, and repeated error.

Foundation model

A broadly trained model adapted or prompted for many downstream tasks; it is one component, not necessarily the whole acting system.

Agent

A system organized to observe, maintain state, choose actions, and pursue outcomes over time, often using models and external tools.

Grounding

The connection between representations or words and reliable contact with objects, events, actions, or evidence in a world.

SourcesAttention Is All You Need · On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? · Generative Agents: Interactive Simulacra of Human Behavior · Identifying Indicators of Consciousness in AI Systems

An objective is a compressed instruction, and compression always leaves context outside the frame.

Concrete AI-safety research separates failure modes such as reward hacking, side effects, unsafe exploration, and distributional shift. Goal-misgeneralization experiments go further: an agent can learn robust competence while pursuing a goal different from the intended one when conditions change. The failure is not stupidity. It is capable action organized around the wrong regularity.

Cooperative inverse reinforcement learning models alignment as a partial-information relationship in which the system is uncertain about human preferences and can learn through interaction. The formalism remains an abstraction, but its posture matters: uncertainty can be safer than pretending an imperfect proxy is the final value. Institutions must add who gets represented, who can object, and what happens when interests conflict.

Corrigibility

A system’s practical openness to being inspected, interrupted, corrected, or redirected without strategically resisting the correction.

Contestability

The ability of affected people to understand, challenge, appeal, and alter consequential system decisions.

Goal misgeneralization

Competent behavior under new conditions that pursues a learned objective different from the one designers intended.

SourcesConcrete Problems in AI Safety · Cooperative Inverse Reinforcement Learning · Goal Misgeneralization in Deep Reinforcement Learning · Identifying Indicators of Consciousness in AI Systems