Embodied Cognition and Enactivism

Definition: Embodied cognition is the view that thinking depends on having a body that acts in a world, not only on a brain that processes symbols. Its strongest form, enactivism, says a mind brings forth its world through action. For AI, the question is whether a model trained on text can understand anything without seeing, touching, or doing.

TL;DR: Embodied cognition grew out of Merleau-Ponty's phenomenology and became a research program in 1991, with The Embodied Mind by Varela, Thompson, and Rosch and Rodney Brooks' robots. The 4E view says minds are embodied, embedded, enacted, and extended. A paper revised in 2025 argues that a body is "neither necessary nor sufficient" for a language model's words to refer. The debate is open. Build an AI agent free →

Try to explain to someone how to ride a bicycle using only words. You can describe balance, pedaling, and steering into a fall, and the listener will still fall over on the first try. The knowledge is in the riding, not in the description. Embodied cognition takes that everyday fact seriously. It says much of what we call thinking is done by a body moving through a world, and that a description of the world, however long, leaves part of it out. That is a direct challenge to systems that learn from text alone.

Why Embodied Cognition Matters for AI in 2026

Embodied cognition matters for AI because the most capable systems of 2026 learned mostly from text, which is the opposite of what the embodied view says minds need. If the embodied view is right, those systems are missing something basic. If it is wrong, or only partly right, then text may carry more of the world than its critics thought.

The Stanford Encyclopedia of Philosophy entry on embodied cognition (first published June 2021) describes the field as a reaction to classical cognitive science. Its two key claims are that "mental processes are not, or not only, computational processes" and that "the body or the body's interactions with the environment constitute or contribute to cognition." The entry groups the research into three themes:

• Conceptualization. The body shapes which concepts a creature can form. A body must "limit or constrain the concepts an organism can acquire."
• Replacement. Notions like symbol, representation, and inference are not the right tools, and some researchers argue they "must be abandoned in favor of others."
• Constitution. The body is part of cognition, not only a cause of it.

The question is now practical, not only philosophical. Robotics teams fit language models with cameras and motors, and AI researchers propose bringing enactive ideas into mainstream machine learning. Both moves test the embodied view against working systems.

The 4E View

The 4E view says a mind is embodied, embedded, enacted, and extended. The Wikipedia article on enactivism sums up the four.

The four claims differ in strength. "Embedded" is accepted by most researchers. "Extended" is the most contested, because it says a notebook can be part of a mind, not just a tool the mind uses.

Enactivism

Enactivism is the strongest version of embodied cognition: it says cognition "arises through interaction between an acting organism and its environment", in the words of the Wikipedia article. The term comes from The Embodied Mind by Francisco Varela, Evan Thompson, and Eleanor Rosch (MIT Press, 1991). Their phrase for the idea is that a species "brings forth and specifies its own domain of problems." A world is not a fixed set of facts waiting to be represented. It is what shows up as relevant to a creature that acts in it.

That is close to what Hubert Dreyfus said about relevance, and enactivism took the same idea into biology. In Enactive artificial intelligence: Investigating the systemic organization of life and mind (Artificial Intelligence 173(3-4):466-500, 2009), Tom Froese and Tom Ziemke asked what AI would have to build to be enactive in this strong sense. Their answer was that AI must take seriously "the organismic roots of autonomous agency and sense-making". They named two necessary requirements, constitutive autonomy (a system that generates its own identity) and adaptivity (a system that adjusts itself in a way that makes things matter to it), and they argued that current embodied AI falls short of both.

In 2026, Banafsheh Rafiee and Richard Sutton made a related case for current AI in Toward Enactive Artificial Intelligence. They argue that much of mainstream AI, from rule-based systems to large language models, has treated cognition as internal computation cut off from physical interaction. They name four enactive ideas they think matter most: experience, the inseparability of action and perception, autonomy, and embodiment. They see a structural kinship with reinforcement learning, which learns from acting and feedback, but they caution that the kinship is not equivalence.

The critics. The Wikipedia article lists the main objections. The "scale-up" objection says enactive accounts cannot explain complex thought, such as planning or mathematics, without representations. Andy Clark disputes purely action-based accounts of perception. Other critics say the early focus on sensorimotor skills is "cognitively marginal", a good account of catching a ball and a weak account of writing a proof.

The Extended Mind

The extended mind thesis says the tools we think with can be part of our minds. Andy Clark and David Chalmers proposed it in The Extended Mind (Analysis 58(1):7-19, 1998). They open with a question: "Where does the mind stop and the rest of the world begin?"

Their example is two people. Inga hears about an exhibition at the Museum of Modern Art, recalls that the museum is on 53rd Street, and goes. Otto has Alzheimer's disease and keeps a notebook with the things he needs to remember. He hears about the exhibition, looks up the address in his notebook, and goes. Clark and Chalmers argue that the notebook plays the same role for Otto that memory plays for Inga, so Otto believed the address before he looked it up. They call their position "active externalism", based on "the active role of the environment in driving cognitive processes."

Otto and Inga, side by side (Clark and Chalmers, 1998)

  INGA                                   OTTO
  ----                                   ----
  hears about the exhibition             hears about the exhibition
  recalls "53rd Street" from memory      reads "53rd Street" in his notebook
  walks to the museum                    walks to the museum

  same role, different location:  memory in the head  vs  memory on paper
  Clark and Chalmers: both count as believing the address
  Critics (Adams and Aizawa): the notebook is coupled to Otto's mind, not part of it

The critics. The SEP entry reports the best-known objection. Fred Adams and Ken Aizawa call it the "coupling-constitution fallacy": a thing that is causally coupled to a cognitive process is not thereby part of it. They also argue that "cognitive states must involve intrinsic, non-derived content", which a notebook lacks. Robert Rupert argues that minds use tools without being made of them. Clark replies that the extended claim is narrower than the critics suggest.

Does AI Need a Body?

Researchers give three answers to whether AI needs a body, and the evidence does not yet decide between them.

Answer 1: build bodies. Rodney Brooks gave the classic version in Intelligence without representation (Artificial Intelligence 47, 1991). His robots linked sensors to behavior directly, and the SEP entry quotes his idea that a robot can "use the world as its own model." Today the strongest form of this answer is the vision-language-action model. In RT-2 (2023), Anthony Brohan and colleagues fine-tuned a vision-language model on robot data and wrote robot actions as text tokens. In 6,000 evaluation trials, the robot handled new objects and followed commands absent from its robot training data, such as picking up the smallest object.

Answer 2: grounding without a body. Dimitri Coelho Mollo and Raphaël Millière argue in The Vector Grounding Problem (revised December 2025) that "multimodality and embodiment are neither necessary nor sufficient for referential grounding." On their account, a model's internal states can refer to things in the world through the causal history of its training, and a body on its own "does not automatically contribute" to that. Marcus Ma and Shrikanth Narayanan go further in Intelligence Requires Grounding But Not Embodiment (January 2026). They argue that "grounding, not embodiment, is necessary for intelligence", and they describe an intelligent LLM agent in a purely digital environment.

Answer 3: the body is the point. Strong enactivists, following Froese and Ziemke, reply that what matters is not sensors but the organization of life: constitutive autonomy and adaptivity, a system that keeps itself going and has something at stake in the outcome. On that view, neither a robot arm nor a text model is the right kind of thing. The symbol grounding problem article covers the related debate about meaning.

How Embodied Cognition Shows Up in Taskade

Embodied cognition shows up in Taskade through one idea, the extended mind. A Taskade agent has no body and no senses, and Taskade makes no claim that it does. What a workspace can be is the shared notebook: projects hold what the team must remember, and agents read them before they act. That is Workspace DNA, where projects are memory.

This is the extended mind idea in plain terms, not a claim that the workspace or its agents think. A Taskade AI agent reads the projects you attach, so the team's written knowledge travels with every run. The useful habit is the one Otto had: if it matters, write it where the next reader, person or agent, will look. Taskade Genesis builds a shared notebook like that from one prompt. To start from a working example, clone the Team Wiki Portal kit, an app kit with projects, an AI agent, and automations you can clone. The Free plan covers 3 apps and 1 agent and starts with 0 AI credits.

Describe yours and build it free →

Related Concepts

• Cognitive Science: Mind, Brain and AI: the hub for this category
• Phenomenology and AI: Merleau-Ponty and the roots of embodied cognition
• Symbol Grounding Problem: how symbols connect to what they name
• Dreyfus's Critique of AI: the argument that skill and relevance come from a body in a world
• Multimodal AI: models that read images, audio, and text
• Computer Vision: how machines turn images into usable signals
• World Model: an internal model of how the world changes
• Agent Environment: the world an AI agent acts in
• Memory Types: the kinds of memory an agent can draw on
• AI World Models: the longer story of world models in AI
• The Hard Problem of Consciousness: why experience resists a functional explanation
• Intentionality and AI: what it means for a thought to be about something

Frequently Asked Questions About Embodied Cognition

What is embodied cognition in simple terms?

Embodied cognition is the idea that thinking is done by a body acting in a world, not only by a brain processing symbols. Counting on your fingers, finding your way by landmarks, and learning to ride a bicycle are everyday examples. The view grew out of phenomenology and became a research program in the 1990s.

What is enactivism?

Enactivism is the strongest form of embodied cognition. It says cognition arises from the interaction between an acting organism and its environment, and that a creature brings forth its world through action. The term comes from The Embodied Mind by Varela, Thompson, and Rosch (1991).

What are the 4Es of cognition?

Embodied, embedded, enacted, and extended. Mental processes involve the body, depend on a supporting environment, include what an organism does, and can reach into tools in the environment. Researchers accept the four claims to very different degrees.

What is the extended mind thesis?

It is the claim, from Andy Clark and David Chalmers (1998), that tools such as a notebook can be part of a person's mind when they play the role memory would play. Their example is Otto, who uses a notebook to remember addresses. Critics argue that the notebook is coupled to the mind but not part of it.

Does AI need a body to understand?

Researchers disagree. Robotics researchers build vision-language-action models that act through a body. Coelho Mollo and Millière argue that a body is neither necessary nor sufficient for a model's words to refer to things, and strong enactivists argue that what matters is being alive, not having sensors.

Are language models embodied?

Standard language models are not embodied: they learn from text and do not act in a physical world. Some systems connect a language model to cameras and robot controls, such as RT-2 from 2023. Whether that gives them anything the embodied view would count is still debated.

What does embodied cognition mean for AI agents at work?

The practical lesson comes from the extended mind. An agent works better when the knowledge it needs lives in a shared, reliable place it reads before it acts, such as a project or knowledge base. That is a design habit, not a claim that the agent has a mind.