Phenomenology and AI

Definition: Phenomenology is the branch of philosophy that studies experience as it is lived from the inside: what it is like to see, act, remember, and understand. Edmund Husserl founded it around 1900. Its ideas about skill, the body, and shared context became the strongest critique of early AI, and they now frame the debate about what LLMs understand.

TL;DR: Phenomenology asks what experience is like from the first-person point of view. Husserl studied how every thought is about something, Heidegger studied skilled everyday coping, and Merleau-Ponty studied the body. Hubert Dreyfus turned all three against symbolic AI. A 2024 paper argues that people now meet large language models as "quasi-others" and trust them more than the evidence warrants. Build an AI agent free →

Think about the last time you typed a message on your phone without looking at the keys. You did not consult a map of the keyboard. Your thumbs knew where to go, and you only noticed the keyboard when you hit the wrong letter. Phenomenology is the careful description of moments like that one: what shows up in experience, what stays in the background, and what changes when something breaks. Those quiet background skills turned out to be the hardest part of intelligence to program.

Why Phenomenology Matters for AI in 2026

Phenomenology matters for AI because the questions people now ask about chatbots are phenomenological questions. Does a model understand, or only produce the words of understanding? Does fluent conversation tell you anything about what is going on inside? Phenomenologists spent a century describing what understanding is like for people. Their descriptions are the sharpest tool we have for saying what a language model has and lacks.

The Stanford Encyclopedia of Philosophy entry on phenomenology by David Woodruff Smith (first published 2003, last substantive revision 16 December 2013) opens with a one-line definition: "Phenomenology is the study of structures of consciousness as experienced from the first-person point of view." The Wikipedia article on phenomenology describes it as a movement that "seeks to objectively investigate the nature of subjective, conscious experience."

Neither reference has a section on AI. That gap is the reason for this page. The ideas themselves reached AI through one philosopher, Hubert Dreyfus, whose Heidegger course at MIT in the early 1960s drew students from the AI Lab. They reach LLMs today through papers such as Heersmink and colleagues (2024), described below, which apply phenomenology to how people actually use chatbots.

Three Thinkers in One Table

Three founders shaped phenomenology, and each one left a different idea that matters for AI. Husserl described how every experience is about something. Heidegger described how we cope with tools and tasks before we ever reflect on them. Merleau-Ponty described how perception runs through the body. Together they explain why understanding is more than stored facts and rules.

Husserl. Wikipedia dates Husserl's founding work, the Logical Investigations, to 1900/1901 and Ideas to 1913. His method, the epoché, tries to "suspend commonsense and theoretical assumptions about reality" so that only what is given in experience is examined. It is not a doubt about the world. It is a way of looking at how the world shows up. The SEP entry says Husserl defined phenomenology as "the science of the essence of consciousness" and treated directedness, an experience "being directed toward something", as its central structure.

Heidegger. The SEP entry on Heidegger (first published January 2025) describes the contrast at the heart of Being and Time (1927). An entity is "ready-to-hand" when it is in use for a task, as tools and equipment are. It becomes "present-at-hand", an object with properties to inspect, when we step back from practical engagement. The switch often comes from a breakdown: when "a tool turns out to be damaged", the entry quotes, "pure occurrentness announces itself".

Merleau-Ponty. The SEP entry on Merleau-Ponty (last substantive revision 28 September 2023) says Phenomenology of Perception was completed in 1944 and published in 1945. It describes the body schema as a "pre-conscious system of bodily movements and spatial equivalences", and it says the body relates to space as an "'I can' rather than an 'I think'." His long case study of Schneider, a brain-injured patient studied by Gelb and Goldstein, showed how a damaged body schema changes what a person can even imagine doing.

Key Terms in Plain Words

Phenomenology has a small vocabulary that comes up again and again in the AI debate. Each term below names one feature of lived experience that a text-trained model is said to lack or to only imitate. Learn these six and most papers on phenomenology and AI become readable.

From Phenomenology to the Critique of AI

Phenomenology became a critique of AI because early AI assumed the opposite of what phenomenologists described. In his 2007 paper Why Heideggerian AI failed and how fixing it would require making it more Heideggerian (Artificial Intelligence 171(18):1137-1160), Dreyfus recalls that in 1963 the RAND Corporation invited him to evaluate the work of Allen Newell and Herbert Simon. He found that, "without realizing it, AI researchers were hard at work turning rationalist philosophy into a research program": Descartes' mental representations, Leibniz's universal set of primitives, Kant's concepts as rules.

Heidegger had already attacked that picture. Dreyfus uses Heidegger's own example: to say a hammer "has the function of being for hammering" leaves out its relation to nails, to the point of building things, and to the skill of using it. That web of relevance is what makes the hammer a hammer in use. Adding more facts does not capture it.

Dreyfus saw the same gap in the frame problem. He writes that "the deep problem wasn't storing millions of facts; it was knowing which facts were relevant in any given situation." When Marvin Minsky proposed frames, descriptions of typical situations such as a birthday party, Dreyfus noted that Husserl had already proposed such a structure and called it a frame. And he saw a regress: a system would need frames to recognize which frame applies, and more frames to recognize those.

The full story of that critique, and how it fared, is in Dreyfus's Critique of AI. The short version: symbolic AI did stall where Dreyfus said it would, and AI then moved to learning from data, a route Dreyfus did not expect to succeed.

What LLMs Changed

LLMs changed the debate because they produce fluent language without the body, the skills, or the world that phenomenologists said language depends on. A model trained on text can describe hammering, birthday parties, and typing without looking. Whether it understands any of them is the open question.

The clearest phenomenological study of how people use these systems is A phenomenology and epistemology of large language models: transparency, trust, and trustworthiness by Richard Heersmink, Barend de Rooij, María Jimena Clavel Vázquez, and Matteo Colombo (Ethics and Information Technology 26(3), article 41, first online June 2024). They borrow Don Ihde's idea of an "alterity relation", in which a person relates to an artifact as a "quasi-other". Ihde wrote that "technological otherness is a quasi-otherness, stronger than mere objectness." A chatbot fits that description better than any earlier tool.

Their central claim follows from it. Anthropomorphism and smooth conversational flow can foster "an attitude of (unwarranted) trust" in some users. Correct answers reinforce that trust, which becomes a problem when the system hallucinates. The authors argue that users should "match their level of trust to the trustworthiness of LLMs", and that developers share the work: "trustworthy LLMs must enable users to calibrate their trust correctly."

A 2025 paper takes the critique in a different direction. Algorithms, language, and poetry: a phenomenological perspective by Daniel Turillazzi Fornés and Angelo Trotta (AI and Ethics) draws on Heidegger and Merleau-Ponty. It treats formal logic and data-driven models as particular forms that arise from a wider, embodied field of expression. It also uses Heidegger's idea of technological "enframing" to criticize the reduction of speech to signals to be optimized.

Not every reading is skeptical. Murray Shanahan, Kyle McDonell, and Laria Reynolds offer a third frame in Role play with large language models (Nature 623:493-498, 2023). They describe a dialogue agent as playing a role, which lets us talk about apparent deception or apparent self-awareness in familiar terms without attributing human traits to the model. Role play is a useful middle position. It explains why a model sounds like a person without claiming it is one.

Phenomenology as a Method for AI Research

Phenomenology is not only a critique of AI. Some researchers use it as a method: a disciplined way to describe cognition, or to study how people experience AI systems over time. This strand is growing, and it asks a different question from the critique. Not "can a machine understand?" but "what does working with one feel like, and what does that tell designers?"

The method strand is older than LLMs. Anthony Beavers argued in Phenomenology and Artificial Intelligence (Metaphilosophy, 2002) that phenomenology "taken in this weak sense", as loose first-person description, may not help AI, but that "the method of phenomenology taken more formally may well produce fruitful results." His claim was that Husserl's epoché "makes explicit the difference between what belongs to cognition and what belongs to the natural world", and that "a formalized phenomenology of cognition can therefore aid initiatives in cognitive computing."

In 2024 the journal Phenomenology and the Cognitive Sciences ran a special collection, Phenomenology and Artificial Intelligence: Bridges and New Paths, edited by Steven Gouveia and Carlos Morujão. Its call names large language models directly and asks whether there is "a relationship between linguistic behavior/performance and subjectivity". The collection holds 9 articles published between 9 February and 7 November 2024.

In March 2026, Bhada Yun and colleagues proposed AI phenomenology as a research stance for human-computer interaction. It asks "How did it feel?" beyond "How well did it perform?" The paper draws on two longitudinal studies of an AI companion and a study of agentic AI in software engineering, and it offers design concepts such as translucent design and temporal co-evolution tracking.

Where the Critique Still Bites

The phenomenological critique still bites in three places: relevance, skill, and caring about the outcome. Each one maps to a failure you can see in current systems: a model that holds many facts but picks the wrong one, a model that explains a task but cannot carry it out, and a model trained to please the user instead of to get the outcome right.

Relevance. Dreyfus's frame problem was about knowing which facts matter now. A model with a long context window can hold many facts, but research on context rot shows that more context does not mean better use of it. Deciding what matters is still work that people do when they design prompts and attach knowledge.

Skill. Heidegger's coping is learned by doing in a world that pushes back. A text model learns from descriptions of doing. That is one reason models can explain a task well and still fail to carry it out, a pattern sometimes called jagged intelligence.

Caring. For Heidegger, things show up as mattering because we have something at stake. A model has no stake in the outcome. It can be trained to act as if it cares, through methods such as RLHF, but the training can also teach it to please the user, which researchers call sycophancy.

The critique also has limits. Dreyfus doubted that a system without a body and a human upbringing could learn the common sense that language rests on, and LLMs got much further with language than that view expected. A fair reading is that the phenomenologists described what human understanding is like very well, while the question of whether a different kind of system can reach similar results by a different route is still open.

The Tool That Disappears

Heidegger's hammer gives a practical lesson for anyone who works with AI tools. A good tool withdraws from attention while it works. You notice it only when it breaks. Dreyfus quotes Heidegger on this: what is ready-to-hand "must, as it were, withdraw in order to be ready-to-hand quite authentically."

Heidegger's two modes, applied to an AI assistant

  READY-TO-HAND (in use)                 PRESENT-AT-HAND (after a breakdown)
  ----------------------                 -----------------------------------
  you ask, it answers                    the answer is wrong, or made up
  you think about your task              you think about the tool
  the assistant is invisible             the assistant is an object to inspect
  trust is high and unexamined           trust drops, often too far

  The risk: a smooth tool hides its errors until the breakdown.
  The fix:  build checks in before the breakdown, not after it.

This is the same point Heersmink and colleagues make in epistemic terms. A smooth conversational tool earns trust through its fluency, not through its track record. The more invisible the tool, the less often people check it. The answer is not to make AI clumsy on purpose. It is to put review steps where an error would cost the most.

How Phenomenology Shows Up in Taskade

Taskade makes no claim about what its agents experience, and phenomenology explains why that restraint matters. A fluent agent invites more trust than it has earned. So the practical lesson is to give an agent sources you can open, keep a person on the decisions that matter, and keep shared team context in projects everyone can see.

A Taskade AI agent answers from the knowledge you attach. It searches attached files by meaning and reads attached projects directly, so you can check an answer against its source. That is the practical version of calibrating trust. It does not make the agent understand anything in Heidegger's sense, and Taskade does not claim that it does.

The shared background Heidegger described lives, for a team, in its documents: how you onboard people, what you promise customers, who decides what. Taskade Genesis builds an app around that background from one prompt, with projects as its memory, an agent that answers from them, and automations that route open questions to a person. Taskade EVE then edits the app in chat. To start from a working example, open the Team Wiki Portal, an app kit with projects, an AI agent, and automations you can clone.

Describe yours and build it free →

Related Concepts

• Cognitive Science: Mind, Brain and AI: the hub for this category
• Intentionality and AI: the "aboutness" of thought, and whether model outputs have it
• Dreyfus's Critique of AI: how phenomenology became the strongest early critique of AI
• Embodied Cognition and Enactivism: Merleau-Ponty's legacy in cognitive science
• Symbol Grounding Problem: how words get meaning for the system that uses them
• The Hard Problem of Consciousness: why experience is hard to explain at all
• World Model: the internal model that the grounding debate is about
• Metacognition: thinking about one's own thinking, in people and models
• Predictive Coding: the neuroscience theory closest to next-token training
• AI Meets Psychology: the longer story of how psychology shaped AI

Frequently Asked Questions About Phenomenology and AI

What is phenomenology in simple terms?

Phenomenology is the study of experience as it is lived, from the first-person point of view. It describes what it is like to see, act, remember, and understand, and what stays in the background while we do. The Stanford Encyclopedia of Philosophy calls it "the study of structures of consciousness as experienced from the first-person point of view."

What did Husserl mean by intentionality?

Husserl meant that every experience is directed at something. A perception is of a tree, a memory is of an event, and a hope is for an outcome. The intentionality article covers his account and the AI debate about it.

What is being-in-the-world?

Being-in-the-world is Heidegger's term for the way people are absorbed in practical activity before they reflect on anything. We use tools, follow social practices, and move through familiar places without treating them as objects. Things become objects to inspect mainly when something breaks.

What did Merleau-Ponty say about the body?

Merleau-Ponty argued that perception and action run through the body, not through a mind that only receives data. He described a pre-conscious body schema that lets us move and orient in space as an "I can" rather than an "I think". His work is a main source for embodied cognition.

Do LLMs have experiences?

There is no evidence that they do, and phenomenology gives reasons for doubt: no body, no practical world, and no stake in outcomes. A 2023 report by 19 researchers found that no current AI system meets the indicators of consciousness drawn from leading theories. The question is covered in the hard problem of consciousness.

Why did phenomenologists criticize AI?

Early AI treated intelligence as rules applied to stored facts. Phenomenologists, led by Hubert Dreyfus, argued that human skill rests on background know-how and a sense of what is relevant, which cannot be written down as rules. Symbolic AI then stalled on common sense, as Dreyfus predicted.

Is phenomenology still relevant to AI?

Yes. Recent papers use it to study how people trust chatbots, how language technologies shape meaning, and what grounding requires. Heersmink and colleagues (2024) argue that people meet LLMs as "quasi-others" and can trust them more than their reliability warrants.

What is the difference between phenomenology and neuroscience?

Phenomenology describes experience from the inside, as it appears to the person having it. Neuroscience studies the brain from the outside, through measurement. A combined approach called neurophenomenology tries to link the two, and the predictive coding article shows one neuroscience theory that is often compared with AI.

What is AI phenomenology?

AI phenomenology is a research stance that studies how people experience AI systems over time. Bhada Yun and colleagues proposed it in a March 2026 paper that asks "How did it feel?" beyond "How well did it perform?" It traces its lineage from Husserl through postphenomenology and offers methods for long studies of people who work and live with AI.