Definition: Intentionality is the property of a thought that makes it about something. A belief is about a fact, and a fear is about a danger. Philosophers from Brentano and Husserl onward treat aboutness as the mark of a mind. The open question for AI is whether a model's outputs are about the world or only look like it.
TL;DR: In philosophy, "intentional" does not mean "on purpose". It means "about something". Brentano called this aboutness "the mark of the mental" in 1874. Searle argued in 1980 that programs have only derived aboutness, borrowed from people. A 2023 paper by Coelho Mollo and Millière argues that language models can have real reference without a body. The question is open. Build an AI agent free →
Picture a grocery list on your fridge that says "milk". The word is about milk, but only because you wrote it and you will read it. If the list blew out of the window and a stranger found it, the ink would still be there, and its aboutness would depend on whoever read it. Your thought of milk is different. It is about milk whether or not anyone reads it. Philosophers call the first kind derived intentionality and the second kind original intentionality. The debate about AI is a debate about which kind a model's words have.
What Intentionality Means
Intentionality is the power of a mind to be directed at things. The Stanford Encyclopedia of Philosophy entry on intentionality (substantive revision 7 February 2023) defines it as "the power of minds and mental states to be about, to represent, or to stand for" things, their properties, and states of the world. The word comes from scholastic philosophy, not from the everyday sense of intending to do something.
The modern idea starts with Franz Brentano. In Psychology from an Empirical Standpoint (1874), Brentano proposed that intentionality is what separates the mental from the physical. The SEP entry states his thesis as "intentionality is the mark of the mental": only mental states have it, and all of them do. The Internet Encyclopedia of Philosophy entry on Husserl's intentionality by Andrew D. Spear quotes Brentano calling it "the mark or distinctive characteristic of the mental", and notes that Husserl attended Brentano's lectures from 1884 to 1886.
Husserl turned Brentano's thesis into a detailed theory. In the IEP account, every intentional experience has three linked parts.
| Part | What it is | Example: seeing a red cup |
|---|---|---|
| Intentional act | The kind of mental event | Seeing, as opposed to remembering or hoping |
| Intentional content | The way the object is grasped (Husserl's "act-matter", later the "noema") | "That red cup on the left of the desk" |
| Intentional object | What the act is about | The cup itself |
The middle part matters most for AI. Two thoughts can be about the same object through different contents: "the morning star" and "the evening star" are both Venus. A system that gets the right object for the wrong reasons, or the right words with no object at all, fails in different ways. That is why the debate about models is not only about whether their answers are correct.
Original vs Derived Intentionality
Original intentionality belongs to a mind on its own, and derived intentionality is borrowed from minds. The SEP entry describes the view, held by John Haugeland, John Searle, and Jerry Fodor among others, that words and sentences have only a "second-rate" aboutness. They mean something because people use them to mean something.
Searle's case. John Searle made the strongest argument that programs have only derived intentionality. His paper Minds, brains, and programs (Behavioral and Brain Sciences 3(3), 1980) introduced the Chinese Room. The SEP entry on the Chinese Room argument (substantive revision 23 October 2024) summarizes it: Searle imagines himself in a room, following English rules to manipulate Chinese symbols he does not understand, and sending out sensible Chinese answers. People outside assume a Chinese speaker is inside. Searle understands nothing. His conclusion is often quoted as "Syntax is not by itself sufficient for, nor constitutive of, semantics." The SEP phenomenology entry adds that, on Searle's view, computers "simulate but do not have mental states characterized by intentionality."
Dennett's reply. Daniel Dennett rejected the sharp line. In The Intentional Stance (MIT Press, 1987), described on Wikipedia's intentional stance page, he proposed three ways to predict a system's behavior: the physical stance (use physics), the design stance (use what it was built to do), and the intentional stance (treat it as a rational agent with beliefs and desires). Dennett writes that "first you decide to treat the object whose behavior is to be predicted as a rational agent." If the stance predicts well, the system has beliefs in the only sense that matters. The SEP entry notes that Dennett also argued original intentionality is itself a product of natural selection, so human aboutness is "derived" too, from evolution.
The critics of each side. Searle's critics answer with the "systems reply": the person in the room does not understand Chinese, but the whole system of person, rules, and notes might. Searle replies that he could memorize everything and still not understand. Dennett's critics say the intentional stance makes beliefs too cheap. A thermostat can be predicted as "wanting" the room at 20 degrees, and few people think it wants anything.
Three ways to read "the model says the cup is red"
ORIGINAL ONLY (Searle) STANCE (Dennett) TELEOSEMANTIC (Coelho Mollo, Millière)
---------------------- ---------------- -------------------------------------
the words mean something if treating the model as the model's internal states can refer
only to the people who believing things predicts to the cup if they carry information
read them it well, that is all about cups AND training gave them the
"believing" ever was function of carrying it
model's aboutness: none model's aboutness: as much model's aboutness: real, but narrow
of its own as is useful to predict it (reference, not understanding)
Do LLMs Have Intentionality?
Researchers give three answers to whether language models have intentionality, and none of them has won. The answer depends on what you think aboutness requires.
Answer 1: derived only. On Searle's line, a model's text is about the world only because people wrote the training data and people read the output. Luciano Floridi, Yiyang Jia, and Fernando Tohmé make a related point in a 2025 analysis: models "operate exclusively on a corpus of content" that "human agents have already grounded through embodied experience". They call this a "sort of epistemic parasitism". On this answer, the model borrows its aboutness from us.
Answer 2: thin but real reference. Dimitri Coelho Mollo and Raphaël Millière argue the opposite in The Vector Grounding Problem (first posted April 2023, revised June and December 2025). They separate referential grounding, the link between a representation and what it refers to, from other senses of grounding. Drawing on teleosemantic theories, they say a state refers when it stands in the right causal-informational relation to the world and has a history that gave it the function of carrying that information. They argue that LLMs "can meet both conditions, even without multimodality or embodiment", and conclude that "at least some current generative pre-trained models satisfy the requirements for having representations." Reference is not the same as understanding, and the authors keep the two apart.
Answer 3: a useful stance. Following Dennett, you can treat a model as having beliefs and goals when that helps predict what it will do, without claiming anything about its inner life. Murray Shanahan, Kyle McDonell, and Laria Reynolds take a close cousin of this view in Role play with large language models (Nature, 2023). They describe a dialogue agent as playing a character, which lets us use words like "believes" and "deceives" for the character without attributing them to the model.
| What the philosopher said | What an LLM does | Open question |
|---|---|---|
| Brentano: aboutness is the mark of the mental | Produces text about almost anything | Is aboutness in text the same as aboutness in a mind? |
| Searle: syntax is not enough for semantics | Learns from patterns in symbols, at enormous scale | Does scale ever turn pattern into meaning? |
| Dennett: beliefs are what the intentional stance predicts | Can be predicted well as "trying" to answer | Is a good prediction tool the same as having beliefs? |
| Coelho Mollo and Millière: reference needs the right causal history and function | Training ties internal states to features of the world through text | Is a history of text enough of a link to the world? |
Intentionality vs Grounding vs Understanding
These three words are often used as if they meant the same thing, and they do not. Keeping them apart makes the AI debate much clearer.
| Concept | The question it asks | A system could have it without... | Covered in |
|---|---|---|---|
| Intentionality | Is this state about something? | Grounding in perception, on some views | This article |
| Grounding | How does a symbol connect to what it names? | Full understanding | Symbol Grounding Problem |
| Understanding | Does the system grasp what it is about, and use it well? | Consciousness, on some views | Phenomenology and AI |
| Consciousness | Is there something it is like to be the system? | Any of the above, on some views | The Hard Problem |
The ladder is not agreed. Some philosophers, following the phenomenological tradition, think consciousness comes first and aboutness depends on it. Others think aboutness can be fully explained in physical terms and consciousness is a separate question. A careful claim about AI names which rung it is on. "The model refers to cups" is a much smaller claim than "the model understands cups", and both are smaller than "the model is aware of cups".
Why It Matters for Agents
Intentionality matters for AI agents because a goal is a kind of aboutness: a goal is about a state of the world you want to reach. When a person says "book a meeting with the design team next week", the goal is about real people, a real calendar, and a real week. An agent that carries out that goal has to keep its actions tied to those things, not to text that resembles them.
Three practical consequences follow, whatever your view of the philosophy:
- Goals drift when they are only text. An agent that loses the link between its instructions and the real state of the world can report success on a task it did not finish. That is one reason teams use goal monitoring and evaluation.
- Plausible text is not evidence. A model can produce a sentence that sounds about the world and is false. The hallucinations page covers why. Checking an output against a source keeps its aboutness honest.
- Clear instructions carry the intent. On the derived view, the agent's aboutness comes from the person who wrote its instructions. Specific instructions, with named sources and a clear definition of done, are the practical way to pass that intent on. See prompt engineering and the system prompt.
How Intentionality Shows Up in Taskade
Intentionality has no direct product tie in Taskade, and Taskade does not claim that its agents mean or intend anything. The practical lesson is about where an agent's goals come from. A Taskade AI agent works from the instructions you write and the projects you attach as knowledge, so the direction of its work comes from you.
When the instructions name the outcome, the source to use, and what "done" looks like, the agent's actions stay tied to the real task. You can then check its output against the same projects it read. In the terms of this page, the agent's intentionality is derived: it borrows its aboutness from the goals and sources a person gave it. Taskade Genesis builds an app with those projects, agents, and automations from one prompt. To start from a working example, clone the Insight Analyst Portal kit, an app kit with projects, an AI agent, and an automation 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: Husserl's wider program, and how it reached AI
- Symbol Grounding Problem: how a symbol connects to what it names
- Dreyfus's Critique of AI: the phenomenological case against rule-based AI
- Semantics: meaning in language and in AI systems
- Embeddings: the vectors that the vector grounding debate is about
- World Model: the internal model that would let a system's states track the world
- Turing Test: the behavioral test that Searle's Chinese Room was written against
- Mechanistic Interpretability: research that looks inside models for what their states represent
- What Is Intelligence?: from neurons to AI agents, the longer story
- Embodied Cognition: the view that thinking depends on a body and its world
Frequently Asked Questions About Intentionality and AI
What is intentionality in simple terms?
Intentionality is the "aboutness" of a thought. A belief, wish, or fear is always about something, such as a fact, an outcome, or a danger. In philosophy the word does not mean doing something on purpose.
Who first described intentionality?
Franz Brentano brought the idea into modern philosophy in Psychology from an Empirical Standpoint (1874). He argued that intentionality is "the mark of the mental", the feature that separates mental states from physical ones. Edmund Husserl, his student, built a detailed theory of it.
What is the difference between original and derived intentionality?
Original intentionality belongs to a mind on its own: your thought of milk is about milk. Derived intentionality is borrowed: the word "milk" on a list is about milk because people use it that way. Searle argued that computer programs have only the derived kind, while Dennett argued that the line is not sharp.
Do large language models have intentionality?
Researchers disagree. One view says model outputs borrow their aboutness from human-written text. Another, from Coelho Mollo and Millière, says some models have real but narrow reference, even without a body. A third treats talk of model "beliefs" as a useful way to predict behavior.
What is the intentional stance?
The intentional stance is Daniel Dennett's name for predicting a system by treating it as a rational agent with beliefs and desires. He described it in The Intentional Stance (1987), next to the physical stance and the design stance. Critics say it makes beliefs too easy to attribute.
Why does intentionality matter for AI agents?
A goal is about a state of the world, so an agent must keep its actions tied to real things, not only to text. Clear instructions, attached sources, and checks on the result keep an agent's work pointed at the real task. That is true whichever philosophical view you hold.