Speech Acts

Learning Objectives

By completing this book, you will be able to:

1. Explain Austin's central thesis—that many utterances do not describe the world but transform it—and why this idea remains revolutionary for human-AI interaction.

2. Distinguish the three levels of every speech act—locution, illocution, and perlocution—and diagnose which of them a specific misunderstanding occurs in.

3. Classify speech acts according to Searle's taxonomy (assertives, directives, commissives, expressives, declarations) and recognize which one is operating in a concrete request.

4. Apply Austin's and Searle's felicity conditions to diagnose why a speech act fails, distinguishing misfires from abuses.

5. Identify indirect speech acts and explain why they are the norm—not the exception—in human communication.

6. Analyze prompts and requests to AI systems as sequences of speech acts, decoding what the user says, does, and seeks to produce.

1. Austin and Performatives: To Say Is to Do

Until the middle of the twentieth century, the philosophy of language lived under an assumption that seemed self-evident: the function of language is to describe the world. Sentences are true or false according to whether they correspond with reality. Language is a mirror; reality is what it reflects. This view, called descriptivism, dominated logic and analytic philosophy for decades.

J.L. Austin, an Oxford philosopher, blew this view apart in a series of lectures at Harvard in 1955, published posthumously as How to Do Things with Words (1962). His observation was devastatingly simple: there are sentences that describe nothing. They are neither true nor false. And yet they are perfectly meaningful—because they do something.

“When a judge says ‘I now pronounce you husband and wife,’ he is not describing a marriage that already exists. He is creating the marriage. Before the sentence, there was no marriage; after the sentence, there is. The sentence is neither true nor false—it is successful or unsuccessful. If the judge lacks legal authority, the sentence fails, but not because it is ‘false’—because the conditions for it to work are not met.”

Austin called these utterances performatives: to promise, to baptize, to bet, to declare, to sentence, to apologize, to inaugurate. But his deeper discovery was that the distinction between descriptive (constative) and performative utterances did not hold up under close examination. Even apparently descriptive sentences do something: assert, inform, warn, suggest. Saying “it is raining” does not merely describe the weather; it can function as a warning, a justification, or a refusal, depending on context—exactly what you learned about implicature in Topic 1. Austin concluded that every utterance is, simultaneously, a description and an action.

1.1 The Three Levels of the Speech Act

Austin decomposed every speech act into three simultaneous levels. This trichotomy is the central analytic tool of this topic, because most communication failures—between humans, and especially between humans and machines—occur by confusing one level with another.

Locutionary act: what is said. The literal content of the utterance: the words, the grammar, the reference, the semantic meaning. It is what a dictionary and a grammar can decode without knowing anything about the speaker, the hearer, or the situation. “It's cold in here” as a locution is a statement about the temperature.

Illocutionary act: what is done in saying it. The action the speaker performs in uttering the sentence: asserting, questioning, ordering, promising, suggesting, warning. Illocution is the force of the utterance—what Grice would call speaker meaning, but structured as action. “It's cold in here” as illocution can be a request (close the window), a complaint (turn on the heating), or a warning (dress warmly).

Perlocutionary act: the effect produced. The real result in the hearer: convincing, frightening, motivating, irritating, consoling, confusing. Perlocution is not under the speaker's total control—you can try to convince and produce rejection, try to console and produce irritation. “It's cold in here” can produce a hearer who closes the window (desired effect), one who takes offense (undesired effect), or one who does nothing because they missed the indirect request (failed effect).

“A user writes to an AI system: ‘Our sales fell 15% this quarter.’ Locution: a factual statement about a metric. But what is the illocution? It could be: a request for analysis (‘tell me why’), a request for action (‘help me reverse it’), an expression of concern (‘I am alarmed’), or the opening of a narrative (‘let me give you context before my real question’). A system that processes only the locution will answer: ‘Understood, your sales dropped 15%.’ A system that catches the illocution will respond to what the user is doing, not only to what they are saying.”

2. Searle's Taxonomy: Five Fundamental Actions

John Searle, Austin's student and later a professor at Berkeley, took his teacher's intuitions and systematized them into a formal theory. His book Speech Acts (1969) and the later essay “A Taxonomy of Illocutionary Acts” (1975) established that every action performed with language falls into one of five fundamental categories, defined by the direction of fit between words and world.

Direction of fit is an elegant concept: when you speak, what must change for the act to be successful? Must the words change to correspond with the world (description), or must the world change to correspond with the words (action)?

2.1 Assertives: Saying How the World Is

Assertive acts commit the speaker to the truth of a proposition. To assert, inform, report, describe, diagnose, classify, predict. The direction of fit is words-to-world: if the assertion does not correspond with reality, the assertion is wrong.

“A user says: ‘Our team has 12 people and we use Salesforce as our CRM.’ This is an assertive establishing context. The speaker's responsibility is that the description be true. But remember from Topic 1: assertions contain presuppositions (the company has a CRM), can violate the maxim of quantity (perhaps there are 15 people but 3 are temporary and he does not mention them), and the user can sincerely believe something false. A rigorous system treats the user's assertives as useful but not infallible premises.”

2.2 Directives: Trying to Get the Other to Do Something

Directive acts attempt to get the hearer to perform an action. To request, order, beg, suggest, invite, challenge, ask. The direction of fit is world-to-words: the words express how the world ought to be, and success is measured by whether the world changes in that direction.

“‘Generate a Q3 sales report segmented by region.’ This is an explicit directive with clear parameters. But many directives are ambiguous: ‘I need to understand what happened with sales.’ The locution expresses a cognitive need; the illocution is a directive, but what action does it request? A report, a causal analysis, a comparison with competitors, a recommendation? The vague directive is the most common source of technically correct but practically useless responses.”

Directives vary in intensity: from the soft suggestion (‘perhaps we could consider…’) to the direct order (‘do it now’). This intensity, which Brown and Levinson analyze in terms of threat to the hearer's public face, determines the social conditions of the interaction. In interaction with AI, the user frequently modulates intensity without being aware of it: ‘could you…’ is more tentative than ‘I need you to…’, which is more tentative than ‘generate…’ The system should calibrate its level of certainty and commitment according to this intensity.

2.3 Commissives: Committing Oneself to Do Something

Commissive acts commit the speaker to a future action. To promise, swear, guarantee, offer, threaten, bet. The direction of fit is world-to-words, but the action falls on the speaker, not the hearer.

“‘If you give me a good marketing plan, I guarantee I'll get board approval.’ The user commits to a contingent future action. For Intention Architecture, commissives reveal expectations and implicit conditions. The user's guarantee declares a level of control over the board that may be illusory—it connects with the distinction between declared and real intention from Module I. An attentive system would notice the gap between the force of the commitment and the real capacity to fulfill it.”

Threats are negative commissives: ‘If the service doesn't improve, we'll change providers.’ The speaker commits to an action harmful to the hearer. For the AI system, detecting an implicit threat is critical because it changes the tone, the urgency, and the type of response required.

2.4 Expressives: Manifesting an Inner State

Expressive acts manifest a psychological state of the speaker with respect to something. To thank, apologize, congratulate, regret, complain, celebrate. They have no direction of fit: they neither try to change the world nor describe it, but to express how the speaker feels.

“‘I'm desperate, nothing I try works.’ The locution describes an emotional state. The illocution is complex: is it only an expressive (venting), does it contain an implicit directive (help me), or an implicit assertive (my situation is serious)? It connects with Damasio's somatic markers (Module I): emotion is not noise in the communication but critical information about the urgency, the importance, and the real state of the user. A system that ignores the expressive in order to jump to the implicit directive loses the human dimension of the interaction.”

Expressives also function as calibration signals: a user who expresses enthusiasm (‘this is incredible, I want to explore further’) is in a different motivational state from one who expresses skepticism (‘I'm not convinced, show me evidence’). The optimal response differs not because the question differs but because the speaker's state does.

2.5 Declarations: Creating Reality with Words

Declarations are the most powerful acts: they change reality by the mere fact of being successfully uttered. To declare war, appoint to an office, hand down a sentence, inaugurate, baptize, resign, dismiss. They have a double direction of fit: the words change the world AND the world changes to correspond with the words, simultaneously. But they require the speaker to have specific institutional authority.

“When a CEO says ‘I approve the project,’ she is not describing a pre-existing approval; she is creating it. But if an intern says the same words, the locution is identical and the illocution fails: the intern lacks the institutional authority for those words to create that reality. The declaration is the act in which the relationship between language and institutional power is most explicit.”

Declarations are relatively rare in interaction with AI, but they appear in important contexts: ‘I decide we're going with option B’ (declaration of decision), ‘I withdraw the previous request’ (annulment), ‘From now on, all queries must include ROI data’ (establishment of a rule). Recognizing them as declarations helps the system treat them with the institutional seriousness they carry.

3. Felicity Conditions: When the Act Fails

Austin introduced and Searle formalized the concept of felicity conditions: the requirements that must be met for a speech act to be successful. If an assertion is evaluated by truth or falsity, a speech act is evaluated by felicity or infelicity—by whether it works or does not work as an action.

3.1 Types of Failure

When the conditions are not met, Austin distinguishes two types of failure:

Misfires: The act is not performed at all. The attempt fails before it begins. Someone without authority tries to declare something; a promise is made about something impossible; an act requires a conventional procedure that is not executed correctly. The misfire is like a key that does not enter the lock: the action is never consummated.

Abuses: The act is formally performed but insincerely or defectively. Someone promises without intending to fulfill; apologizes without remorse; congratulates without meaning it. The key enters the lock but is hollow: the action is executed on the surface but lacks substance.

“A user says to an AI system: ‘We need to double sales next quarter.’ This is a directive act. Felicity conditions: (1) The user must have the authority to set that goal. (2) The goal must be at least potentially achievable. (3) The user must be genuinely committed. If the goal is structurally impossible (the market is contracting by 30%), it is a misfire: it asks for something that cannot be fulfilled. If the user knows it is impossible but says it to demonstrate ambition before his superiors, it is an abuse: the apparent illocution (directive) masks the real one (an expressive of organizational performativity).”

3.2 Searle's Four Conditions

Searle systematized the felicity conditions into four categories, each connected with concepts you have studied in previous modules:

Propositional content condition: What is said must make sense in the context of the act. A promise must refer to a future action of the speaker (you cannot promise what has already happened). A request must refer to an action of the hearer. This connects with Module II: if the content does not correspond with the reality of the system, the act fails before it begins.

Preparatory condition: The circumstances must be appropriate. To request, you must believe the hearer can do it. To warn, there must be a danger the hearer does not know about. This connects with feasibility: does the system have the capacity to do what the user asks? Does the user have the authority their act presupposes?

Sincerity condition: The speaker must have the mental state corresponding to the act. In promising, the intention to fulfill. In asserting, belief in the truth. In requesting, the desire that it be done. This connects with the gap between declared and real intention from Module I, and with Libet's experiments: the sincerity of a speech act is not guaranteed by the words.

Essential condition: The act must count as what it purports to be for both parties. A promise must be recognized as such. An order must be accepted as an order. This connects with declared vs. real function from Module III: what an act declares itself to be and what it really is can differ.

4. Indirect Speech Acts: The Form Deceives

If all of language worked directly—if questions always asked for information, assertions always described, and orders always used the imperative—communication would be simpler and duller. But the reality is that most everyday speech acts are indirect: the linguistic form performs one act and the illocutionary force performs another.

Searle devoted a seminal work to this phenomenon: “Indirect Speech Acts” (1975). The mechanism is as follows: the speaker issues a primary speech act (what they really want to achieve) through a secondary act (the literal form). The hearer processes both and, using context and the Cooperative Principle you studied in Topic 1, identifies the primary illocution.

“‘Can you pass me the salt?’ Secondary act (literal form): a question about the hearer's ability. Primary act (real force): a request to pass the salt. Nobody answers ‘Yes, I can’ and remains seated, because the hearer recognizes that the question about ability is the vehicle of a request. The Cooperative Principle operates here: the hearer assumes the speaker is not genuinely interested in their motor capacities but is asking for something, and responds to the primary act.”

4.1 Why We Speak Indirectly

Indirect acts are not an anomaly but the norm, for deep social reasons that Brown and Levinson systematized in their theory of politeness. Every request is an imposition on the other's autonomy. Indirectness softens that imposition: “could you close the door?” is socially more acceptable than “close the door.” Indirectness protects the face of both: the speaker does not appear authoritarian, the hearer does not appear subordinate.

But indirectness also arises from uncertainty. When the user does not know exactly what they want, they formulate indirectly: ‘I wonder whether it would be a good idea to…’ is not a definite request but a tentative exploration. Indirectness preserves flexibility: if the interlocutor responds negatively, the speaker can retreat (‘I was only thinking out loud’) without having committed their position. It is a strategy of linguistic optionality—it connects with the real options you studied in Topic 3 of Module III.

4.2 Indirect Acts in Interaction with AI

Indirect speech acts are the territory where human-AI communication meets its greatest difficulties. Humans formulate indirectly out of habit, politeness, uncertainty, or because they do not know exactly what they want. The system receives the literal form and must infer the real force.

“A user writes: ‘I wonder whether it would be a good idea to change our pricing strategy.’ Locution: an internal reflection shared aloud. Possible primary illocutions: (1) A request for analysis (‘evaluate whether we should’). (2) A request for validation (‘I've already decided but I want confirmation’). (3) The opening of a deliberation (‘I want to think aloud with you’). (4) A social probe (‘is it acceptable to question this?’). A system that answers ‘Yes, it would be a good idea’ or ‘No, it wouldn't’ treats the expression as a literal yes/no question. The pragmatically competent response depends on which is the real illocution—and when it is ambiguous, the right move is to make the interpretation explicit.”

The solution is not to demand that users speak directly. Indirectness is constitutive of human communication, and eliminating it would destroy the social functions it serves. The solution is to develop the capacity to recognize the probable illocution behind the form and, in the face of genuine ambiguity, to ask—without turning every interaction into an interrogation.

5. Prompts as Sequences of Speech Acts

A typical prompt to an AI system is not a single speech act. It is a sequence of interwoven acts that the system must decompose, classify, and respond to in an integrated way. Mastering this decomposition is the most direct application of speech act theory to Intention Architecture.

5.1 Anatomy of a Complex Prompt

“Prompt: ‘We're a 15-person startup at seed stage [ASSERTIVE: establishes context]. We've tried three acquisition strategies and none worked [ASSERTIVE + EXPRESSIVE: informs and conveys frustration]. We need to find an acquisition channel that works before our runway runs out [DIRECTIVE: asks for action with temporal urgency]. What do you recommend? [DIRECTIVE: requests advice].’”

The prompt contains at least four speech acts. A system that responds only to the last directive (‘I recommend X, Y, Z’) ignores the assertive context, the expressive frustration, and the temporal constraint. A pragmatically competent system integrates all the acts and responds: ‘Given your size, stage, prior failures, and runway constraint, I recommend…’ The response is better not because it is longer but because it answers the totality of what was communicated.

5.2 Four Errors of Illocutionary Decoding

Error 1: Treating everything as a directive. Not every sentence is an order. Sometimes the user is thinking aloud (expressive), establishing context (assertive), or exploring possibilities (tentative directive). Responding with action when the user was reflecting is as wrong as not acting when they were asking.

Error 2: Ignoring the expressive. When the user communicates frustration, urgency, or enthusiasm, that emotion is information about priority, seriousness, and the type of response required. Ignoring the expressive produces answers that are technically correct and emotionally disconnected.

Error 3: Failing to detect the hidden illocution. The user who asks ‘how much does X cost?’ frequently does not want a number but an assessment of feasibility. The one who says ‘show me the data’ frequently does not want a dump but an interpretation. The hidden illocution is the real request.

Error 4: Missing the perlocution. The user wants the response to enable them to make a decision (the desired perlocution). If the response overwhelms with options, confuses with technicalities, or fails to prioritize, the real perlocution is paralysis. The system's speech act also has three levels, and it must be designed for the desired perlocution.

6. Practical Application: The Request That Says Three Things

Daniela, a product director at a SaaS company, writes: “Our churn rate rose to 8% last month and the executive team is putting on pressure. I need an analysis of the main causes and a retention plan to present on Thursday.”

6.1 Analysis by Speech Acts

Act 1 — Assertive: ‘Churn rose to 8%.’ Establishes the fact. Contains presuppositions (that 8% is high, that it was previously lower, that churn is measured in a specific way).

Act 2 — Assertive + Expressive: ‘The executive team is putting on pressure.’ Informs and communicates urgency and pressure. Reveals that the desired perlocution is not only a correct analysis but one that protects her against executive pressure.

Act 3 — Directive: ‘I need an analysis of causes.’ Requests causal investigation. But it contains an implicit causal hypothesis: that there are identifiable ‘main causes’ (they may be systemic and diffuse, not isolable—as you studied in Module III).

Act 4 — Directive: ‘A retention plan.’ Requests a recommendation of actions. Presupposes that retention is the correct area of intervention (and not, for example, the acquisition of different clients).

Act 5 — Implicit directive: ‘To present on Thursday.’ This is not only a temporal constraint; it reveals the perlocution: the plan must be in a format presentable to a board under pressure. This changes the format, the tone, and the optimal level of detail.

“A system with good architecture would respond: ‘Daniela, before analyzing I need two things so the analysis holds up before the board. First: is the 8% churn against a stable baseline, or was it already rising? That changes the diagnosis. Second: do you have cohort data and data on the last interaction before abandonment? With that I can give you causality, not just correlation. I'll format it as a 5-slide executive presentation—which is what the situation calls for.’ This response integrates the assertives (uses the context), acknowledges the expressive (the pressure), responds to the directives with the necessary clarification, and designs for the real perlocution (looking good before the board, not merely having a correct analysis).”

6.2 A Speech Act Analysis Protocol

1. Decompose into constituent acts. What assertives (context), directives (requests), expressives (emotions), commissives (promises), and declarations (decisions) does the request contain?

2. Classify each act. Use Searle's taxonomy. Which is the dominant act that should guide the response?

3. Detect indirectness. Are there primary acts hidden beneath secondary acts? Is the question really a request? Is the assertion really a complaint?

4. Audit the felicity conditions. Is what is being asked possible? Does the evidence support the assertions? Is there sincerity between what is declared and what is real? Are the preparatory conditions met?

5. Infer the desired perlocution. What effect does the user want the response to produce? To understand, to decide, to convince others, to feel understood, to look good?

6. Respond to the illocution, not only to the locution. Integrate all the acts into a response that addresses what the user is doing, not only what they say. When the illocution is ambiguous, make the interpretation explicit.

7. Glossary of Key Terms

8. Final Reflection and Self-Assessment

Before advancing to the next topic (Context and Meaning, with Wittgenstein), verify that you can apply speech act theory as a tool for decoding intentions.

9. Bibliography and Recommended Readings

Essential Readings

• Austin, J.L. (1962). How to Do Things with Words. Oxford University Press.

• Searle, J. (1969). Speech Acts: An Essay in the Philosophy of Language. Cambridge University Press.

• Searle, J. (1975). “Indirect Speech Acts.” In P. Cole & J. Morgan (eds.), Syntax and Semantics, Vol. 3. Academic Press.

Complementary Readings

• Searle, J. (1975). “A Taxonomy of Illocutionary Acts.” In K. Gunderson (ed.), Language, Mind, and Knowledge. University of Minnesota Press.

• Searle, J. (1979). Expression and Meaning: Studies in the Theory of Speech Acts. Cambridge University Press.

• Brown, P. & Levinson, S. (1987). Politeness: Some Universals in Language Usage. Cambridge University Press.

• Sbisà, M. (2009). “Speech Act Theory.” In J. Verschueren & J-O. Östman (eds.), Key Notions for Pragmatics. John Benjamins.

• Grice, H.P. (1975). “Logic and Conversation.” [Connects with Topic 1]

Next Topic

The next book addresses Context and Meaning: the revolution of the later Wittgenstein. If Austin and Searle showed you that to speak is to do, Wittgenstein will show you that meaning does not reside in words but in their use within concrete forms of life. You will study language games, the private language argument, and the radical thesis that ‘the meaning of a word is its use in the language.’ You will discover that the same word can mean radically different things in different contexts—and that context, not the dictionary, is the true determinant of meaning.