Speech acts and performativity: why continuous ingestion follows

Speech acts and performativity: why continuous ingestion follows

Some sentences describe the world. Others change it. "The cat is on the mat" is true or false depending on the cat. "I now pronounce you married" is not true or false at all — said by the right person, in the right setting, to the right people, it does not report a marriage, it creates one. Say the same words on a stage, in rehearsal, or without legal standing, and nothing happens. The sentence is grammatically identical and institutionally void.

This class of utterance does not fail by being false. It fails by misfiring. A cheque signed by someone with no account, a verdict read by someone not a judge, a resignation offered to someone with no authority to accept it — each produces the form of an act with none of its force. The utterance is not mistaken about the world; it fails to alter it. Philosophy needed a different currency than truth to assess these cases, because truth was built for descriptions and these are not descriptions.

That currency is felicity. A performative succeeds or fails against conditions: the speaker must hold the relevant office or standing, the form of words must be the one convention assigns to the act, and the hearer or institution must take the utterance up as that act. Miss any condition and the utterance misfires — not lied, not erred, just void. The world remains as it was, and the speaker who thought they had married someone, fired someone, or cleared someone for take-off discovers that thinking so did not make it so.

Where this came from

The philosopher J. L. Austin worked this out delivering the William James Lectures at Harvard in 1955, published after his death as How to Do Things with Words in 1962. His target was a piece of orthodoxy in the philosophy of language: that a sentence is meaningful only insofar as it can be verified true or false. Austin pointed at an entire category the orthodoxy could not place. Marriage vows, bets, verdicts, namings, bequests — all plainly meaningful, all plainly consequential, none of them assessable as true or false. He proposed felicity conditions in place of truth-conditions, and by the end of the lectures had generalised the point further: even ordinary assertions carry illocutionary force, doing something as well as saying something. John Searle later systematised the conditions into something closer to a checklist, and legal theory absorbed the whole apparatus — a contract, a plea, a statute, are performatives with named felicity requirements written into doctrine.

The turn

Put a language-producing system into a loop where its output is taken up as an act — an order sent to an exchange, a discharge instruction sent to a ward, a clearance read back to a cockpit — and Austin's apparatus applies to it whether or not anyone built it to.

A Large Language Model, trained on a corpus fixed at some cutoff, can produce the exact wording of a discharge order, a payment instruction, a takeoff clearance, with total fluency. What it cannot produce is any of the three conditions that would make the wording an act rather than a script. It has no office — nothing confers standing on it. It has no addressee — there is no hearer whose uptake could be checked. And crucially, it has no ledger: no record of what has already been committed, what has already been revoked, what is currently true of the account, the licence, the patient. It is not that the model lacks a body. It is that felicity is not a property of wording at all, and wording is the only thing a frozen corpus supplies.

A Large World Model adds a scene. It senses the room the act would land in: the counterparty is present, the runway is occupied, the patient's monitor shows a rhythm. This genuinely supplies uptake and immediate context — something the language-only system could never check. But felicity is not only a fact about the instant of utterance. It is also a fact that must remain true afterwards, and a fact that was made true by something prior — an authorisation granted last week, a consent given last month, a mandate that could be revoked at any time between grant and use. A scene that dissolves when attention moves cannot hold that. It checks felicity once, at one moment, and forgets.

This is the sense in which a Large Universe Model — every relevant stream still running, held as beliefs with provenance and revisable on new evidence — names the last position on this axis rather than a further improvement on it. What acting on the world felicitously requires is a maintained, sourced, revisable record of who currently may do what to whom, and of what has already been committed and by whom it might yet be undone. Once intake is that — everything still running, sourced, open to correction — there is no further category of evidence a felicity condition could demand. There is only more of the same kind: more streams, longer memory, better provenance, faster revision. The ladder does not grow a fourth rung; it runs out of new kinds of rung.

The misreading to disown

The common objection here is that language systems cannot act because they only produce text, and that giving them a body — sensors, actuators, embodiment — is the fix. This gets the diagnosis backwards. Text is the ordinary medium of institutional action. A warrant is words. An order is words. A dismissal, a diagnosis, a discharge, a transfer of title — all made of words arranged in an accepted form, all already indistinguishable in shape from the acts they constitute. A language model producing fluent tokens is not failing to act because it has no body. It is failing because it has no standing, no addressee, and no record of what currently holds. Bolt actuators onto that and the system does not gain felicity. It gains the ability to misfire with consequences in the physical world rather than only on a screen.

Tenerife shows the cost of a misfire that was purely verbal. On 27 March 1977, KLM 4805's captain transmitted "we are now at take-off" while beginning the roll, without a clearance from the tower. The phrase sat ambiguously between report and request; uptake failed on the other end; 583 people died. Aviation's response was not to add sensors but to fix the felicity conditions themselves: "cleared for take-off" became a reserved phrase, readback mandatory, hearback required. The fix was institutional discipline over what counts as the act, not a richer description of the runway.

Three objections, taken straight

The first: standing is granted by institutions, never inferred from observation. A registrar marries people because a state confers the office, and no volume of sensor data confers that office on a machine. This is correct, and should not be blurred. But grant the office, and the demand does not disappear — it sharpens. A delegated actor still has to establish, at the moment of acting, that the conditions of its delegation currently hold: that the account is not frozen, the consent not withdrawn, the mandate not superseded. Authority answers who may act. Intake answers whether the act, once performed, will be happy. Institutions settle the first question once. They leave the second question open every time, and it stays empirical.

The second: separate the concerns. Let the model describe; put a thin, deterministic, audited layer between description and commitment, holding live state itself. This is sound engineering and usually the right design. It relocates the requirement rather than eliminates it — the deterministic layer still needs a current, sourced picture of the felicity conditions to decide against, and someone still has to maintain that picture continuously. Where the conditions are few and enumerable, the sandwich works cleanly. Where they are too numerous and too fluid to write as rules — a shifting sanctions list, a network of revocable consents — the intake demand reappears inside the supposedly safe layer.

The third genuinely narrows the claim. Continuous intake enlarges the surface for a different failure: a spoofed stream, or action taken faster than review permits. Knight Capital's 2012 malfunction sent valid, well-formed exchange orders for forty-five minutes from an obsolete flag on one server, costing roughly $440 million; the messages were institutionally happy and nobody intended them. Bounded, auditable input is a real safety property, and freezing intake is sometimes the correct control, not merely a deficiency to be engineered away. But freezing does not remove misfire; it changes its species, from the spoofed-input misfire to the classic Austinian one — acting on a consent already withdrawn, an order already superseded — which stale intake produces reliably, as coroners' findings on DNACPR orders make plain often enough. The remedy for spoofing is provenance and revisability, not staleness. Both are properties this position supplies, not arguments against occupying it.

What this does and does not establish

The concept establishes that felicity is a live, perishable fact about the present, not a property of wording, and that any system whose output is taken up as an act inherits Austin's conditions whether or not it was designed with them in mind. It establishes that the intake axis — corpus, scene, running record — tracks exactly the resources felicity-checking requires, and that the third position closes the axis because there is no further kind of evidence felicity could ask for.

It does not establish that standing can be earned by observation, that architecture cannot relocate the problem usefully, or that wider intake is safe by default. It does not make continuous ingestion sufficient for right action — only necessary for checking whether an action, once taken, will be happy rather than void.