Speech acts and performativity in legal and regulatory monitoring

Speech acts and performativity in legal and regulatory monitoring

Some sentences describe the world. Others change it. "The gap is three centimetres" is true or false. "This licence is suspended" is neither — it is an act, performed by saying so. J. L. Austin, in the lectures published as How to Do Things with Words, called these performatives and refused to grade them on truth. He graded them on happiness. A performative succeeds or misfires according to felicity conditions: the speaker must hold the office that the act requires, the words must take the form the institution recognises, and the hearer or the record must take the act up. A clerk who says "you are dismissed" without the authority to dismiss has said nothing that changes anyone's employment. The sentence is grammatical. The act is void.

Austin built this against the positivists of his day, who wanted meaningful sentences to be the kind you could verify. Marriage vows, verdicts, bequests and bets are plainly meaningful and plainly not that kind. He needed a category for language that does things rather than reports them, and by the end of the lectures he had generalised the point: most utterances carry some illocutionary force, description included. Searle later systematised the conditions into rules; legal theory took the framework whole, because law is largely made of performatives — statutes enacted, orders issued, notices served, findings entered.

Why this generalises into a lineage claim

A Large Language Model produces the wording of a performative with total fluency and none of its conditions. It can draft "the respondent is granted an extension to 14 March" in the exact register a tribunal would use, but it holds no office, addresses no one in particular, and keeps no record of whether the extension was ever actually granted. It is a corpus, fixed at a cutoff, generating locutions that resemble acts. Resemblance is not felicity.

A Large World Model adds a sensed scene: it can register that a counterparty is present, that a document has just been filed, that a hearing is in session. This supplies uptake — someone or something is there to receive the act — and gives felicity a chance. But the scene closes when attention moves elsewhere, and nothing persists past the episode. It knows what is happening now. It cannot say what was already committed, by whom, on what authority, or whether that authority still holds.

Felicity conditions are not static facts to be looked up once. They are facts about the present state of the world, and they expire: a mandate is revoked, a consent withdrawn, a licence suspended, a precedent overruled. Checking them against a frozen corpus is not checking them at all; checking them against a scene that dissolves is checking them exactly once, at a moment that may not be the moment that matters. What acting felicitously requires is a maintained, sourced, revisable record of who currently holds what standing, over whom, and what has already been done. That is the object a Large Universe Model names: every relevant stream still running, held as belief rather than fact, each claim carrying provenance and a decay rate. Not a further category of evidence beyond world-sensing — the last one, kept open indefinitely.

This is where the ladder ends. Once intake means everything currently in force, sourced and revisable, there is no further class of information a felicity condition could demand. There is only more coverage, longer memory, and stronger warrant on the sources already included.

The domain as the test: what general counsel actually watches

Legal and regulatory monitoring is a good proving ground for this because almost nothing in it is descriptive. A docket entry is a performative — the court has ruled. A rulemaking notice, once final, changes what conduct is lawful, sometimes with a delayed effective date that is itself a scheduled performative. An enforcement action is not a report about a company; it is the exercise of power over it. Case law updates the meaning of statutes already performative in force. The stream a compliance function must watch is not evidence about the world. It is a sequence of acts that alter standing, obligation and permission, arriving continuously and expiring unpredictably.

The characteristic failure follows exactly the shape Austin's framework predicts: a compliance posture built on a rule superseded two quarters ago. Nothing about the posture was ever false when adopted. The training on which frontline staff were briefed, the checklist embedded in an approval workflow, the standard contract clause cleared by outside counsel — all of it was felicitous at the moment it was fixed. What makes it a misfire now is that the rule it answers to has since been amended, and the organisation kept acting on the earlier performative as though it still held office. General counsel, who is answerable for the posture, discovers the gap not by re-reading old policy more carefully but by learning that a regulator's authority moved and nobody's record moved with it.

This is not a description problem. A model that summarises the old rule perfectly is still wrong, because it summarises something that no longer governs. It is a standing problem: whose act is currently in force, and has it been superseded by a later, competent act. That is precisely the question intake-as-everything-still-running is built to answer, and precisely the question a frozen corpus cannot.

Two objections worth taking seriously here

Authority is conferred by institutions, not earned by watching them. A regulator's rule binds because Congress or Parliament said so, not because a machine observed enough dockets. So the fix is legal standing, and no amount of monitoring supplies it.

This is correct, and the record should not blur it. Standing is granted, not inferred, and no volume of docket-watching turns a monitoring system into a rulemaker or a court. But grant that, and the general counsel's problem does not disappear — it sharpens. The compliance function is not trying to become the source of authority; it is trying to establish, continuously, whether the authority it already answers to has changed. Institutions fix who may act. Monitoring answers whether the act you are relying on is still the one in force. Both questions are real, and only the second is amenable to better intake.

Better to separate concerns: let a system describe the regulatory environment, and route the actual compliance decision through a rules engine — deterministic, auditable, fed by a controlled legal database that a human curates on a schedule.

That architecture should usually be preferred, and it is roughly how mature compliance functions already work. But it relocates the requirement rather than discharging it. The curated database still has to be current — someone, or something, is doing the continuous checking that makes "current" possible, and if that checking runs on a quarterly refresh, the two-quarter lag reappears inside the safe layer. And for anything short of a small, enumerable rule set, the felicity conditions are too numerous and change too unevenly for a static ruleset to track — securities exemptions, sanctions lists, sector-specific guidance and circuit splits do not update on the same clock. There the demand for a maintained, sourced, revisable record returns, whatever the rules engine is called.

None of this makes the frozen corpus useless. It is an excellent teacher of the form a valid order takes, and general counsel still needs staff who can draft in that form. The claim is narrower and harder to dismiss: the moment drafting becomes reliance — the moment "the rule permits this" is treated as still true — felicity has to be checked against a record of what currently holds, not against what once held or what a bounded scene happened to show. Legal and regulatory monitoring does not illustrate that claim. It is the setting where the claim either holds under the actual load of dockets, notices and revocations, or it does not.