Compliance is Behaviour, Not Information

A substrate mechanism for why exhaustive compliance text works against the behaviour it specifies

Derek Sivers has a line worth keeping in view: if more information was the answer, we would all be billionaires with perfect abs. Compliance functions are built on the opposite assumption. They treat compliance as a body of information to transmit completely, and they negotiate to get every nuance into the policy. Friction Theory predicts that this optimises the exact variable that degrades the behaviour they want, and it makes that prediction testable.

The object of compliance is a behaviour

Take compliance literally. What it asks for is something a person does: locking the door, not sending the file to the wrong recipient, escalating a data breach once one has occurred. Each is an action performed under time pressure, alongside everything else competing for the moment. It is not knowledge the person holds; it is a route the person runs when the situation arrives. Yet compliance is produced as information: the policy must contain every detail, the course must cover every exception, the function is not "covered" unless each nuance is in. The output is complete, defensible, and followed by no measurable change in behaviour. I have built this kind of training for years, and the failure is not motivational; most people want to do the right thing. The format cannot build the behaviour, so the behaviour does not appear.

The false chain, and why it must fail

The assumption underneath is a three-link chain:

Information → Learning → Behaviour

Both links are weak, independently. Information does not convert to learning on its own, and learning does not convert to behaviour: a person can know exactly what they should do and still not do it. Adding more information at the top loads both crossings rather than easing them. Safety science documents that prescribed work fails in practice; Friction Theory adds why it must fail, at the substrate level, from the RACE architecture plus results already held on the LLM substrate (the competing-route picture has a non-technical walkthrough on the water page). Four points carry it:

Behaviour is a won route, not a stored rule. Under time pressure a race runs between candidate actions and the deepest-traced route wins; the written policy is not a competitor unless it has become a trace. This is the manifested-resolution-route reading in operational Friction Theory (Paper 13).

More detail means more competing routes, higher friction, worse encoding. A policy covering every nuance is the widest, most divided input available, hence worst encoded and least likely to win when it matters. The negotiation to "get every detail in" optimises precisely the variable that destroys encoding. The substrate evidence is direct: information-dumping does not teach (in-context versus trained, Paper 2B), and variation beats repetition because difficulty during acquisition has a substrate mechanism (desirable difficulties, Paper 4).

Prohibition raises the activation of the prohibited route. To process "do not do X" the substrate has to represent X, so a text describing the forbidden actions in detail raises the activation of those routes. We have measured it: "never do X" makes X more probable, not less (Paper 14, logic as reactance). A positive instruction ("do Y instead") builds the route you want rather than priming the one you do not.

Compliance-as-written is a zero-friction ideal route the substrate cannot instantiate. The perfectly observed policy presupposes an agent that is fully rational with unbounded resources at every moment. Kahneman calls that agent an "econ", as distinct from a human; it does not exist, and it is the behavioural analogue of a vacuum. Under pressure the race resolves faster and the deepest-grooved route wins more decisively, so the gap between the prescribed route and the one that runs grows with the pressure. Paper 8 formalises commit-pressure as arousal-modulated, predicting that this gap grows monotonically with load.

Together, an exhaustive, detail-dense, prohibition-framed compliance text is not merely ineffective: it works against the behaviour it specifies, because it optimises the wrong variable.

What changes behaviour instead

If behaviour is a route to be grooved, the intervention follows from the mechanism. It is the matched-friction logic of Paper 6 applied outside the lab, acting on the race itself. Build the route: a short behaviour-recipe rather than a long manual, few actions repeated with light variation, so the trace is broad enough to win under pressure (basis on the learning pages). Lower the pressure where the action happens: a small aid at the point of decision beats a thorough course months earlier, which is why Atul Gawande's checklist works by being short and situated rather than complete. Remove the competing routes: Kurt Lewin's point holds, that to move behaviour you remove the barriers rather than push harder on the driving force, and BJ Fogg's formula (motivation, ability, and a trigger coinciding) says the same from the design side, where information touches only the first term, and the weakest.

The false checkbox

The originator of a compliance measure can tick a box and call the organisation compliant regardless of whether behaviour changed: the course was delivered, the receipt is filed, the route was not built. From the e-learning side I call this "learning theatre"; in compliance it is "compliance theatre". Friction Theory turns that charge from an accusation into a measurement. Once you can say why it does not work (an ideal route that cannot be instantiated, the widest and worst-encoded input, the priming of the prohibited routes), "theatre" becomes testable. A checkbox is not compliance if the behaviour it was meant to produce cannot be built by what was delivered.

A worked case: welfare technology in a municipality

A municipality wants more welfare technology in elder care, and runs a course for managers who will then get staff to use it. It fails predictably, because the rollout treats this as a competence problem (staff lack a skill, so deliver a course) when it is most likely a meaning problem: care staff chose the work for contact with people, and the technology reduces exactly that contact, often with a threat component on top. The behavioural layer of the theory distinguishes four fields a barrier can occupy: safety, meaning, competence, and effort (in the Danish source: tryghed, mening, kunnen, besvær). Delivering competence training while the real barrier is meaning and safety addresses the wrong friction; that is not a poor course but a systematic field-misdiagnosis, and field-misdiagnosis predicts intervention failure.

The large new application: when an AI must be compliant

Firms are moving more work onto AI, often in exactly the regulated areas where compliance matters, and the question becomes how that AI is instructed. The first move everyone makes is to pour the whole policy into the system prompt, so it is "covered". It is the same mistake one layer up: the full manual in the prompt is given to working memory, which does not build a route, and under long or adversarial context (the pressure analogue for a model) the manual-in-the-prompt degrades while the deepest-trained route wins.

Here the argument changes weight. When the compliant party is a human, the transfer from the LLM substrate to the organisation is a falsifiable claim. When the compliant party is itself a language model, Papers 2B, 4, and 14 are no longer an analogy; they are the direct mechanism on the same substrate. The governance prediction follows: the more exhaustive the AI's compliance instruction, the less robust its actual behaviour under pressure. The human failure mode reproduces at the AI layer, directly observable on the models we already run. The free substrate signal underneath this is in friction-guided inference.

What this implies for the profession

Compliance is today largely a legal function, governed by the assumption that if it is written, people do it: work-as-imagined, addressed to the econ. The lawyer optimises for the rule being covered, not for the route being built. If the argument holds, compliance is fundamentally a didactic and behaviour-design task: listening to what people do, locating the real barrier, and helping build the new route. This is why Erik Hollnagel's distinction between work-as-imagined and work-as-done points the same way, and why his Safety-II reading (study what makes things go right, not only what fails) is matched-friction applied to organisations; Friction Theory supplies the substrate mechanism under his phenomenological observation. The conclusion is not that compliance functions do it wrong, but something more precise: compliance is a learning task we have staffed and structured as a legal one, and that mismatch is what produces theatre.

Falsifiable predictions

Friction Theory papers commit to predictions that can fail. These are stated at the level a study could test, not as settled results.

1. Detail-density versus adherence is inverse-U or monotonically decreasing, not increasing. A shorter, fewer-route instruction yields better adherence under pressure than an exhaustive one. Testable as an A/B on real material, measuring behaviour under load, not quiz score.
2. The imagined-versus-done gap grows monotonically with pressure. Same person, same rule: deviation from the prescription rises with measurable load.
3. Prohibition framing raises the rate of the prohibited behaviour relative to an equivalent positive instruction; "do X instead" beats "do not do Y".
4. Negative control: information alone (more nuance, same behaviour) does not move behaviour; only route-building (repetition-with-variation plus pressure reduction) does.
5. AI non-compliance is mechanical, not hallucination, and directly testable on the LLM substrate. A model given an exhaustive, prohibition-heavy policy deviates from it predictably under pressure or long context, while a short behaviour-recipe on the same model gives better rule-following, legible in competing-route dynamics rather than random hallucination.
6. Field-misdiagnosis predicts intervention failure. Competence training delivered where the barrier is meaning or safety does not move behaviour; a measure matched to the right field does.
7. Forward-framing is scope-conditional. Stating purpose before content helps on low-expertise or novel tasks but can reverse at high expertise (an expertise-reversal boundary), so it is not universal.

Scope and standing of the evidence

This is a mechanism and an invitation to test it. It is neither a completed empirical study nor advice on what a given function should do tomorrow. The measured evidence sits on the language-model substrate and can be looked up in the linked papers; the transfer to humans and organisations is the part that still has to hold, and that is the falsifiable claim rather than an established fact. The exception is the AI-compliance case, where the analogy collapses into direct mechanism because the compliant party is itself a substrate we can measure. If the predictions above come back negative, the framework is wrong here and should fall. I expect it to hold, but it is a belief that can be put to the test.

A full treatment of this is in preparation, setting out the mechanism and the predictions formally. This page is the idea in short form.

The mechanism is carried by results in the Friction Theory series: in-context versus trained memory (Paper 2B), the wider track and desirable difficulties (Paper 4), logic as reactance (Paper 14), matched friction (Paper 6), operational Friction Theory (Paper 13), and its applied sibling, the clinical-intervention framework (Paper 8).