The Sorites Collapse Principle

A heap has no first grain, and a workflow has no first replacement

One grain of sand is not a heap. Adding one grain to a non-heap cannot make a heap. Run the induction and heaps cannot exist, yet they plainly do. The paradox exposes not the absence of heaps but the absence of a boundary. No grain count marks the change, so any line drawn is arbitrary.

The same structure runs through any single use of AI at work. Spell-check helps. Autocomplete helps more. Draft generation helps a great deal more. Multi-step composition under light review goes further. Autonomous composition with a rubber-stamp approval is replacement in all but name. Nothing breaks the gradient. Assistance and replacement are not separate events with a crossing between them. They are positions on one path of cognitive substitution.

This essay opens Premise Four, the claim that no available preservation route restores the wage-demand circuit, and takes the first route: regulation by category. Any regime that relies on a clean line between assistance and replacement fails as a defence of the circuit, because that line does not exist. That holds for statutes, professional norms, treaties and human-only zones alike.

The gradient runs unbroken through a single legal brief

Take a legal brief. A few years ago a lawyer or a small team wrote it. It can now be produced in stages, each defensible as assistance.

First, the lawyer has AI summarise the case law, reads the summaries, and writes the brief. Nobody calls this replacement. Second, the lawyer gives bullet-point instructions, the AI drafts sections, and the lawyer revises. The output still carries the lawyer’s judgement. Third, a high-level prompt returns a complete brief, which the lawyer edits lightly and files. The lawyer carries legal responsibility and has read every page, but did not author the document in any meaningful sense. The role has moved from author to verifier.

Fourth, an agentic system takes the case file, drafts the brief, assembles the filing package, prepares replies to opposing counsel and reports back. The lawyer signs off at key decisions. The lawyer is in the loop and no longer doing the work. Fifth, the firm runs the system across its caseload. One lawyer supervises many matters, routine cases run end to end, and the lawyer steps in only when the system flags a hard call. The job has become exception handling.

The evidence register’s entry on OpenAI’s GPT-5.6 release in July 2026 records early adopters describing its legal documents and financial models as ready for production without human polish. That is second-hand vendor material, and the register files it as such. It matters here only as a sign that the later stages are being travelled in public, each step sold as better assistance.

Every move between stages is small. No stage marks where the lawyer stopped authoring and began verifying. By the fifth, the work bears little resemblance to the first, yet no single step crossed a visible line.

That is the Sorites collapse. Inside the workflow there is no point at which assistance becomes replacement. The boundary is nowhere and everywhere. The regulator’s question, where to draw the line, has no answer from inside the gradient.

Substitution is not a quantity any instrument can read

The obvious objection is that law settles definitional problems every day. Speed limits, the age of majority and tax brackets are all arbitrary lines. Why not draw one for AI assistance, admit it is arbitrary, and enforce it?

Because those lines sit on quantities of a different kind. Velocity is measurable and a radar gun says which side of the limit a car is on. Age is measurable and moves one way at a known rate. Income is defined over an accounting period and can be audited.

The quantity an assistance rule would need is the degree of cognitive substitution inside a workflow. No instrument reads it. It is a relation between the work, the worker, the model and the task, and it moves continuously with each. Two lawyers using the same model on the same task can sit at different points on the gradient because their baseline competence differs. One lawyer can sit at different points on Monday morning and Friday afternoon.

So this is not the ordinary Sorites case of an arbitrary but workable cut-off. The underlying quantity does not admit clean measurement, and what cannot be cleanly measured cannot be cleanly regulated as a defence of the circuit.

Proxies measure manifestations, not substitution

A regulator can still measure proxies: review time, the share of machine-generated text, output per employee, approval rates, provenance logs, edit distance between draft and final version, audit trails, task allocation records, headcount. These are real numbers that can be enforced.

They serve fraud detection, safety documentation, intellectual property protection and accountability after harm. They do not capture the variable that matters for the wage-demand circuit, which is whether human labour is still economically necessary to production. A worker can edit every paragraph and still be a verifier. A worker can approve every decision and still be ornamental. A firm can hold headcount for a while after the productive need for it has gone. The proxy survives. The circuit does not.

The best public measures share the limitation. The Federal Reserve’s Real-Time Population Survey, as recorded in July 2026, puts generative AI assistance at 1.7 per cent of US work hours, up from 1.4 per cent in late 2024. It is self-reported, and its unit is hours assisted. It cannot say whether an assisted hour sits at the first stage of the legal brief or the fifth, and no survey question could, because the respondent faces the same gradient as the regulator.

This is the within-face form of the formulation that anchors the next essay. Regulation attaches to manifestations. Capital moves through relations. The mismatch is the failure. Inside a workflow, regulation sees only what leaves a trace, and the traces substitution leaves track the appearance of human involvement, which is exactly what a well-designed deployment is built to preserve. The trace becomes the compliance product. The economic reality moves beneath it.

Every category-based regime fails at the line it needs

A rule that prohibits AI replacement of workers cannot say what it prohibits. It can ban AI use outright, which forfeits the productivity gains and will not survive politically. Or it can ban use above some level, which needs the line the gradient denies. Whatever line is chosen becomes the new ceiling, workflows are redesigned to sit just beneath it, and the ceiling drifts upward. The regulator either follows the drift and grows steadily more permissive, or holds the original line while workflows move past it.

A rule that mandates human oversight meets the same problem from the other side. Oversight is not on or off. It is an allocation of attention. One reviewer can check one output a day, ten an hour or ten thousand a shift, and all three satisfy “human oversight” unless the rule fixes an intensity. Once it does, the regulated activity shifts into sampling, escalation, exception handling, peer review, audit trails or managerial sign-off. The human stays formally present while the labour content of the role tends towards nothing.

A rule that licenses particular uses hits the boundary at the level of description. Clinical decision support becomes differential consideration, which becomes case-by-case workflow assistance. The same activity passes in and out of the regulated category depending on how it is framed. Regulating the framing only returns the problem to definition.

Rules that act on outcomes are different. A requirement that firms hold labour costs at a fixed ratio to revenue never has to separate assistance from replacement. It acts on the value flow directly, and the Sorites problem does not defeat it. But it is no longer regulation by category. It is structural intervention, and the question it raises is what such intervention preserves. The Successor System takes that up later in this premise. In brief, a mandated wage share preserves the wage form rather than the wage mechanism, and the difference decides what kind of system results.

The claim is therefore narrow. No category-based instrument that depends on separating assistance from replacement succeeds as a defence of the wage-demand circuit. Documentation duties, liability rules, safety standards, conduct codes and audit requirements all do legitimate work even where the distinction collapses. The thesis does not say regulation is futile.

Law can manage gradients. It cannot preserve productive necessity inside a workflow that is dissolving.

Professional norms drift with the deployment they were meant to bound

Professions are sometimes offered as the answer statute cannot give. Lawyers, doctors and accountants work under codes of conduct that could, in principle, set an acceptable level of AI use and enforce it through licensing, peer review and reputation.

The gradient applies here too, with an added difficulty. Norms are written by bodies made up of practitioners, and practitioners compete. A lawyer who limits AI use to preserve personal authorship competes with one who uses it heavily and delivers equivalent work for less. The first must lower the standard, charge fees clients will not pay, or lose the clients. A body composed of people under that pressure moves towards norms that accommodate AI rather than restrain it.

This is the Multiplayer Prisoner’s Dilemma operating one level down, inside the profession. Norms follow deployment. The Sorites structure hides the shift, because no single rule change announces it. What counts as competent legal work, diagnosis or accounting slides towards AI-assisted production without anyone deciding it.

In May 2026 KPMG announced a global alliance with Anthropic to embed Claude across a workforce of more than 276,000 in audit, tax, legal and advisory work. The announcement comes from the vendor and describes intent, not measured substitution. Its relevance is structural. When a firm that shapes professional practice builds the tool into the software its people and clients use, the implicit standard of competent work moves with it.

At the scale of the circuit, professions adapt to the technology more than they restrain it. Professional rules can delay or shape adoption in high-stakes settings but cannot hold the assistance-replacement line as a defence of mass labour absorption.

Treaties inherit the definition problem with less knowledge of the work

International agreements are sometimes offered as a higher tier of coordination. The Sorites problem applies to them in full. A treaty banning AI replacement of workers in regulated industries cannot define the ban, because no boundary separates augmentation from replacement. A treaty capping use above a capability threshold cannot hold the threshold, because capability grades continuously across model generations and deployments. A treaty reserving human-only zones of employment cannot say what human-only means once the human works with AI tools.

The usual template is arms control, which has a moderate record on discrete objects such as warheads and chemical agents. It fails for AI for the reasons set out in the previous essay: no discrete inputs, no observable defection, no credible enforcement. This essay adds one more. Even if every institutional condition were met, the thing to be governed cannot be defined. A treaty without a definition cannot be enforced. A treaty with one is overtaken by the gradient and gamed within months.

No tier of coordination dissolves the problem, because it belongs to the gradient, not to the level of governance. Higher tiers face the same definition problem with less knowledge of the workflows they govern. They are worse placed to draw the line, not better.

Within the face and across the faces, the regulatory surface closes

This essay has dealt with the within-face problem, the gradient inside any single use of AI. Categorical Recursion is its across-face companion. Even a line drawn inside one face is left behind when the deployed object migrates to a face the regime does not cover. The within-face gradient makes each category ungovernable as a defence of the circuit. Across-face migration makes the choice of category arbitrary. Together they exhaust the regulatory surface, which is the first route Premise Four closes.

What would refute this

The principle would fail if a jurisdiction adopted a category-based rule, such as a human-authorship requirement or an oversight mandate with fixed intensity, and the rule held its line against workflow redesign for several years without being redescribed around or relaxed. The test lies in the canonical signals, not compliance filings. Entry-level hiring in AI-exposed work, the junior-to-senior ratio and the wage share in covered sectors would need to hold or recover against comparable uncovered work. A direct, auditable measure of substitution, rather than of human involvement, would also count against it. Proxies that rise while those signals deteriorate would confirm it.