4. The Sorites Collapse Principle

The ancient paradox of the heap goes like this. One grain of sand is not a heap. Two grains are not a heap. Adding a single grain to a non-heap cannot make it a heap. Therefore, by induction, no number of grains constitutes a heap. The paradox is not that heaps do not exist. They obviously do. The paradox is that no specific grain count separates heap from non-heap, which means the boundary is undefined, which means any line we draw is arbitrary.

The paradox has a serious economic application. Inside any single use of AI, there is a continuous gradient from minimal assistance to near-total replacement. Spell-check is assistance. Autocomplete is more assistance. Draft generation is much more assistance. Multi-step composition with light human review is more still. Autonomous composition with rubber-stamp approval is replacement in all but name. The gradient is unbroken. There is no specific point at which assistance becomes replacement, because the categories are not discrete events. They are stages on a single trajectory of cognitive substitution.

This is the Sorites problem applied to AI deployment, and it is the structural reason that no regulatory boundary, no treaty, no professional norm, and no human-only zone can hold against the gradient as a defence of the wage-demand circuit. Any regime that depends on a clean line between assistance and replacement fails because no such line exists. This essay establishes the principle and traces its implications.

The within-face gradient

Consider a single professional task: writing a legal brief. Five years ago, this task was performed by a lawyer or a team of lawyers, with research support and possibly a paralegal drafting initial sections. Today, the same task can be performed in a sequence of stages, each of which is technically defensible as a form of assistance.

Stage one: the lawyer uses AI to summarise relevant case law. The AI reads cases and produces summaries. The lawyer reads the summaries and writes the brief. This is assistance. Nobody objects. The lawyer is still doing the cognitive work.

Stage two: the lawyer uses AI to draft individual sections of the brief based on bullet-point instructions. The AI produces draft text. The lawyer revises. This is also assistance. The lawyer is still in control. The output reflects the lawyer’s judgement.

Stage three: the lawyer uses AI to draft the entire brief based on a high-level prompt. The AI produces a complete brief. The lawyer reviews it, makes some edits, and submits. This is harder to characterise. The lawyer is still legally responsible. The lawyer reviewed the document. But the document was not authored by the lawyer in any meaningful sense. The lawyer is functioning as a verifier rather than an author.

Stage four: the lawyer uses an agentic system that takes the case file, drafts the brief, prepares the filing package, drafts responses to opposing counsel’s filings, and updates the lawyer on developments. The lawyer reviews periodically and approves at key decision points. The lawyer is still in the loop. The lawyer is no longer doing the work.

Stage five: the firm deploys the agentic system at scale. One lawyer supervises the output of the system across many cases. The system handles routine matters end to end. The lawyer intervenes only when the system flags a complex decision. The lawyer’s role is now exception handling. The system does the work.

Each transition between stages is small. Each stage is defensible as a continuation of the previous one. There is no specific stage at which the lawyer stopped being the author and started being the verifier, because the transition is gradual. By stage five, the lawyer is doing fundamentally different work than at stage one. The transition happened, but no individual step crossed a clear line.

This is the Sorites collapse. There is no point inside the workflow at which assistance becomes replacement. The boundary exists nowhere and everywhere. The regulatory question of where to draw the line cannot be answered from inside the gradient.

Why the gradient is not just a definitional problem

A critic might respond that this is merely a definitional problem, of the kind law solves all the time. Speed limits draw arbitrary lines. Age of majority draws arbitrary lines. Tax brackets draw arbitrary lines. Why can AI assistance not be regulated by drawing a similar arbitrary line, accepting that the line is arbitrary, and enforcing it?

This response misunderstands the structure of the problem. Speed limits work because the underlying quantity (vehicle velocity) is measurable, discrete, and relatively stable. A vehicle is going either above or below the limit, and which side it is on can be measured by a radar gun. Age of majority works because age is measurable, discrete, and changes only by aging. Tax brackets work because income is measurable in a defined accounting period.

The quantity that would have to be measured to regulate AI assistance is not measurable, not discrete, and not stable. It is the degree of cognitive substitution within a workflow. This is not a number that can be read off any instrument. It is a relational property of the work, the worker, the AI, and the task, which varies continuously across uses, across tasks, across workers, across time of day, across deadline pressure, across the worker’s expertise level. Two lawyers using the same AI for the same task may have different degrees of substitution because they have different baseline competencies. The same lawyer using the same AI for the same task may have different degrees of substitution on Monday morning versus Friday afternoon.

The Sorites problem with cognitive substitution is not the ordinary Sorites problem of arbitrary boundaries. It is a deeper problem in which the underlying quantity is not the kind of thing that admits clean measurement, and therefore not the kind of thing that admits clean regulation as circuit defence.

Proxies measure manifestations, not substitution

A regulator can measure proxies. It can measure review time, AI-generated text, output per employee, approval rates, provenance logs, edit distance between AI draft and final version, audit trails, task allocation records, or headcount changes. These proxies are real. They produce numbers. They can be tracked, reported, and enforced.

The proxies are useful for fraud detection, safety documentation, intellectual property protection, and accountability after harm. They do not measure the thing that matters for the wage-demand circuit, which is whether human labour remains 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 maintain headcount temporarily while the productive necessity of that headcount disappears. The proxy survives. The circuit does not.

This is the within-face equivalent 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 can only see what produces a measurable trace. The substitution of human labour by AI produces measurable traces that do not track the substitution itself. They track the appearance of human involvement, which is exactly what the deployed workflow is engineered to preserve. The trace becomes the compliance product. The economic reality moves underneath it.

The implications for regulation as circuit defence

This has direct consequences for any regulatory regime that depends on distinguishing AI assistance from AI replacement to preserve the wage-demand circuit.

Regulatory regimes that prohibit AI replacement of human workers cannot define what they are prohibiting. The regulation either prohibits all AI use (which prevents productivity gains and is politically unworkable) or prohibits some specific level of AI use (which requires drawing a line that the gradient prevents). Any line that is drawn becomes the new ceiling for permissible AI use, with workflows redesigned to sit just below it. The line drifts upward as workflows adapt, and the regulation either follows the drift (becoming reactive and increasingly permissive) or holds the original line (becoming impossible to enforce as workflows evolve past it).

Regulatory regimes that require human oversight of AI decisions face the same problem from the opposite direction. Human oversight is not a binary. It is a bandwidth allocation. A human can review one output per day, ten outputs per hour, or ten thousand outputs per shift. All three satisfy the phrase “human oversight” unless the regulation specifies intensity. Once intensity is specified, the regulated object moves to sampling, escalation, exception handling, peer review, audit trails, or managerial approval. The human remains formally present while the labour content of the human role approaches zero.

Regulatory regimes that license specific uses of AI face the boundary problem at the use definition level. Each use can be redescribed as a different use. AI for clinical decision support becomes AI for differential consideration becomes AI for case-by-case workflow assistance. The same underlying activity moves between regulated and unregulated categories depending on how it is framed. The regulator’s only response is to regulate the framing, which returns the problem to definition.

Regulatory regimes that operate on outcome measures (productivity per worker, hiring patterns, wage levels) face a different kind of problem. They can succeed where they directly mandate the outcome they want to preserve. A regime that requires firms to maintain a specific ratio of labour costs to revenue does not need to distinguish assistance from replacement. It regulates the value flow directly. This is a real category of intervention, and it is not defeated by the Sorites problem. It is, however, no longer category-based regulation. It is structural intervention on the value flow, and it raises the question of what such intervention preserves. The Successor System essay addresses that question directly. The short answer is that mandatory wage-share schemes preserve the wage form, not the wage mechanism, and the difference matters for what kind of system results.

There is no category-based regulatory instrument that depends on distinguishing assistance from replacement and still succeeds as a defence of the wage-demand circuit. Other regulatory purposes survive the gradient: documentation duties, liability rules, safety standards, professional conduct codes, and audit requirements can all serve legitimate functions even where the assistance-replacement distinction collapses. The thesis is narrow. It says the wage-demand circuit cannot be defended by category-based instruments that depend on the gradient, not that all regulation is futile.

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

The implications for professional norms

The argument is sometimes made that professional norms can do what regulation cannot. Lawyers have professional ethics. Doctors have professional standards. Accountants have professional codes. These norms could, in principle, define what level of AI use is professionally acceptable and enforce that limit through licensing, peer review, and reputational sanction.

The Sorites problem operates here too, with an additional difficulty. Professional norms are typically defined by the profession itself, through its representative bodies. The profession’s representative bodies are made up of practitioners. Practitioners face the same competitive pressure as everyone else. The lawyer who limits AI use to maintain a high standard of personal authorship competes against the lawyer who uses AI extensively and produces equivalent work at lower cost. The high-standard lawyer either lowers their standard, raises their fees beyond what clients will pay, or loses clients to the other lawyer. The professional body, made up of practitioners under this pressure, drifts toward norms that accommodate AI use rather than restrict it.

This is the same Multiplayer Prisoner’s Dilemma operating one level inside the profession. The norms shift to match the deployment, rather than the deployment being constrained by the norms. The Sorites problem makes the shift hard to detect, because no specific norm change is visible. The implicit standard for what counts as competent legal work, or competent medical diagnosis, or competent accounting, drifts gradually toward AI-assisted production. The drift is not a decision. It is the cumulative effect of practitioners adopting AI at the margins to remain competitive.

At circuit scale, professions adapt to the technology more than they constrain it. Real professional rules can delay or shape adoption in specific high-stakes contexts. They cannot hold the assistance-replacement boundary as a defence of mass labour absorption.

What about treaties and international agreements

International treaties are sometimes invoked as a higher-level coordination mechanism that might succeed where national regulation fails. The Sorites problem applies here with full force. A treaty banning AI replacement of human workers in regulated industries cannot define what is being banned, because the boundary between augmentation and replacement does not exist. A treaty regulating AI use above a certain capability threshold cannot maintain the threshold, because capability gradates continuously across model generations and across deployments. A treaty defining human-only zones in employment cannot specify what counts as human-only when the human is using AI tools to perform the work.

The standard model for international technology governance is arms control, which has a moderate track record on discrete physical objects (warheads, missiles, chemical agents). The model fails for AI for the reasons named in the previous essay (no discrete inputs, no observable defection, no credible enforcement) and for the additional reason given here. Even if the institutional conditions for treaty-based governance were satisfied, the substance of what would be governed cannot be defined. The treaty either lacks definition (in which case it is unenforceable) or imposes a definition the gradient invalidates (in which case it is gamed within months).

There is no level of international coordination at which the Sorites problem dissolves. The problem is inherent to the gradient, not to the level of governance. Higher levels of governance face the same definition problem with less detailed knowledge of the workflows being governed. They are worse positioned to draw the line, not better.

The relation to the across-face problem

This essay covers the within-face Sorites problem: the gradient inside any single use of AI. The next essay covers the across-face problem: the migration of AI deployment between legal categories. The two problems compound. Within-face Sorites makes any single category individually ungovernable as circuit defence. Across-face recursion makes the choice of category arbitrary. Together they exhaust the regulatory surface available for preserving the wage-demand circuit.

The Sorites Collapse Principle is the within-face foundation. Categorical Recursion is the across-face foundation. The next essay completes the regulatory closure.