Thesis
Taylorism is built on a single structural move: separate the thinking from the doing. IP work is a craft in which the two cannot be separated without destroying the value. AI patent drafting recreates the Taylorist separation under a new name. The quality it destroys is invisible at the point of destruction and stays invisible for years. By the time the damage surfaces, it can no longer be traced to the decision that caused it, and in many cases can no longer be repaired at all.
My argument runs in this chain. IP work is craft. Taylorist decomposition fails when applied to a craft. AI drafting is Taylorism wearing new clothes. The thing AI cannot supply is experiential, tacit knowledge, which by its nature cannot be extracted into a corpus that the AI is trained on.
Even if it could, patent applications are non-ergodic objects, and ensemble-averaging tools are structurally mismatched to non-ergodic tasks regardless of how good the averages get.
The patent system’s quality signals and the economics of professional services actively hides the resulting degradation rather than correcting it.
The final turn is damage to the craft. Skilled practitioners are turned into reviewers.
The Taylorist frame, and Mintzberg
Frederick Taylor’s scientific management worked by putting the thinking in a planning department and handing the doing to executors who follow the plan. Frederick Taylor is the bloke from the early 1900s who brought efficiency, standardisation of best practices, timing how long you take a shit for, and all those delightful scientific management principles. Taylor made most of his income from licensing his patents so may have found this essay interesting.
Taylorism is genuinely effective where the work decomposes neatly into a duality of thinking and doing where no feedback is required. Tightening a bolt to a specified torque is separable from the engineering that determined what the torque should be. The action is pure execution of a prior decision. Decompose, standardise, measure, reduce variance: in the right setting it works. The success of Taylorism in those narrow domains has led to Taylorism being considered as suitable for everything.
The early 1900s saw Taylorism transform the factory floor, mechanical and linear systems into highly efficient machines. Naturally such a batch of successful ideas was transplanted from the factory floor to other realms irrespective of its applicability. Taylorism was imported into the executive suite as “strategic planning” where it promptly failed.
I don’t think that Taylorism can be equally applied to knowledge work as it can to the factory floor. Knowledge work does not decompose in a linear manner because it is not a mechanical and linear system. You can’t separate knowledge work into thinking and doing; it’s knowledge work, thinking is doing and doing is thinking!
Mintzberg’s The Fall and Rise of Strategic Planning (1994) Identified this in one of the early areas where Taylorism was unsuccessfully applied: strategic thinking. His diagnosis is that strategic planning is not strategic thinking but they got muddled up. Strategic thinking is synthesis, intuition, the integration of soft and hard information into a vision, where the magic lies. Once this is done, then the strategic planning could begin. Strategic planning is analysis, the breaking-down of a goal into formalised steps. These are two different things. You could arguably use Taylorism for planning. If you try applying Taylorism to strategic thinking It’s not going to work. Strategic thinking needs the space that Taylorism is designed to remove.
The grand fallacy of strategic planning, he argues, is the assumption that because analysis can break a strategy into steps, analysis can produce a strategy. It can’t. Synthesis cannot be formalised and handed down. Strategic planning was to the boardroom what Taylor’s work-study methods were to the factory floor. They are a way to circumvent human idiosyncrasy by capturing knowledge into administrative systems so that thinkers could be separated from doers.
I like the irony that Mintzberg calls out , strategic planners actually missed Taylor’s own most important lesson. Work processes must be fully understood before they can be formalised. The planners formalised before understanding. They did not take the time to understand that thinking and planning around strategic goals are not the same thing.
As we’ll see, IP work is in danger of making the same mistake, systematising the prosecution workflow while never properly mapping the strategic judgement layer that should sit above it. I wonder will this have knock-on effects to the overall profession? To my mind, a good patent attorney builds in strategic thinking to their patent drafting. If you’re not doing both the thinking and the doing will is there any space for the thinking to happen. Do you build an adequate mental schema of the invention and the milieu it fits within?
Why IP work is craft, not engineering
I think I should concede that some IP work is mechanical, surely patent prosecution is largely procedural, and only high strategy is craft. This is the first thing to turn over in my mind.
There is of course a fairly mechanical formalities layer, you need to fill in the forms and record the deadlines. This is fine for Taylorism. It is important but mechanical. It is not why someone pays an attorney to handle their case. But that is not what I want to look at here. I want to look at the craft of patent work. A definition would be a good place to start.
An attorney I worked with once called patent work “word engineering”. I liked the term at the time but now I disagree that engineering is the correct term. My thesis is that it is instead “word craft”.
So what is the difference? I am going to define them in terms of a knowledge transfer.
Engineering is a realm where knowledge can be passed with high fidelity to another practitioner through external artifacts, the CAD drawing, the design document. An engineer has a document as a terminal output that the manufacturing technician can use to make the tangible item again and again. The thinking and the doing become separated. If there is an issue the technician can raise a ticket to get the engineer to think again. I am happy flying on an Airbus A320 because it is the product of high quality engineering.
Craft is a realm where knowledge travels with experience and cannot be formalised into a design document. Craft is tacit knowledge (Polanyi). This definition is more phronesis than techne. The thinking and the doing are the same thing. If there is an issue the craftsman deals with it based on their craft. When the apprentice spots an issue he learns the solution from the master by working with them, not by the master giving them an instruction manual.
In engineering knowledge travels outside the doing and in a craft knowledge travels inside the doing. The knowledge of patent prosecution travels inside the doing of patent prosecution. If it was something that you could write a manual for then such a manual would exist. I am not aware of any such manual but if someone writes one I would love a copy!
This needs to be teased out. Why not just write down how to write a patent.
The claims are the heart of the patent. They are what defines your protection. They define what you get to call yours, your legal monopoly. How does a patent attorney write these? They have to understand the data to know the bounds of the invention, they should know what is commercially relevant to the client, they should know what it likely to get in the way of the clients competitors, they should know the law, around the world, to know what is acceptable to the different patent offices. To keep this interesting these things do not typically align. The client wants to protect something that they know will work but lack the evidence for, and it’s borderline unpatentable subject matter in two of their main markets.
Then the description, the meat of the patent. This is a bit more mechanical because at some level you have done the strategic thinking when drafting the claims. However you are in a feedback loop, as you write the description you realise that the claims can be read in a way that narrows the scope so you need to tweak the claims.
The difficulty is that there is no right answer about what to do. There are trade-offs. A constant judgement on where to put there thumb on the scales. The mechanics of writing a patent may be mechanical but the layers of judgement is a craft.
OK, I am happy to defend the position that you can’t write down how to draft a patent. It is a craft, a tacit knowledge.
The patent attorney doing prosecution well is doing something far closer to Mintzberg’s synthesis than to Taylor’s execution. They integrate technical knowledge, legal doctrine, examiner psychology, and commercial context into a position that no defined sequence of steps would produce.
AI drafting recreates the separation — and there is no mechanical layer to separate
So where does AI patent work fit into the craft. I am not a Luddite burning a loom, although I do think that history has not been kind to the Luddites. I think AI will find a place because it is a tool, and, deep down, humans are tool loving monkeys. I am using AI to mean LLM’s mainly because I cannot be arsed typing LLM’s repeatedly. For the first time ever a non-human can use human language as a tool. How can a curious monkey resist!
AI can reproduce the form of prosecution-quality text. The claims look like claims, the arguments that look like arguments, the descriptions that conform to convention. This is pattern completion over a vast corpus of prior work. It has read all the patents, it knows the shape and can pull it from it’s tensors, those dense mathematical structures that they build.
The form is not where the value lives though. The value in a claim is the boundary it draws. Drawing that boundary requires understanding what is being protected, why this boundary and not a wider or narrower one, how an examiner and later a court will read it, and what a competitor will do to get around it. That chain of reasoning is not recoverable from the finished text, because the reasoning is not in the text.
The AI trained on outputs that do not contain the signal that is required to recreate the output. It cannot learn the judgement that produced them, because the judgement was never written down.
AI is reliable precisely where reliability barely matters, producing text that looks like a patent, and unreliable precisely where it matters most, the boundary decisions that determine whether the patent is worth anything.
The reason there’s no safe “mechanical layer” for AI to own is that, unlike manufacturing, IP work has no real executory stratum beneath the judgement. The genuinely mechanical slice (formalities) is small and the rest only looks decomposable.
Why AI cannot supply the missing layer — the tacit-knowledge argument
The hopeful reply is that AI will get better and eventually supply the judgement too. This misunderstands the kind of knowledge involved.
Experiential reasoning is not pattern recognition at scale — which is what AI does superbly. It is the accumulation of situated judgements whose consequences were lived through. Fighting an examiner for two years over a single word, or having your claims narrowed until the point they are useless carries that experience as a recalibration of judgement. It changes how you read the next problem. It is not a rule; it is a tuning of intuition that exists only because the person lived the consequences of prior judgements.
AI has no consequences. It produces output and never learns whether the output was good from the downstream result. A claim that looked elegant at drafting but failed in enforcement generates no corrective signal in the model that produced it. The feedback loop that makes experiential reasoning possible simply does not exist.
The AI zealots will say that it’s an easy fix because it’s what they always say. More data, the world would be perfect if only we had more data. We can close the loop, train on litigation outcomes, enforcement histories, prosecution records paired with commercial results. It doesn’t rescue the position, for a number of reasons. First, the outcomes are extraordinarily delayed and noisy: a patent drafted today may not be litigated for fifteen years, in a landscape that didn’t exist at drafting, against a product nobody anticipated. The signal is too temporally distant and contextually transformed to teach what the drafter should have done. But that faces us all, the attorney also has the same delay. Second, and deeper, the protection of patents is silent, the signal is that the competitor never launched a competing product or the competitor spent an extra year on R&D to work around the patent. The world never sees these signals. How do you train on an invisible feedback loop. Thirdly, the signal of success that you would train on, the application is granted, doesn’t carry the information whether the strategic patent goal was achieved, yes it granted but the scope was narrow, or doesn’t read onto the competitor anymore, and there is no direct causation between grant and commercial success.
Finally, even in a perfectly closed loop with large enough data the AI sees only a statistical correlation between textual features and outcomes across a population. That is not what the experienced practitioner has. They don’t know that narrow claiming tends to lose commercially because they’ve seen the correlation in a dataset; they remember the specific case where narrow claiming cost them coverage, and they understand the mechanism — the claim read on the embodiment rather than the function, so when the competitor changed the embodiment they fell outside it. Causal, situated, mechanistic understanding. Correlation does not get you there.
This is Polanyi. Michael Polanyi — Hungarian-British chemist turned philosopher, whose philosophy of knowledge grew out of watching how science actually works — argued that “we know more than we can tell.” The knowledge that lets a skilled practitioner perform exceeds anything they can articulate about what they’re doing: the cyclist who can’t explain how they balance, the surgeon who can’t fully say how they perform an operation, the blind man with a cane who can navigate the world. Polanyi called this tacit knowledge, distinct from explicit knowledge, which can be written down and transferred. Engineering versus craft. His deeper point is that explicit knowledge always rests on a tacit foundation. To create the engineering documents the engineer will apply their craft, determine the best torque for the bolt, but once then they can document their thinking and hand it off the the doers. Polanyi was equally clear on transmission: tacit knowledge, craft, moves not through instruction but through proximity to practice, master to apprentice, by absorbing corrections and accumulating one’s own situated experience.
The consequence is that the floor on craft judgement in IP work is permanent, not a temporary limitation that scale will erase. AI will keep improving on the explicit layer, form, structure, prior-art retrieval, consistency, procedural compliance. The tacit layer stays out of reach regardless of model capability, because the problem isn’t capability; it’s that tacit knowledge by definition isn’t in the corpus.
So the craft can’t be pulled out of the practitioner and into the model. The obvious reply is to leave it in the practitioner — let AI draft, and keep the human judgement in the loop. That escape only holds if reviewing a draft builds the same knowledge as writing one. It doesn’t.
Why review is not doing — the schema-building argument
The strongest version of the “AI as tool” defence is worth stating at full strength: the practitioner uses AI for a first draft, then brings full craft judgement to interrogate, reshape, and own it. Thinking and doing stay integrated; the tool has merely cleared the mechanical work. If that holds, my whole argument collapses. Let’s turn there next.
A patent is a business instrument, and a good practitioner drafts it against a live mental schema: the previous filings in the portfolio, the competitor’s last three applications, the examiner’s tendencies in that art unit, the R&D pipeline, the commercial roadmap the patent is meant to serve. A myriad of competing factors that pull in different directions. What the patent office wants conflicts with what the client wants and at the same time the prior art is pushing one way while the competitors filings pull another way. There is no right answer, or more correctly there are a myriad of right answers depending on who is asking what question.
The AI bro’s will say just capture all this into the AI’s 1 million token context, capture the client’s roadmap, the prior art, the commercial strategy, the previous patent prosecution histories in the art area, the engineer’s thoughts, give it all to the AI. I think this is fucking mental.
Imagine asking a client to write all this down before you write a patent. Even if it was possible, and remember there is no canonical knowledge here, no right answer, they will just move the work to non-crazy patent attorney who will just draft the goddamn patent.
But let’s say you have a crazy client who is happy to work with the crazy attorney. It’s always good to test edge cases.
I think the essence I am trying to capture is that drafting is constructive. It forces the practitioner to reach into that schema, pull out the layers that bear on this filing, and assemble them into a document — and the reaching-in is not incidental to the work, it is the work. Each new application, drafted from the blank page, becomes part of the practitioner’s schema because they got up close and personal with it. The knowledge formation happens in the gap between the blank page and the first draft, not between the first draft and the final one. The new knowledge reforms the schema. It’s called legal practice for a reason.
Reviewing is evaluative. It draws on the schema the practitioner already has to assess something already made. A skilled reviewer catches errors, improves phrasing, spots a strategic gap — but they never did the reaching-in, so their schema doesn’t grow. You do not build experiential knowledge by checking another’s work; you build it by doing the work. When the AI fills the blank page, the practitioner is handed a finished object to judge, and the cognitive act through which craft knowledge forms never happens.
This is what turns a contingent market observation into a structural cognitive claim. It no longer matters how conscientious the practitioner is or how the work is priced. Even the diligent expert who reviews AI output carefully will, over time, know less, but not through laziness, but because reviewing doesn’t build what drafting builds.
The tacit-knowledge argument established that craft judgement can’t be pulled out of the practitioner and into the model. The construction point goes deeper: experiential knowledge isn’t even built through second-order engagement with the outputs. You have to do the reaching-in, or the knowledge never forms. But why should doing build the schema when reviewing doesn’t? Something has to make the difference between constructing and evaluating more than cognitive. That something is exposure: the drafter builds judgement because the drafter is on the hook.
AI has no skin in the game
Nassim Taleb wrote about skin in the game. The idea that upside and downside should be borne by the same person. There should be a symmetry in the payoffs.
All practitioners have skin in the game. They have invested years becoming practitioners; even a trainee has skin in the game, they have a STEM degree, a training contract, a career staked on getting good. When a practitioner drafts a claim, they are exposed to what happens to it. If the scope is wrong, it is their name on the file, their client who thinks they are an idiot for drafting it, their reputation and possibly their liability when the patent fails in enforcement. That exposure is not incidental to the judgement, it is what makes the judgement serious. A practitioner reasons hard about the irreversible move because they will personally bear the cost of getting it wrong.
This is Nassim Taleb’s point sharpened for our domain. Skin in the game is not a slogan about fairness; it is an epistemic mechanism. Exposure to the downside is what disciplines judgement, separates the reckless from the careful, and tunes the very intuition the tacit-knowledge section describes. You do not develop craft judgement in the abstract; you develop it because your errors cost you, and the memory of that cost recalibrates the next decision.
AI has no skin in the game. It drafts the claim and bears nothing, no upside, no downside. It has not invested anything getting to this point, the model suffers no loss, feels no pride, carries no memory. This is the same absence the tacit-knowledge argument identified as a broken feedback loop, but the point here cuts deeper than learning. It is that the entity doing the doing is the one entity in the chain wholly insulated from the consequences of the doing. It is a decision-maker with no upside and no downside, an inert agent.
So when the workflow becomes AI drafts, practitioner reviews, the AI does the doing but bears no consequence. The practitioner bears the consequence but didn’t do the doing, and, per the previous section, didn’t build the schema either. What remains is a practitioner reviewing the polished and plausible output of an inert agent. The separation Taylor achieved by putting the thinker upstairs and the doer downstairs, AI achieves by putting the doing in a system that has no stake in whether the doing was any good.
The obvious objection is that the practitioner who signs off does still have skin in the game. Their name goes on the file, the client knows who did the work. The skin in the game isn’t gone, the objection runs; it has simply moved to the review step. But this misunderstands how skin in the game disciplines judgement. Exposure is corrective only when it is coupled to the act that carries the risk. The practitioner who drafts feels the weight of each irreversible move as they make it. The boundaries they’re creating, the scope they’re surrendering, the future that they are trying to protect, and that felt weight is what forces the reasoning to go deep at the exact points where it matters. Exposure at the review step is decoupled from construction. The reviewer is accountable for boundaries they didn’t draw and can’t fully see, because the reasoning that set them was never theirs. That produces anxiety without the corrective information to act on it, liability without the epistemic access that would let liability sharpen judgement. Skin in the game that floats free of the doing doesn’t discipline; it just allocates blame.
And here the feedback loop turns a bad situation into a self-concealing one. If the failure surfaced next week, the misallocated exposure would at least generate a signal the reviewer would learn, painfully, what they missed. Skin in the game was supposed to be the mechanism that makes quality observable to the person producing it, sever exposure from the act, stretch the feedback loop past the point of traceability, and that mechanism is gone. The result is a market in which no one, not the client, not the reviewer, not the model, holds a live, corrective stake in whether the boundary was drawn well. Quality becomes unobservable to everyone who could act on it. The full lemons dynamic this sets in motion is developed below; for now the point is simply that skin in the game, decoupled from the doing, cannot counter it from the inside.
That drift toward the unobservable would be tolerable if the errors it hid were the averaging-out kind, small, recoverable, washed away over an ensemble of patents. They are not. Each patent walks it own path. Part of the skin in the game is that the practitioner is trying to create a path that does not lead to death.
The ergodicity argument
Here is the sharpest way to see the structural mismatch. Statistically, an LLM writes a patent application that is the statistical average of thousands of other applications. That is why they look, at first sight like good applications. They nail the average boilerplate because it is the average. The secret sauce is not in the boilerplate though. The rest of the application should be anything but average. No one ever needs an application that is part of the ensemble — you need your application, this application, protecting your lead asset, to be exactly the one that isn’t average, because the value lives in the divergence.
Borrow the frame from ergodicity economics (Peters, 20191). A system is ergodic when the time average for a single trajectory equals the ensemble average across many trajectories. When what happens the average across the population tells you what happens, on average, to any individual over time. In a non-ergodic realm the ensemble average does not equal the time average. When you follow an individual trajectory across time there will be points of ruin, points of no return that are hidden when you look as the ensemble average.
Think of the difference between 1000 people running through a minefield and one person running through a minefield 1000 times,. Let’s say that there is a 50/50 chance of getting to the other side. The ensemble average is when 500 from the 1000 people make it across safely.
The time average is a lone runner, their next run is dependent on surviving the previous run. Each time they run matters, and the order that they run in matters. Once a run goes wrong they never run again. Their odds of surviving over the 1000 runs is 0.5 to the power of 1000 or essentially zero.
Running across the minefield is non-ergodic because the ensemble average (50%) does not equal the time average (0%). Looking at the ensemble average tells you nothing about the fate of an individual over time. Your path matters, the order of events matter, and you face ruin along the way.
Patents are non-ergodic. A claim drafted too narrow can’t be broadened after grant. Scope surrendered through prosecution-history estoppel can’t be unsurrendered. You lock in during drafting and then you run the application across the minefields of prosecution, and litigation, and opposition, and due diligence. Even worse, you run it across minefields where there is another patent attorney actively putting mines in your path.
These are not temporary deviations from a mean that wash out over time. they are permanent path dependencies that lock in consequences. The individual trajectory is all that ever exists in practice, and it can diverge sharply from the ensemble precisely where it matters. A claim drafted to look like most claims will perform like most claims — and most claims are mediocre, because that’s what an average is.
In an ergodic world you can safely learn from the population and apply it to the individual. LLM training optimises toward the ensemble average: the central tendency of what patent applications are. That is exactly the wrong target for a non-ergodic object.
The experienced practitioner is not trying to hit the ensemble average. They are reasoning about the specific path this application is on and where the irreversible moves are, which scope locks in a defensible position regardless of how prosecution unfolds, which wording is going to hold up as it navigates a myriad of minefields. That is trajectory reasoning about a single non-repeating path through a space where some moves are permanent. AI have no concept of irreversibility; they emit the next token with no model of which textual choices close off unrecoverable futures.
For the AI the training corpus systematically excludes the very information trajectory reasoning needs: it records what was filed, never what was lost by filing it that way, the roads not taken are invisible. So the model is optimising on the ensemble, blind to individual trajectories, and trained on data that omits the signal that would make trajectory reasoning possible. These are not separate problems. They are one problem seen at three levels: ensemble-blindness, irreversibility-blindness, and a corpus that hides the counterfactuals.
This is also what disarms the strongest counterargument, that an expert using AI will outperform an expert working unaided, because AI clears the mechanical layer and frees time for judgement. Granted at the ensemble level, and probably true on average for the best practitioners in the short run. But it is irrelevant at the individual level, which is the only level that exists for a non-ergodic object. The practitioner whose path runs through one AI-assisted error on a critical application doesn’t get to average that away across their other matters; the client whose patent fails in enforcement is not compensated by the fact that AI-assisted prosecution produces better average outcomes across the market. Non-ergodicity means the variance is what matters, and AI does not reduce variance on the cases where variance matters most, the high-stakes, novel, boundary-pushing applications where craft judgement is most critical and ensemble statistics least informative. The counterargument doesn’t just fail; it misdescribes the nature of the problem, which is itself a restatement of the thesis.
A cost this high and this irreversible should be exactly what a market prices out. It isn’t, and the reason is that the cost is invisible at the only moment anyone could act on it.
Why IP is specifically exposed (and not just any profession)
Most professional domains carry one or two of the relevant features. Surgery has irreversibility but immediate feedback — you know quickly whether it worked. Architecture has temporal delay and physical embodiment — the building stands or falls partly on its own terms, somewhat independent of the drawing. Financial advice has delay but the position can often be unwound.
IP has all three. Temporal delay between decision and consequence or even a complete lack of consequence. Formal irreversibility of certain choices. The document is the entire asset — there is no physical artefact carrying part of the value, as a building does for an architect. This trifecta defeats the feedback loop that would normally discipline quality. The irreversibility means that if and when feedback finally arrives, in litigation, in diligence, in a competitor’s clean design-around, there is nothing to be done about the original choice. There’s no middle ground where a slightly weak claim still partly protects because the underlying invention is real: the claim either covers what matters or it doesn’t. This combination makes IP uniquely resistant to the ordinary market mechanisms that discipline quality in professional services, where poor work becomes visible and costs the practitioner clients. In IP any signal of quality is delayed or maybe non-existent, expensive to obtain, and hard to trace to its cause that the disciplining loop barely operates. The IP market can sustain poor quality at scale.
The lemons problem — why the market selects the degradation in
There is an information structure underneath everything described so far that has its own name. It is Akerlof’s market for lemons.
In Akerlof’s original formulation, when buyers cannot distinguish a good used car from a bad one at the point of sale, they rationally discount what they’re willing to pay to the average. That discount drives good cars out of the market — sellers of good cars can’t get a price that reflects their quality, so they withdraw. What remains is lemons. Quality is driven out not by preference but by unobservability: the market cannot price what it cannot see.
IP services have the same structure, sharpened by the temporal features described above. A client cannot distinguish a well-drafted patent from a poorly-drafted one at the point of delivery. They look identical to a non-expert, the claims are formatted the same way, descriptions of similar length, both stamped “granted” two to four years later. Think of the patents that fall because a stray comma changes their scope. The feedback that would reveal the difference is years away and expensive to obtain, surfacing only in litigation, diligence, or a competitor’s clean design-around. By the time the signal arrives, the original drafting decision is buried under layers of prosecution history, amended claims, and changed commercial context. A path full of micro decisions that make it impossible to trace the outcome to its source.
The lemons dynamic means market pressure runs toward the cheaper option not despite the quality gap but because it is invisible. When AI makes drafting faster and cheaper, the rational buyer, unable to observe quality differences, selects on price. The rational provider, unable to signal quality credibly, competes on price. The virtuous practitioner who drafts carefully and charges for the time cannot demonstrate to the market that their patent is better than the AI-assisted shallow draft, because “better” won’t be observable for a decade and may never be tested at all. The market selects the cheap gap in, exactly as Akerlof predicted.
This is the structural reason the degradation described in earlier sections does not self-correct. Ergodicity showed that the losses are ruinous rather than averageable. The lemons structure shows why ruinous losses do not generate the corrective feedback that would normally drive them out of a functioning market. The information never reaches the buyer in time to inform the next purchase.
You would expect the institution built to certify patents to supply the missing quality signal. It supplies one, grant, and that signal is precisely the problem.
The patent system subsidises the degradation
The system’s primary quality signal, grant, is the wrong signal for the thing that matters. A granted patent signals that the application survived examination: that it was formally compliant and the examiner found no disqualifying prior art in the time available. It says nothing about whether the claims read on the actual product, whether a competitor can design around them cheaply, whether they’d survive inter partes review, or whether they map to any commercial objective the company actually holds. Examination is under heavy resource pressure; offices process enormous volumes; the prior-art search is necessarily incomplete. So “granted” increasingly means little more than “formally compliant and not obvious or anticipated in the hours an examiner could spend.”
The result is that the system inadvertently creates a protected space for quality degradation. The cheap, fast route, file something, get it granted, report the portfolio size, is indistinguishable from the expensive, slow route where a practitioner thinks hard, drafts strategically, builds a portfolio that can deal with the minefields, right up until enforcement or challenge, which may never come.
There’s a historical mechanism underneath this. The patent system was designed around the assumption that a granted claim represented a meaningful moving forward of the art, a reasonable proxy for value in an era when filing was expensive and slow, and therefore self-selecting for serious inventions. The cost barrier did much of the filtering. When filing becomes cheap and fast, first through process efficiency, now through AI, that self-selection mechanism disappears, but the system hasn’t adapted its quality signals to compensate. The barrier that used to do the filtering is gone, and grant alone can’t tell the difference.
So nothing in the market and nothing in the institution stops the degradation while it is happening. So where does that leave us.
Where this leaves us — and what’s deliberately left open
The diagnosis is that Taylorist management logic and AI drafting fail at the same point: the separation of thinking from doing, in work where they cannot be separated.
The danger is that failure is structurally hidden by temporal delay, by irreversibility, by the patent system’s weak quality signals, and by the lemons structure of the market. The craft will degrade without generating the corrective feedback that would normally stop it.
Taylorism is pushing on 100 years old. Can we look back at history to see where rhymes, to see what Taylorism broke and how it was fixed? That feels like another topic. This essay is the diagnosis only.
Related
Footnotes
-
Peters, O. The ergodicity problem in economics. Nat. Phys. 15, 1216–1221 (2019). ↩