Between roughly 1600 and 1800, England’s common lands were fenced off and turned into private property. Many people, then and now, call it theft at a societal scale. In a new 14-page arXiv paper, “AI and The Enclosure of the Scientific Commons,” Harvard astronomer P. K. G. Williams argues that AI finally gives someone the tools to do the same thing to shared knowledge: take what no one owns and everyone uses, and convert it into private wealth. That sounds academic until you look at your own plant.

Your quality records, SOPs, deviation history, and supplier know-how form a commons too, and it is already being fenced in by vendors, platforms, and copilots trained on your data. Below, what enclosure looks like inside a manufacturer, how to spot it in a contract, and how to keep ownership of the knowledge your margins depend on.

A Harvard Astrophysicist Just Named What Everyone Feels About AI

Williams’ paper (arXiv:2610.27251, submitted 19 October 2026) makes one claim that should interest anyone running a plant: AI provides “precisely the tools needed to enable a previously-infeasible enclosure of the information commons.” Not new theft. Newly possible theft, at a scale the mechanics never allowed before.

Think about where your engineers actually get answers. ASTM and ISO standards, published research, vendor application notes, trade forums, your own internal wiki. Every one of those was built by people contributing knowledge back, on the assumption the pool stays open. AI changes who captures the value of that pool without changing who paid to build it.

That makes this a procurement question, not a philosophy seminar. Your 2026 contracts decide whether your quality data stays yours.

Enclosure Is Not the Tragedy of the Commons, and the Difference Matters

Most people reach for the wrong metaphor here. The famous worry about shared resources is overuse: too many sheep, not enough grass, everybody loses. Williams points out that this is not what happened to England’s common lands. They disappeared because they were fenced, surveyed, and signed over to private owners between roughly 1600 and 1800.

Knowledge behaves differently from grass. My reading of your SOP library does not leave less of it for your process engineers. Williams calls information commons among humanity’s greatest achievements precisely because they hold up under unlimited use. The danger has never been consumption. It is conversion.

Why overuse arguments mislead leaders about data risk

Plenty of manufacturing executives govern data as if depletion were the threat. Access gets restricted, internal wikis get locked behind role permissions, and knowledge-sharing between sites slows to quarterly review meetings. That protects nothing of value and costs you speed.

Meanwhile the actual exposure goes unmanaged. A vendor ingests your deviation history into a model you do not control, and the pattern knowledge your quality team built over fifteen years now sits inside someone else’s product. Nobody used up your data. Someone fenced it. Check your contracts for training rights and derived-model ownership, not your read permissions.

What ‘no one owns it and everyone shares it’ means for your supply chain knowledge

Williams defines an information commons as an intangible resource that

no one owns and is shared by everyone

Your supplier knowledge fits that description more closely than most leaders admit. Material behavior under specific process conditions, which heat treat partner runs tight, the tribal fixes passed between your supplier quality engineers and theirs. No one holds title to any of it. Everyone in the chain draws on it daily.

That is why it is enclosable. Unowned knowledge has no defender, so the first party to capture it in a model effectively claims it. Write down what your network knows collectively, decide which parts stay inside your walls, and put naming conventions and ownership terms around the rest before a platform does it for you.

How AI Converts Shared Knowledge Into Private Wealth

Extraction: the commons as training input, the output as private product

The first mechanic is simple. A model ingests decades of published method papers, standards commentary, application notes, and forum answers, then sells access to a synthesis of all of it. The contributors get nothing back: no citation, no traffic, no reason to keep contributing. Williams frames this as “turning the commons into private wealth,” and the phrasing is exact.

Inside a plant, the same shape appears when you point a vendor tool at your own records. Your deviation write-ups, CAPA narratives, and tribal fixes become training data for a system you rent rather than own. If the contract does not say who owns the derived model and the embeddings, you have funded somebody else’s product with twenty years of your engineers’ judgment. Check the data clause before you check the demo.

Slop spam: why the signal-to-noise cost lands on your engineers

The second mechanic is degradation. Williams argues AI undermines the social systems that support the commons through what he calls “slop spam,” a flood of plausible, unverified content into journals, forums, peer review, and documentation. Those channels run on volunteer attention. Drown them and they stop functioning, which costs more than any single bad article.

Your team already pays this tax. Search a torque spec or a material compatibility question and you get five confident pages that cite nothing. Supplier datasheets arrive polished and unverifiable. A technician’s five-minute lookup becomes a thirty-minute verification job, and sometimes nobody verifies at all.

The practical defense is boring and it works. Decide which sources count as authoritative (the controlled standard, the signed drawing, the validated test report) and make everything else advisory by default. Require a traceable citation on any AI-assisted answer that enters a quality record. The cost of distinguishing a real answer from a plausible one is now a real line item, so govern it like one.

The Same Enclosure Is Happening Inside Your Own Operation

No fence goes up on a plant floor. The contract does the work instead. Twenty years of deviation reports, CAPA closure notes, inspection images, maintenance logs, and the half-documented tribal knowledge sitting in Teams threads get uploaded into a vendor platform under default terms, and the asset quietly stops being fully yours.

Nobody who wrote those records was thinking about data rights. That is exactly the condition Williams describes: a resource built by many contributors who never imagined it would be converted into someone else’s product.

Five contract clauses that decide who owns your quality knowledge

Read these before signing anything. Most procurement teams check security posture and skip all five.

• Ownership and deletion: Do you retain title to raw and derived data, and can you compel full deletion within a defined window?
• Training rights: Does your data improve a shared model sold to your competitors, and can you opt out without losing functionality?
• Fine-tuned weights: If the vendor tunes a model on your corpus, is that artifact yours, theirs, or jointly held?
• Export format: Can you pull everything out in structured, documented form, including labels, annotations, and embeddings, not just a CSV dump?
• Exit terms: What happens on termination, acquisition, or price increase? How long do you keep access, and at what cost?

If a vendor resists the weights question or the embeddings question, that resistance is your answer. Those two items are where the accumulated value actually lives.

The ROI case for keeping your corpus portable

Portability is leverage on price. A vendor who knows you cannot move your labeled inspection data is a vendor with no reason to hold a renewal flat. Teams that negotiated export rights up front switch platforms in weeks rather than rebuilding a dataset from scratch.

It also protects the compounding asset. Your corpus gets more valuable every quarter as defect classes and supplier failures accumulate. Keeping it portable means that value accrues to your balance sheet, not a vendor’s roadmap.

Building AI That Pays Back Into the Knowledge You Depend On

The manufacturers who come out ahead in 2026 will treat their knowledge base the way they treat tooling: an asset that appreciates when maintained and depreciates when neglected. That means structured, owned, exportable, and continuously fed by the people who generate the signal in the first place. A model trained on your data is a snapshot. The underlying record is the thing with compounding value.

Four moves make the difference. Inventory what your internal commons actually contains and where it lives. Build contribution incentives so engineers keep adding context instead of treating documentation as overhead. Require portability clauses in every AI procurement, including the right to export embeddings and labelled data in a usable format. And measure value in hours returned and defect escape rates, not licence seats, because seat counts tell you nothing about whether the knowledge got better.

A 90-day knowledge-ownership audit for operations leaders

First 30 days: map it. List every repository holding quality, process, and supplier knowledge, name an owner for each, and flag which ones are already sitting inside a third-party platform. Most teams find two or three systems nobody formally owns. Those are the ones that leak.

Days 30 to 60: fix the contracts. Read the data clauses in your existing AI and MES agreements, and write exit terms into the next one before signing. If a vendor cannot describe how you get your structured data back on day one of a termination, that is your answer. Days 60 to 90: pick one high-friction workflow, instrument it, and baseline the metrics you intend to improve.

Williams argues this is not just a smash-and-grab but “an attack upon a whole way of life.” Inside a plant, that translates to a straightforward question about which way of working survives: one where your people keep building the knowledge, or one where they rent access to a version of it. Manufacturers still get to choose, and the choice is made in procurement terms, not in principle.

Source: arxiv.org