AI Has a Supply Chain Problem. The Supply Is Human Work.

An internal warning at Microsoft has become an unusually blunt way to describe one of the anxieties surrounding artificial intelligence.

A publishers’ legal brief made public on September 17, 2026 quotes Microsoft director of applied science Brent Hecht describing the possibility of the “largest theft of labor in human history.” The filing also quotes his concern that people would see models absorbing their work as theft on an unprecedented scale.

The brief comes from news organizations including The New York Times in their copyright litigation against Microsoft and OpenAI. It is the plaintiffs’ argument for summary judgment, not a ruling that the companies committed theft.

Microsoft told Ars Technica that Hecht’s comments represented an individual perspective, not the company’s views or a legal analysis. It defended its products as transformative fair use and disputed that they substitute for news sites. OpenAI also publicly argues that training on available internet material is fair use.

That disagreement matters. A striking internal phrase does not settle a copyright case.

But a separate passage in the brief points to an economic problem worth examining on its own. It quotes a Microsoft document warning of a “doom loop”: AI products could undermine the financial foundations of the content suppliers they depend on.

The workplace question underneath that warning is uncomfortable: what happens when a technology becomes more useful by drawing on work whose future production it may make harder to sustain?

Start with publishing. Reporting requires money, time, access and judgment. An answer assembled from that reporting may be convenient enough that a reader never visits the organization that paid to produce it.

If that substitution reduces revenue without an offsetting source of income, the publisher has fewer resources for the next investigation. Better access to existing information can coexist with worse incentives to create new information.

That is a risk to examine, not a prediction that every publisher will fail or that every AI answer costs someone a job. Different products, licensing arrangements and reader habits can produce different outcomes.

The same tension offers a useful way to think about a much smaller setting: the office.

Imagine a company asking an experienced employee to document how they resolve unusual customer problems. The guide becomes searchable through an internal AI assistant. Colleagues can get an answer without waiting for that employee to become available.

From the company’s perspective, this can be sensible. Knowledge that only one person holds can disappear when that person leaves, takes vacation or gets pulled onto another project.

From the employee’s perspective, something else has happened. Expertise that once made the organization depend on them is now easier for the organization to distribute.

There is an important technical distinction here. Putting documents into an AI assistant does not necessarily train its underlying model. A system can retrieve relevant documents and use them while answering a question, without retraining the model. Microsoft’s documentation describes that approach as retrieval-augmented generation.

So “you are training your replacement” is too sweeping a description of ordinary workplace AI use. Documenting a process, supplying context to an assistant and training a model are different activities.

The management question survives that distinction. If a system makes an employee’s knowledge available more widely, how does the organization reward the person who helped make that possible?

Knowledge sharing has always been part of work. Training a colleague and writing a good manual can also make a team less dependent on one person. AI adds another way to reuse that material, but it does not erase the need to decide what the contributor gets in return.

A company could recognize that employee’s contribution, expand their responsibilities or give them more time for difficult cases. It could also decide it needs fewer people doing the original task. Those are organizational choices, not features selected automatically by the software.

Consider a deliberately simplified example. Ten employees gain 25 percent more capacity with an AI tool. At the same quality and working hours, they could now produce the output that previously required twelve and a half people.

Management could use the additional capacity to produce more, improve service or reduce working hours. If the work divides neatly and the gain holds across every task, eight people could instead produce the old team’s output.

Real jobs rarely divide that neatly. Reviews, handoffs and unexpected problems still consume time. The example is an illustration of the choice, not a forecast or a measured productivity result.

The software may create additional capacity. Management decides what to do with it.

This is where the employee’s incentive becomes important. Asking people to share their best methods is easier when they expect to benefit from the result. If they expect only a higher target or a smaller team, the request can sound less like collaboration and more like an audition for their own redundancy.

That does not mean workers should hoard knowledge. It means organizations should be able to explain the bargain they are offering.

What happens to the time saved? How is the person who documents the process recognized? Who remains responsible for checking an answer when the assistant encounters a case the guide never anticipated?

The publishing dispute and the workplace example involve different rights, relationships and technologies. A copyright claim about news articles does not establish what will happen to an employee who writes an internal manual.

They do, however, point toward a shared economic question: how do you keep people producing valuable knowledge when your product makes that knowledge easier to reuse?

For a newsroom, the answer may involve subscriptions, licensing or other ways to fund original reporting. For an employer, it may involve pay, promotion, staffing decisions and a credible commitment about how productivity gains will be used.

Calling the result efficiency does not answer who receives the benefit.

We keep asking what AI can produce. The harder management question is what keeps people willing and able to produce the work it draws on.