In October 2026, Current Affairs wrote one of publishing’s strictest anti-AI policies into its union contract. No AI text, no AI illustration. Editor Nathan J. Robinson’s reasoning was not sentimental. His readers pay for human thought, so automating the writing would destroy the only thing the magazine sells. Meanwhile the CEO of Business Insider says his publication is “all in on AI,” and Sports Illustrated got caught running articles by journalists who did not exist.

Both are strategic decisions, and only one of them was made on purpose. If you run quality or operations, you face the same question on a shop floor instead of a masthead. This article breaks down how to decide where AI belongs in your process, where it quietly erodes the value your customers actually pay for, and how to tell the difference before you scale.

A Magazine Just Wrote an AI Ban Into Its Union Contract

Notice what Robinson did not argue. He did not claim the tools produce bad text, or that they fail at illustration, or that the technology will plateau. He skipped the capability debate entirely. His position is that a magazine selling human thought cannot automate the thinking, no matter how good the output gets.

Compare that to the Wall Street Journal, which ran an AI-generated op-ed under Stan Druckenmiller’s name and then defended publishing it because the piece reflected his genuine opinion. Same technology. Opposite conclusion about what the product is.

That is the only question worth arguing about. Not whether AI works, but whether the thing you automate is the thing your customer is paying for. Get that wrong in either direction and the tools are irrelevant.

The Real Calculation Behind ‘All In on AI’

Robinson spells out the arithmetic publishers are actually running. Producing 100 posts with AI costs almost nothing and takes four minutes. Paying a writer for 20 posts costs $5,000. On per-story economics, the writer loses every time.

What makes the math uncomfortable is that it survives failure. Quality can drop, errors can creep in, readers can leave, and the spreadsheet still works. As Robinson puts it, you could lose 60 percent of your readership and still make more per story because you can pump them out in seconds. Futurism reports creative industries being “eviscerated,” with more than 200,000 jobs gone in four years.

Why per-unit cost logic breaks when the output is the product itself

Cost-per-unit thinking works when the unit is interchangeable. A bolt is a bolt. Ten thousand identical bolts at a lower cost each is straightforwardly better, and that is exactly where automation earns its keep on a factory floor.

It stops working when the unit is the thing customers are buying for its own sake. An article is not a bolt. Neither is an audit finding, a deviation investigation, or a root cause analysis that a regulator will read. In those cases cheaper units do not mean more value, they mean you have quietly swapped out the product and kept the invoice the same.

The Druckenmiller op-ed and what ‘genuine opinion’ defences actually concede

Defending that op-ed on the grounds that it reflected the author’s genuine opinion gives away more than it protects. It admits the writing itself carried no value, that only the position mattered, and that the prose was packaging.

Operations leaders make the same concession more often than they notice. When a supplier assessment gets generated and nobody reads it closely, you have declared the document a formality. That may be true, in which case automate it and stop pretending. If it is not true, the generated version is a liability waiting for an auditor to find it.

Where the Publishing Analogy Holds for Manufacturing, and Where It Breaks

A magazine sells human thought. You sell conforming parts, shipped on time, with documentation that survives an audit. Nobody on your customer’s receiving dock cares whether a human wrote the deviation summary. They care whether the part is in spec.

That difference means a blanket ban would be the wrong lesson to take from publishing. The right lesson is the method: decide which part of your work the customer is actually paying for, then keep machines out of that part and nowhere else.

Process work vs. judgment work: the dividing line that holds in any industry

Process work is retrieval, formatting, cross-referencing, and first-draft assembly. Deviation triage, pulling the right work instruction out of a 4,000-document library, drafting a CAPA from investigation notes, assembling the evidence pack before a supplier audit. None of it touches the product. All of it eats your team’s week.

Judgment work is different. Root-cause reasoning, the engineering call on whether a borderline measurement means something, the signature releasing a lot. Those are the things you are personally accountable for, and no model shares that accountability with you.

The quality-system equivalent of a fake byline

Robinson describes freelance submissions arriving with “suspicious numbers of em-dashes” and tell-tale sentence patterns. Quality systems have the same tell. A root-cause section that reads fluently, cites the right clause, and explains nothing anyone actually investigated.

That is the failure to guard against. Not AI in the workflow, but AI output passed off as human analysis inside a record someone signed. An investigation nobody conducted is a fake byline with a lot more legal exposure.

So require attribution. Every AI-assisted artifact gets flagged as a draft, and a named person confirms the reasoning before it enters the record.

Writing Your Own AI Policy Before Someone Else Writes It For You

Most plants are operating with no written position at all, which means every engineer is inventing one privately. Somebody is pasting nonconformance data into a chatbot right now. You do not know who, and neither does your quality director. Current Affairs wrote its rules into a union contract, naming specific formats and specific prohibitions, and it still admits enforcement is imperfect: “We can’t promise that nothing will ever slip through the cracks, but we’re vigilant!” That is the right posture. Write the boundary, then assume leakage.

The permitted/prohibited list every quality function should have by Q1

Keep it to one page and two columns. Permitted: drafting work instructions from existing approved documents, summarising audit findings for internal review, translating procedures, clustering complaint text, first-pass data extraction from supplier certs. Prohibited: anything that constitutes a release decision, root cause conclusions, CAPA effectiveness verification, and any signature-bearing document where the signer did not read the underlying evidence.

Then name a human owner for every AI-assisted output that leaves the building. Not a department, a person, with their name on the record. The practical effect is speed. Engineers stop guessing whether they are allowed to use the tool, stop hiding it when they do, and stop waiting on a governance committee that meets quarterly.

Auditing supplier and contractor documentation for undisclosed AI use

Your suppliers are running the same unwritten experiment. Add a disclosure clause to the quality agreement requiring suppliers to declare AI involvement in 8D reports, PPAP narratives, validation protocols, and corrective action responses. Disclosure is not prohibition. It tells you which documents need a closer read.

Build the detection habit into incoming document review. Current Affairs flags freelance submissions with suspicious em-dash density and “this is not an X, it’s a Y” phrasing. Your tell is different: 8D narratives with fluent root cause language that does not match the process data, or corrective actions written in a register the supplier has never used before. Flag, question, verify against the raw evidence.

The Competitive Advantage of Being Specific About AI in 2026

The plants that struggle over the next two years will not be the ones that adopted AI or the ones that refused it. They will be the ones that never separated the product from the overhead around it. Without that split, every AI question becomes a debate, and every debate stalls a decision that should take ten minutes.

Specificity pays three ways. Adoption moves faster where it compounds, because nobody has to re-litigate whether a tool is allowed. Rework drops to zero on documentation that would otherwise need unpicking after the fact. And when a customer or auditor asks how a decision was reached, you have a written answer instead of a shrug.

Three questions to settle with your leadership team this quarter

Put ninety minutes on the calendar and come out with answers in writing. Not a framework. Answers.

• What are we actually selling?: Name it in one sentence. If you cannot, you have no basis for deciding what machines may touch.
• Which tasks are overhead?: Shift reports, first-draft work instructions, supplier correspondence, data cleanup. These are the fastest wins and the easiest to approve.
• What must never be automated, and who enforces it?: Root cause conclusions, disposition calls, anything with a signature attached. Name the owner, not the department.

Write it down, date it, and tell everyone. A publishing magazine put its position into a union contract rather than a blog post, which is what made it real. Your version can be a one-page policy pinned to the quality management system. The format matters less than the fact that it exists and has an owner.

Revisit it every six months, because the capability line will move and your boundary should not move with it by accident. The point of an AI ban in one area is that it buys you permission to move aggressively everywhere else. Teams that know exactly where the fence is stop asking and start building.

Source: currentaffairs.org