In October 2026, Current Affairs wrote a total AI ban into its union contract. No AI text, no AI illustrations. Meanwhile the CEO of Business Insider says his publication is “all in on AI.” Two opposite decisions, same underlying mistake: neither one names the specific work where human judgment actually changes the outcome. One bans the tool everywhere. The other removes the human everywhere. Both are blanket rules standing in for a decision nobody wanted to make.

You face the same choice on your factory floor. Deviation reports, CAPA write-ups, supplier audits, batch record reviews: some of that is judgment you are paid for, and some is expensive manual labour you have never questioned. This article covers how to tell them apart and write an AI usage policy that holds up.

A Magazine Banned AI Outright While Business Insider Went ‘All In’

Nathan J. Robinson framed the Current Affairs ban as a direct challenge to what he calls “the slopification of media.” He is not wrong about the problem. A total prohibition is just a crude answer to it.

The opposite extreme is worse, and Sports Illustrated showed why: AI-written pieces published under the bylines of journalists who did not exist. Nobody decided that was acceptable. Nobody decided anything. The tool filled a gap where editorial judgment used to sit, and the failure only surfaced after publication.

Most plants are parked at one of these poles. Either AI is informally banned because quality has not sanctioned it, or it has crept into reporting and documentation with no sign-off rules at all. Neither position is an AI usage policy. Both are the absence of one.

What Current Affairs Actually Banned, and What It Left Alone

Read the policy closely and it is narrower than the headline suggests. No AI-generated text in the magazine. No AI-generated art used as an illustration. That is it. The ban lands on published output, the words and images readers pay for and attribute to a named human being.

Nothing in it stops the magazine from using software to schedule, transcribe, format, or search. Robinson’s objection is to the “slopification of media” and to publishers replacing writers with cheaper output. He drew the line around the artefact, not around every computer in the building.

Output-level bans versus blanket technology bans

This distinction decides whether a policy survives contact with reality. A technology ban (“no AI tools here”) dies the first week, because half your MES, vision inspection, and planning software already has machine learning inside it. Nobody can enforce a rule that would shut down the plant.

An output-level rule holds up because it is checkable. Name the deliverables where a human signature means something: the root cause statement in a CAPA, the disposition decision on a nonconformance, the risk rating in a supplier audit. Then say plainly that a machine may draft around those, but a qualified person owns the conclusion. That is a workable AI usage policy, and it leaves the boring text assembly alone.

The detection problem Robinson openly admits he cannot fully solve

He is honest about enforcement, which is rarer than it should be. Current Affairs screens freelance submissions by watching for suspicious numbers of em-dashes and “this is not an X, it’s a Y” sentences. These are tells, not proof.

We can’t promise that nothing will ever slip through the cracks, but we’re vigilant!

Assume the same on your floor. You will not reliably detect which deviation narrative was drafted by a model. So stop trying. Build controls that work regardless of authorship: named approvers, evidence attached to every claim, and sampling of signed records against the underlying data.

The Cost Math That Drove Publishers to AI, and Where It Breaks in Manufacturing

Robinson spells out the arithmetic that moved the industry. Producing 100 posts with AI costs almost nothing and takes four minutes. Paying a writer for 20 posts costs $5,000. Even if quality drops, errors creep in, and 60 percent of the audience walks, the per-story economics can still look better than before.

That trade works because a lost subscriber is a rounding error. On a production floor, the person reading your output is an auditor, a customer quality engineer, or a notified body assessor. The downside is not churn. It is a recall, a failed certification, or a consent decree.

Accountability erodes from the top first. The Wall Street Journal published an AI-generated op-ed from hedge fund billionaire Stan Druckenmiller and then defended it, on the grounds that it reflected his genuine opinion. Senior sign-off became a rubber stamp on content nobody actually authored.

Why volume economics do not apply to CAPA reports and audit evidence

You do not need 100 CAPA reports. You need one that holds up when an investigator pulls the file three years later and asks why you concluded what you concluded. Volume is not the constraint. Defensibility is.

Drafting, formatting, cross-referencing prior deviations, pulling batch data into a template: that work scales beautifully with AI. Root cause determination, effectiveness checks, and the decision to close do not. An AI usage policy that fails to separate those two categories will either block the drafting (wasting hours) or automate the judgment (creating a finding).

Reputation debt: what the Sports Illustrated episode cost

The damage was not the articles. It was the discovery that nobody could vouch for how they were made. Once that becomes public, every piece of past output is suspect, including the legitimate work.

Manufacturing carries the same exposure, with less forgiveness. A customer who finds one unsourced claim in a supplier audit response re-opens the whole file. Regulators escalate to for-cause inspections. Rebuilding that trust takes years and costs far more than the labour you saved.

Drawing the Line: Which Quality and Operations Tasks AI Should Own

Run this test on every recurring task in your quality system. Ask one question: does this produce a judgment someone has to stand behind by name, or does it move, format, extract, or reconcile information that already exists? Two columns, an afternoon’s work, and you have the skeleton of an AI usage policy that actually tells people something.

The signature test for any AI-assisted output

If a document ends with a name, a date, and a title, the reasoning behind it belongs to a person. That is not a philosophical position. It is what a notified body assessor will ask you to defend, line by line, and “the model suggested it” is not an answer anyone accepts.

The inverse holds just as firmly. Pulling six months of batch data into one view, matching supplier certificates against incoming specs, drafting the first pass of a non-conformance summary: nobody signs those. They are expensive manual labour wearing the costume of expert work, and your engineers know it.

Where the hours actually come back, with realistic numbers

Measure the win in hours returned, not heads removed. A quality engineer who spends a third of her week assembling evidence instead of analysing it does not need replacing. She needs the assembly done before she sits down.

Robinson’s fear is that AI gets used to produce output “at a fraction of the expense that a writer costs,” with quality as the acceptable casualty. He is describing a real failure mode. A policy built on the signature test makes that substitution structurally impossible, because the judgment work was never on the table.

Write Your AI Usage Policy Before Procurement or a Regulator Writes It for You

The thing worth copying from Current Affairs is not the ban. It is the fact that the policy sits inside a union contract. That makes it enforceable, visible to readers, and awkward to quietly reverse next quarter. Most manufacturers have the opposite: a Slack message from someone in IT, a vague line in the acceptable use policy, and twelve people using ChatGPT on deviation text without telling anyone.

That gap is where your risk lives. A customer quality engineer asks how an investigation report was produced and nobody can answer. Procurement signs an AI-embedded MES module with no clause about what it is allowed to conclude. Write the rules first, while you still get to choose them.

A four-clause policy template you can adapt this quarter

• Named authorship: List the outputs that require a human name attached to the conclusion. Root cause statements, risk acceptances, batch dispositions, supplier approvals. The name means that person read the evidence, not that they approved a draft.
• Disclosure: Internally, every AI-assisted document carries a flag. Externally, agree with your top five customers what they want told and when. Do this before an audit raises it, not during one.
• Audit logging: Record which tool, which version, which prompt or input set, and which human reviewed the result. If you cannot reconstruct how an output was produced, it does not go in a regulated record.
• Review cadence: One owner, quarterly. Models change faster than your document control cycle, and a policy written against last year’s capability becomes either a loophole or a blocker.

Four clauses, two pages, signed by the quality director and the operations lead. Not a committee project.

The organisations getting real value from AI in 2026 are the ones that said clearly where it does not go. Robinson drew his line to protect “human thought and creativity.” Draw yours around judgment, then let your team use the tool freely everywhere else. Permission is the point.

Source: currentaffairs.org