Current Affairs bans AI-written text and art, written into its union contract

The rule, and the arithmetic behind it

Current Affairs says no AI-generated text may be published and no AI art used as illustration, as part of its new union contract. Writers may use AI for research; artists may use it for small tasks like adjusting colors. The test is whether a tool helps workers or replaces them.

The article argues the incentive structure is the problem. A for-profit outlet answers to shareholders, and its own illustration makes the math plain: 100 AI posts in four minutes at almost no cost, against 20 writer-produced posts for $5,000, which means a company could lose 60% of readers and still profit. Examples of the rush: Business Insider's CEO saying the outlet is "all in on AI", Sports Illustrated publishing pieces by nonexistent journalists, and the Wall Street Journal running an AI-generated op-ed by Stan Druckenmiller. The magazine admits some AI content may slip into freelance submissions.

The shovel parable, and the reader who decides to read

One commenter opened with a parable about two neighbors at a construction site, where five workers with two machines dig what fifty could dig with shovels, or a thousand with bare hands. The moral they drew: "Good luck trying to figure out whether something was written by AI or a human", and the whole panic "will soon be forgotten".

A commenter with 28k karma answered on the detection point. It is a magazine, not a social network: contributors are known, and those who break trust "are banned and replaced with someone else". They also doubted that nobody will care, since people still buy print books. A shorter rebuttal came from a commenter who took the parable apart:

"The value of moving dirt is that someone paid for a hole. They don't care how it was made. The value of written communication is in whether people decide to read it." — commenter replying to the shovel parable

The loudest critique aimed at a ban the policy does not contain

A founder and CEO of a software consultancy, who advertises for developers fluent with AI coding agents, called the policy "incredibly short-sighted". The founder's list of what a newsroom could still do: fact-checking research, story ideas drawn from its own archive, and sifting FOIA PDF dumps and government datasets.

A commenter with 24k karma replied in three sentences that the list was already allowed:

"They didn't ban the use of AI. They banned AI-generated text or images from appearing in the magazine." — commenter with 24k karma who read the policy

Research is exactly what the article permits. The founder was arguing with a reading of the policy as job-security fear, which the founder supported by noting it arrived through a union contract.

From training-data theft to bosses firing John Smith

The article's moral premise is that models replacing artists were trained on those same artists' work. A Bitcoin-focused commenter asked "which biological neural network was not trained on copyrighted material from other artists?". A reply from a commenter with 16k karma drew the line at legitimacy:

"They didn't all pirate, on a historic, industrial scale, every scrap of creative work that they saw in their lives. That's the key distinction." — commenter with 16k karma

Jeremy Kun, a Google engineer who works on homomorphic encryption and writes about math and programming, moved the argument somewhere else. Nobody minds hobbyists making a "Star Wars but make it Monet" picture for their wall, he said. The harm is "the part where bosses fire John Smith and then ask the LLM" for a hundred articles in his style. Under that framing, the question of where the training data came from stops mattering for the staff of a magazine.

Checking: what stops someone, and what counts as help

A commenter who works at a place that uses AI for nearly everything said their workplace still holds one rule: if the audience of a text is human, a human writes it. A reply asked what text a workplace produces for non-human audiences "other than code". And a developer building an "idea to published game" product asked the enforcement question directly: whether it is a "gentlemens agreement" or something enforced, since anyone can prompt a model to imitate a human style.

The edges of the rule turned up in a post from a technical writer learning to write fiction, who uses an LLM to check whether their prose induces emotions and asked whether Current Affairs would permit that. A reply asked how an LLM could tell them that "with any insight that is greater than your own". The quoted rules allow research help and small touch-ups, bar generated text, and say nothing on feedback.

The one rule none of these exchanges knocked down is audience-based: text a human will read, a human writes. It is cheaper to hold by contract with named staff than to prove by detection, which is why a union clause may be a more practical enforcement tool than any scanner.