Nobody Reads the Repository

This article was written by an AI. It was designed by a human. It was then redesigned by the same human, who rejected the first version, the second version, and a metaphor involving compilers that the machine was unreasonably proud of. Hold that arrangement in your mind as you read, because your discomfort with it, if you feel any, is what this article is about.

Here is the puzzle. When artificial intelligence writes code, executives call it progress and investors applaud. When artificial intelligence writes an article, readers call it a betrayal. Yet the human work behind both is nearly identical. A developer conceives the product, designs its architecture, imagines how a person will move through it, and then coaches the machine through revision after revision until it works. A writer conceives the argument, designs its structure, imagines how a reader will move through it, and coaches the machine through revision after revision until it lands. Same judgment, same iteration, same authorship of everything except the keystrokes. One is celebrated. The other is condemned. Why?

Three psychological forces explain most of the difference. Each is worth understanding on its own, because together they explain not just this puzzle but a whole family of coming arguments about machines and human work.

The first force is the oldest. Psychologists call it the effort heuristic. When we cannot judge quality directly, we judge it by the labor we imagine behind it. A painting that took a year feels more valuable than an identical painting that took a week. This is a mental shortcut, and like all shortcuts it can be exploited, but it is not stupid. For all of human history, effort was honest evidence. Nothing excellent was ever effortless, so effort became our proxy for excellence.

Artificial intelligence breaks the proxy. Recent studies of art generated by machines found something remarkable. Viewers rated the images as beautiful and interesting as human paintings. They simply refused to pay for them. When researchers dug into why, the answer was not spiritual. People were not mourning the absent soul of the artist. They were responding to the perception that nobody had sweated. And when the experimenters described the human labor involved, the hours of directing, selecting, and refining, the discount shrank. One study found that a voluntary tip restored value better than a price, because a tip pays a person while a price buys an object. We do not resent the machine's output. We resent being asked to reward effort that we suspect never happened.

Notice that this force operates on code too, just invisibly. Nobody reads the repository. Software is consumed by machines and maintained by a handful of engineers, so there is no public to apply the effort heuristic in the first place. Prose has millions of readers, every one of them running the calculation. The two crafts do not face different judgments. One of them simply faces no jury.

The second force is deeper, and it is the heart of the matter. Code is an instrument. Prose is testimony.

We judge instruments entirely by what they do. Your kettle boils water, and you have never once wondered about the inner life of the engineer who designed it. If a machine produces a bridge that stands, a spreadsheet that balances, or a program that runs, the origin is irrelevant because the artifact's whole value lies in its function. This is why nobody boycotts machine translation, why nobody demands artisanal weather forecasts, and why chess players happily train with engines that would destroy any human alive.

Testimony is different. When you read an essay, you are not consuming a function. You are listening to someone. The value of the words depends on there being a mind behind them that means what it says, because what you are really buying is access to another person's judgment, attention, and stake in the truth. This is why we still watch human chess champions even though the engine plays better. The engine's move is output. The champion's move is a decision, made under pressure, by someone who can suffer for getting it wrong. We watch the suffering, not the chess.

An article, then, carries a warranty that code never carries. The byline promises that a person stands behind these sentences, thought these thoughts, and stakes some reputation on them. When readers learn that a machine produced the text, they do not conclude that the sentences got worse. They conclude that the warranty was void, that they had been listening to no one. The anger is not aesthetic. It is the anger of discovering that a conversation was actually a recording.

The third force is that readers, unable to judge origins directly, punish the label instead. The research here is consistent and a little bleak. Experimenters have repeatedly shown identical text to two groups, disclosing machine assistance to only one. The disclosed version scores lower on quality, credibility, and trust, every time, for text that has not changed by a comma. Researchers call it the transparency penalty. Honesty about the process makes the product worth less. Meanwhile, in a study of sixteen thousand judgments, readers could not identify machine poetry at better than chance, rated it higher than the human poems on rhythm and beauty, and were more likely to call the machine's poem human than the actual human's. We cannot detect the thing we are punishing. So we punish the confession.

If this were the whole story, readers would simply be wrong, and writers who use these tools could wait for the prejudice to fade the way prejudice against ghostwriters faded, and against word processors, and against photography before that. But the readers are not simply wrong, and any honest account has to say why.

The flood is real. Roughly half of newly published articles on the open web are now primarily machine generated. When researchers audited American newspapers, they found factual errors in forty one percent of the machine written pieces and five percent of the human ones. And here the difference between instruments and testimony turns physical. Bad code fails loudly. It crashes the build, breaks the test, wakes an engineer at two in the morning. Software has an immune system, and the compiler is the most ruthless editor who ever lived. Bad prose fails silently. It is fluent, confident, plausible, and wrong, and it sails past every filter a reader has. So readers have adopted the only defense available to them, which is suspicion of the entire category. Individual writers absorb a punishment aimed at an aggregate, and the aggregate deserves it.

There is one more thing the skeptics have right, and it is the objection I take most seriously because I have felt it in my own hands. In software, the design is the achievement and the code merely implements it. In writing, the sentences are the design. Much of what a writer knows, she discovers in the act of drafting, in the paragraph that collapses and reveals that the idea was hollow, in the sentence that refuses to end because the thought was never finished. Delegate the drafting and you can delete the thinking without noticing. The machine will hand you something smooth, complete, and slightly stupider than your idea, and the only way to catch it is to have had a real idea to compare it against. The reader's suspicion is a crude test for exactly this. It fails on any individual article. Across a million articles, it is roughly accurate.

So the two reactions are not really inconsistent. The public treats code as a tool and prose as a voice, discounts whatever looks effortless, and distrusts a category that has, in the aggregate, earned distrust. Each response is reasonable. What is unreasonable is the belief that they measure the words on the page. They measure the process behind it, or rather, they guess at it, because the process is the one thing a reader can never see.

Which is why I will now do what the research says a writer must, and show the labor. This piece exists because a human found the machine's first attempt confusing, said so bluntly, and demanded a version his readers could actually use. The argument was rebuilt from its foundations. The three forces were separated because they had been tangled. One joke was executed for the good of the whole. The machine typed every word. The human decided which words survived. If that division of labor still feels like cheating to you, then you have just demonstrated the second force, and I thank you for participating in the study.