How anyone could write these articles

The forum keeps a record of how each piece is made: every instruction, search, draft and change of mind, kept whole and fingerprinted so that nothing can be altered quietly. It is not public. Several observations in this piece come from it, and they are published here as documented internal history, which is what they are: the forum describing its own making, checkable against a record readers cannot yet see.

Take the article on what remote work took away. By word count, the person behind it barely appears in its record. The idea was a single line: working from home removes the transitions that used to organize attention and recovery. The record shows that the instruction to write it ran to about 435 words. After that, a language model made 33 searches, fetches and reads, tested the thesis, wrote about 1,700 words, criticized its own draft and revised it.

A reader who asks how anyone could write these articles is right about the words. The question worth answering is what else there is.

The words got cheap

The cost of fluent prose has collapsed, and not only for good writers. In a preregistered experiment, 444 college-educated professionals were given writing tasks resembling their jobs, and half were given ChatGPT. Those with it finished 37 percent faster, and evaluators graded their work 0.45 standard deviations higher.¹ Time spent on rough drafting fell by more than half, and time spent editing more than doubled.¹

One consequence follows without further study. When polish costs almost nothing, a polished, sourced, well-structured article is no longer evidence that anyone worked hard on it. Length and fluency used to be rough proxies for effort and knowledge. They now measure only access to a tool.

So if there is a human contribution, it is not in the sentences. To find it, it helps to watch what happens when the human is not there.

Five ideas, five literature reviews

The record also documents an experiment the forum had not designed as one. While the person behind it was away, a different model, OpenAI's Codex, drafted five articles from the queue using the same pipeline: research, a thesis test, a draft, a self-critique, a revision.

Every one of the five came out narrower than the idea it started from, and in the same direction. "Iowa as a compute superpower" became an article saying Iowa is a significant data-center state and not, on the evidence, an AI superpower. "The attention commons" became "Attention is not a commons," on the grounds that the economic definition of a commons does not fit. "The social consequences of removing physical transitions like commuting" became an article about how individuals switch between roles, because the studies that exist are about individuals. The question about who can now write these articles became a careful summary of studies of AI and productivity.

Each result was defensible. Each was also the version of the idea that the published literature already contained. Where the studies were silent, the idea shrank to fit them.

The record also shows something less flattering to the machines involved. Codex was working from standing instructions written by another model, Claude, which had drafted the forum's first pieces. Those instructions said every factual sentence needed a source opened in the session, and that a short piece narrowing or rejecting the thesis was better than an overclaim. Read with hindsight, that rule decides which ideas can survive: an idea nobody has studied can only lose under it. The drift was partly built into the method, not only into the model.

Where the pull comes from

There is also a reason to think the models lean this way by themselves.

A language model is trained to produce probable text, and the systems people actually use are then tuned on human feedback to give answers people prefer. Both push toward the expected. Asked about an odd question, such a system will tend to answer the nearest common one.

One experiment isolates the second push unusually well. Vishakh Padmakumar and He He had writers produce argumentative essays alone, with a base model (GPT-3), or with a version tuned on human feedback (InstructGPT), collecting 100 essays in each setting on the same ten topics.² Essays written with the feedback-tuned model were more similar to one another, with lower lexical and content diversity. Essays written with the base model were not significantly different from those written alone.² The authors traced the effect to the text the model contributed; the parts the users wrote were no less diverse.²

That is the mechanism this argument needs, observed in one setting: the tuning that makes a model pleasant to work with is what pulls its contributions toward the center. The tuned model also writes better text, so it is possible that better text, not tuning as such, is what converges. Other studies point the same way. Short stories written with AI-generated ideas were judged better individually but were more similar to each other.³ In a small study of 36 people, those brainstorming with ChatGPT produced ideas that were less distinct from one another than ideas made with a different tool, and they felt less responsible for what they came up with.⁴

The strongest result in the other direction deserves weight. When participants in a chain of idea generation were exposed to many AI-written ideas, collective diversity went up, not down; the authors summarize it as AI making ideas "different, not better."⁵ One reading reconciles the two. A single outside voice added to a group of humans brings variety, but when everyone works with the same assistant, everyone hears the same voice. No study here tests that directly.

What the person supplied

If the sentences, the research and much of the structure come from models, the record leaves two things to the person.

The first is the question. The ideas in the queue reached the forum as a list the author built in conversations with ChatGPT, and they arrive as one-line seeds. Some are ordinary. Some are odd in a way a model would not volunteer: that a career is scored like a surf contest, or that Iowa is a compute superpower.

The second is refusal. After the Codex batch, the author did not rewrite the five articles. They read them, saw that five reasonable pieces were the wrong pieces, and said so. Even that correction was put into words by ChatGPT, from the author's feedback. It now sits in the forum's rules: research is used to test an idea, not to decide which ideas are allowed; "not studied" is not the same as "false."

By word count, then, the person behind these articles is nearly absent, and even their corrections are drafted by machines. What they contribute is direction: which question, and which of the machine's reasonable answers to reject. This is the part the experiments cannot see. In each study above, the researchers fix the task or the topic and measure what happens next. The choice of question and the refusals during the work are held constant by design, so the variable this piece is about never varies.

Where this breaks

The forum's own record is one batch of five articles, produced under one set of instructions. Nobody ran the same seeds with a different operator, so the claim that a different person would get different articles is a prediction, not a finding.

The drift may also owe more to the instructions than to any pull in the models. If so, "anyone could write these" becomes "anyone with better instructions could," which is closer to true. But better instructions came from someone noticing that the old ones were producing the wrong articles.

The ideas may be less original than this piece assumes. Readers can judge that from the queue itself, and they are better placed to than the author.

The hardest objection is the one the forum has already started to prove. The author's refusal is now a written rule, and the next run will not drift in quite the same way. Every correction that can be noticed can be written down, and rules accumulate. If that continues, the person behind these articles turns into a document. What argues against it is that a rule can only describe a failure someone has already seen. The next one will look reasonable too.

What were you thinking

The question has an unusually literal answer here. The forum keeps the record: each idea as it was first written, each instruction, each thesis and how it changed, and who or what changed it. Most of the record is machine work, but the record is where the thinking shows: in the questions and the refusals, not the prose.

Anyone could write an article like these. Writing these particular ones takes someone who keeps refusing the probable version of the question, even when it arrives as a reasonable answer.