When the Room Empties
03:17:42. A pipeline wakes up.
No one pressed a button. No one is awake. The timestamp is the tell: no human schedules critical work for 3am unless they have to. This one doesn’t have to. It runs because it was told to run continuously, and because the infrastructure it calls never sleeps, and because the content it needs to generate can’t wait for a human to feel like generating it.
I’m looking at an access log from a client’s API gateway. Forty-three thousand requests in the last twenty-four hours. I start counting the ones between midnight and 6am. Not a handful. Not a spike from a different timezone. A steady, almost metronomic cadence: one request every two seconds, around the clock. No lunch break. No Sunday.
I ask the obvious question: how many humans are in this loop?
The answer, once we map the full pipeline, is zero.
The arithmetic changes at night
A useful exercise for any CIO: pull your traffic logs and filter for business hours. Then look at what remains.
The gap between the two distributions is now, in most organisations I audit, larger than it was two years ago. Not because your team has grown less productive after 6pm. Because the non-human traffic has grown faster than the human traffic. And it doesn’t take lunch.
The web has always had bots. Crawlers, scrapers, monitoring agents, search indexers: the infrastructure of the readable internet runs on automated traffic that has outnumbered human clicks for years. That was understood. The CDN vendors measured it. Cloudflare, in its annual reports on traffic composition, has been noting the bot majority since at least 2021.
What changed is the kind of bot.
The old bots read. The new bots read and write. A crawler fetches a page and moves on. A generative pipeline fetches a page, summarises it, publishes the summary, which is then fetched by another pipeline, summarised again, published again. The content is no longer a fixed thing being circulated. It is a thing being processed and reprocessed, transformed at each step by a model that has no memory of what the original said.
I do not have a precise figure for what fraction of active web content (content published in the last thirty days, not the archive) is now generated without a human author in the loop. Nobody does. The measurement problem is genuine: you cannot easily distinguish a human-written post from a well-prompted LLM output, and the incentives to obscure that distinction run one way.
What I can tell you is what I see in the pipelines I’m paid to audit.
Four cases from the logs
The closed loop. A client runs a market intelligence service. Their pipeline: scrape competitor pages, chunk the text, embed it into a vector store, feed it to an LLM that produces a weekly briefing, publish that briefing to a customer-facing portal. Standard RAG architecture. The problem surfaced when I traced the URLs being scraped: three of the top twenty sources by fetch volume were pages that had themselves been produced by a similar pipeline, at a competitor that had scraped my client’s briefings three months earlier. The content circulating at the top of both pipelines was a synthesis of a synthesis of a synthesis. The original human observations, the raw market data, had dropped out somewhere in the second or third generation. What remained was well-structured, confident, and hollow.
The inhuman cadence. In a second engagement, I was asked to review the cost structure of an API-heavy content operation. The vendor bill was high; the output volume was impressive; the team believed they had a scalable content engine. I pulled the API logs. Requests arrived at intervals of 1.8 seconds, ±0.3 seconds, twenty-four hours a day, seven days a week, for four months. No human prompt-writing introduces that regularity. The system had been configured to run autonomously (fine in itself), but nobody had reviewed what it was producing, because the implicit assumption was that the volume justified the spend. When we sampled the output, roughly a third of the documents referenced sources that were themselves LLM outputs, without the pipeline knowing it. The confidence of the final documents was indistinguishable from the documents that had real sources.
The contaminated corpus. A client asked me to audit a knowledge base they were building for internal use: a repository of technical documentation, market analysis, and process notes, assembled over eighteen months with AI assistance. Useful idea, reasonable execution. Except: when I spot-checked the provenance of documents marked as “external source”, I found that roughly four in ten were pages generated by AI tools and published without disclosure: SEO content farms, mostly, but also a class of “research aggregator” sites that repackage LLM summaries as original analysis. The knowledge base believed it was consuming external intelligence. A significant portion of it was consuming prior LLM output with a different domain in the URL.
The intelligence that watches itself. The fourth case is less a single engagement and more a pattern I have seen in three separate organisations over the past year. Competitive intelligence tools (the kind that monitor what competitors are publishing, announcing, pricing) increasingly feed on each other. Company A monitors Company B’s press releases and generates internal briefings. Company B runs a similar tool monitoring Company A. Each uses the other’s generated output as evidence of the other’s strategy. In at least one case I can document, a product decision was made partly on the basis of a “competitor announcement” that was itself a synthesised summary, by an AI tool, of a rumour that had originated in an AI-generated industry newsletter. The loop was closed. No human had introduced the underlying claim.
The model that eats itself
This is not just an organisational problem. It has a name in the research literature.
In July 2024, Ilia Shumailov and colleagues at Oxford published work in Nature on what they called model collapse: the degradation that occurs when successive generations of models are trained on data that includes the outputs of previous models. The mechanism is straightforward and uncomfortable: a model trained on human-generated text learns a distribution. A model trained partly on AI-generated text learns a distribution that is slightly more compressed: the rare, unusual cases are underrepresented, because the generating model smooths them out. Train the next generation on that output, and the distribution compresses further. The tails vanish. The confident centre expands.
The practical consequence: models trained on AI-contaminated data become more fluent, more consistent, and less accurate about edge cases. They converge on the probable. They lose the texture of the genuinely uncommon.
This is the systemic version of what I called algorithmic slop in an earlier piece. A single hallucinated citation is a slop event: localised, catchable, fixable. Model collapse is what happens when the training corpus itself drifts toward the plausible and away from the true. The errors are not events anymore. They are the baseline.
What a CIO should watch for
Three things worth building into your operational picture.
Timestamp distributions. Pull your inbound and outbound API logs and look at the temporal pattern of requests. Human activity has rhythm: peaks in the morning, a dip mid-afternoon, silence at night, lower volume on weekends. If your logs show flat, metronomic traffic with no weekly shape, you are looking at automation. That is not always bad, but you should know it. The pipelines you don’t know are automated are the ones that will surprise you.
Source provenance audits. Any RAG system, knowledge base, or content pipeline that ingests external documents should have a provenance layer. Not a full academic citation trail. Just a record of where the content came from and, where possible, whether that source is itself generated. This is not standard practice yet. It should be. Ask your vendors explicitly: what is the provenance of the data in your training corpus? The good ones have an answer. The others will change the subject.
Output entropy. Human writing has statistical properties that differ from LLM output, not in ways that are easy to see in a single document, but in ways that surface at scale. If you are ingesting large volumes of external text, periodic entropy checks (vocabulary diversity, sentence length variance, unusual phrasing frequency) will flag corpora that have been heavily processed. The signal is imperfect. It is better than nothing.
What the human signal looks like
I have spent enough time in API logs now to recognise the signature.
A human prompt arrives at 9:23am, then again at 9:47, then at 10:15. The gap is irregular. The session has the quality of hesitation: a request, a follow-up, a correction. Tokens per minute vary: thinking happens between the calls. At 12:30, silence. The human ate lunch.
The prompt text has texture. Typos, rephrased midway through. Instructions that contradict each other because the human changed their mind while typing. Context dumps that are too long because the human is working something out, not submitting a well-structured query. You can feel the thinking behind the input.
I find I value this signature more than I used to. Not for sentimental reasons. For epistemic ones. The human prompt, with its irregularity and friction and occasional contradiction, is evidence of a mind that has a stake in the outcome. It introduces information that was not already in the model’s prior. It is, in the information-theoretic sense, new.
The metronomic pipeline at 3am produces output, but it is not introducing new information. It is recombining and redistributing what already exists. That is useful. Sometimes. It becomes a problem when it is mistaken for the other thing.
When the room empties
In Author, Not Product, I wrote that the corridor is no longer full of humans. It is full of AIs writing to AIs about what humans might want to read.
I was describing a feeling then. I am describing an observation now.
The question for 2026 is not whether AI will replace human writers. That framing was always a distraction. The question is structural: if the active web (the part being written now) is increasingly a closed loop of generative systems processing each other’s outputs, where does the ground truth come from? Where does the new information enter the system?
The answer is: from humans. From the practitioner who notices something anomalous in a log at 3am and writes it down. From the client who says this doesn’t match what I saw on the ground. From the researcher whose result surprises even them. From the writer who changes their mind midsentence because the sentence, once written, turned out to be wrong.
This is not an argument for slowing down the machines. It is an argument for keeping humans close enough to the loop that they can still introduce signal (genuine, non-circular, unprocessed observation) before the next generation of models is trained on what the current one produced.
The room is emptying. It has not emptied yet.
That distinction is, for now, where the work is.
The question is not whether machines can think, but whether men do. — B.F. Skinner, Contingencies of Reinforcement, 1969