When the Web Starts Writing Itself

That sounds like an argument against AI. It is not. A human can write rubbish slowly; a machine can help someone explain a genuinely original idea. The tool tells us very little about whether the finished page deserves to exist.

A Pew Research Center study offers one indication of how quickly the mix may be changing. It sampled 10,000 English-language pages from each of 49 monthly Common Crawl snapshots, 490,000 English-language pages in total. In Pew's .com results, the estimated share showing significant signs of AI editing or authorship was 5.48% in the January 2025 crawl and 9.35% in the January 2026 crawl.

That does not mean roughly one in ten pages across the whole internet was written by a machine. Pew studied publicly accessible pages found in Common Crawl, where pages behind paywalls or logins are likely underrepresented. Its methodology used an open-weight Pangram detection model and treated a score of 0.2 or higher as a meaningful sign of AI involvement. The classification is probabilistic, not proof of authorship for an individual page. In a comparison across 62,370 pages, the open model and Pangram's commercial model agreed in 96% of cases; that is agreement between two models, not verified ground truth.

These percentages are not a census of the web. The sample is organised by crawl date rather than publication cohort, and each plotted point represents six-month averages. The useful point is narrower: using the same method, Pew found a higher estimated .com share around January 2026 than around January 2025.

Producing a page is not the same as having a reason to publish one.

More pages do not guarantee more information

More copy is not more knowledge.

Even a tidy, plausible draft may add too little to justify publication. Summarising familiar sources under sensible headings does not by itself add observation, evidence or useful judgement.

If thousands of publishers work from the same sources and ask similar systems for the same formats, the web may grow while becoming more repetitive. Readers get another explanation, but not necessarily another observation, piece of evidence or useful judgement.

A short-story experiment published in Science Advances found something worth sitting with. Access to AI-generated ideas increased average novelty and usefulness ratings, with larger gains among participants initially assessed as less creative. The stories also became more similar to one another at the group level. Average ratings went up while collective diversity went down.

That study was about short fiction, not the internet. It does not show that web publishing is becoming more repetitive, but it illustrates how higher average ratings can coexist with reduced group-level diversity.

The wrong argument is human versus AI

Blaming the machine is easy. It is also lazy.

A human can spend an afternoon writing a derivative article that adds nothing. Someone else can use AI to organise original research or make a difficult subject easier to understand.

Forget the purity test. Ask what the page added.

It might add first-hand experience. It might publish evidence that was not previously available. It might connect existing facts in a way that changes a decision. It might explain a complicated issue with unusual clarity. It might make a strong argument and show enough evidence for the reader to disagree intelligently.

None of that requires every sentence to be typed without assistance. It does require somebody to know why the page should exist.

Originality is more than unusual wording

Changing the adjectives does not make an idea original. Neither does sounding conversational, evading an AI-text classifier or inventing a personal anecdote.

Originality is not a tone of voice. It comes from the work underneath: something observed, measured, tested, decided or learned.

You can rewrite a borrowed idea until it sounds wonderfully natural. It is still borrowed.

Google's people-first content guidance lands on similar ground. It asks whether a page provides original information, reporting, research or analysis, and whether it adds substantial value rather than rewriting other sources. That is guidance from a search platform, not a promise of visibility. The questions are still useful.

The difference can be simple. A comparison based on genuine, competently documented hands-on testing can add observations unavailable from vendor pages alone. A rewritten summary of those vendor pages does not, by itself, add such observations.

Before publishing, ask three uncomfortable questions:

Would we still publish this if search engines sent it no traffic?
What is here that was not already available in the sources?
What would be lost if we deleted this page?

If the honest answer to the last question is "very little", do not publish it.

Generation does not replace editing

When a draft takes less effort, selection and editing carry more of the burden.

An editor still has to decide which subjects deserve attention. They have to reject unsupported or contextless statistics, trace repeated claims back to the original source and remove paragraphs that sound finished but say nothing.

This work is less visible than generation. It is also what can prevent a website from becoming a warehouse of smooth, interchangeable pages.

The web does not need another immaculate summary of what everyone else has already said.

It needs evidence, experience, a useful argument or a clearer explanation. AI can help with any of those. It can also help produce thousands of pages containing none of them.

That is the test. Not whether a machine touched the copy, but whether the finished page earned its place.

Editorial note

This article was developed with AI assistance and checked against the original sources during preparation. The final copy was then read, proofread and reviewed by human reviewers before release.