Readers Can Smell LLM Prose, and They Stop Trusting You

You know the paragraph. It opens with a claim about how fast things are moving, lists three benefits in a row, and ends on a sentence that restates the paragraph. Nothing in it is false. A reader gets through two lines of it, recognises the rhythm, and quietly files the whole page under "nobody was here." That reaction takes about a second, and it is the most expensive second your content will ever cost you, because the reader does not just skip the page. They revise their opinion of the company that published it.

On October 29, 2026, Martin Fowler published a short piece called LLM Voice and it landed on the Hacker News front page the same day. His advice is blunt. We think he is right, and we think the commercial case for following it is stronger than the aesthetic one. This post makes that case, goes through what the evidence says about AI drafted claims, and ends with an editorial policy short enough to paste into a content team's doc.

What Fowler Actually Said

Read the piece itself, because the summaries flatten it into "AI writing bad." That is not the argument. The line that matters is this one:

Read the word "trap." Fowler is not claiming the draft will be ugly. A modern model produces clean, grammatical, well organised text on any topic in moments. The trap is that the draft looks finished. It has headings, transitions, and a confident register, so the author's job quietly changes from "decide what I think" to "edit what the machine thought." Those are different jobs. The second one is much easier, and it feels like progress.

His second point is the one we keep coming back to. If authors delegate writing to LLMs, he argues, they will also delegate thinking and comprehension to LLMs. That is a claim about what writing is for. Anyone who has tried to explain a system they half understand knows the sentence you cannot finish is a map of where your understanding stops. Writing is the cheapest test of comprehension there is. Skip it and you also skip the test.

We should be honest about what this post does not claim. We are not citing Fowler for a statistic, because he does not offer one, and we will not invent a figure about how often readers detect machine text. Nobody here has run that study. What we can argue from is his reasoning, the sourcing evidence below, and our own judgment about how content programs go wrong. Where the evidence is our own judgment, we say so.

Why Readers Notice

Detection does not need a classifier. It needs a reader who has seen enough of the stuff. By late 2026 that is nearly everyone who reads for work. Engineers, marketers, analysts, and journalists have all spent two years skimming model output in their inboxes, search results, and support tickets. They have built a feel for it the way you build a feel for stock photos. You cannot name the rule, but you spot the violation.

What gives it away is rarely a single word. It is the uniform texture. Sentences of similar length marching in groups of three. Every paragraph ending on a neat takeaway. Hedges placed evenly, like seasoning. Claims with no names, no dates, no numbers, and no sign the author has ever touched the thing being described. Human writers are lumpy. They spend four paragraphs on the one detail that annoyed them and one line on the parts they find obvious. They contradict themselves and correct it in public. A model optimises for the middle of the distribution, and the middle is what readers have learned to skip.

The last test in that right column is the one that matters commercially. Interchangeable text has no value as a signal. When a company publishes a post, the reader is quietly asking what the company knows that they do not. An article that could have come from anyone answers "nothing." It also answers a second question the reader did not ask out loud: do these people respect me enough to say something they thought about?

The Distrust Does Not Stay on the Page

The damage spreads sideways. A reader who decides your blog is machine filler starts applying the same discount to your docs, your changelog, your case studies, and your sales deck. They have no way to tell which parts a person wrote, so they assume none of it was. This is the part that content leads underestimate when they approve a "scale the blog with AI" plan. The unit of trust is the brand, not the post.

The same dynamic shows up in adjacent places. When we covered YouTube's purge of faceless AI channels, the lesson was that platforms and audiences both move against content that has no one behind it, and they move after the publisher has already built a business on the assumption they would not. Our look at authenticity and synthetic voices made a similar point about audio: the technology gets good enough to fool people briefly, and trust damage lasts longer than the illusion.

Text is the same story with a lower barrier. Cloning a voice or producing a video takes some effort. Generating a 2,000 word post takes a sentence. When production is that cheap, the thing that becomes scarce is evidence of a person. A byline with a real history, a claim you can check, a mistake that got corrected. Those are the things AI drafting erodes first, because they cost effort, and effort is what the process was bought to remove.

The Second Problem: Claims Nobody Can Trace

Style is the visible half of the problem. The half that causes real harm is sourcing. Language models are good at producing a sentence shaped like a fact: a percentage, a "study found," a survey with a round number of respondents. If the draft came from a model, the author did not do the reading that produced those claims, which means the author cannot vouch for them. The numbers arrive already looking like citations.

Trakkr published a piece of research the day before Fowler's post under the name zombie stats. The term fits. A zombie stat keeps getting quoted long after anything alive stands behind it. We want to be careful with the numbers, because these are Trakkr's own findings, so treat them as vendor reported and not as an independent audit. With that label attached, they are still worth a close look. For the engineering side, including an eval that checks whether the cited source says what the answer claims, see our write up of the Trakkr study and RAG provenance.

Read that the other way. Trakkr's figure means that most of the statistics AI systems quote did not trace to a matching study. We would not claim the remainder are all fabricated, since the study does not say that, and a stat can fail to match for dull reasons: a source that moved, a number rounded or mangled in transit, a figure from one context quoted in another. The finding is about traceability, not about malice. It still means a reader who goes looking for the original has a better than even chance of coming back empty.

The second and third numbers are the ones aimed at us. In Trakkr's data, company blogs accounted for 54% of citations, and only 37% of those were sourced. Put plainly: more than half of what these systems cite comes from corporate blogs, and fewer than two in five of those blog claims point back to something. By our own arithmetic, roughly one citation in five is a company blog claim that also has a source behind it. That is our multiplication of two vendor figures, not a number Trakkr reported, so hold it loosely.

Here is why it matters beyond SEO. Company blogs are both the biggest source of cited claims and the weakest at backing them up. The loop is easy to see. A company publishes an AI drafted post with a plausible statistic and no link. Another model reads it, treats it as a source, and repeats the number. A second company writes a post citing the first. After a couple of rounds, the number has a lineage that looks respectable and bottoms out in nothing. Every step in that chain was cheap, and nobody in it was ever asked to produce the study.

Why "I Just Edit It Heavily" Does Not Work

The standard defense is that nobody publishes raw output. A human edits. This is the claim to push on, because it sounds responsible and it fails in a predictable way. Editing a draft and writing a draft use different mental muscles. When you edit, you react to the text in front of you. You fix awkward phrasing and trim the long sentences. You are very unlikely to ask whether the entire structure is the right one, because the structure is already there and looks plausible.

Anchoring does the rest. The model's first framing of the topic becomes your framing. Its list of three points becomes your three points. Its example becomes your example, even when you know a better one, because rewriting costs more than accepting. Fowler's warning about delegated thinking bites right here. The edits get made at the sentence level, and the thinking that mattered happened at the outline level, where nobody was looking.

There is also a plain mechanical effect. Light editing preserves the texture. You can swap out a dozen stock phrases and still leave the rhythm, because rhythm is spread across the whole piece and no single sentence carries it. A reader who feels the rhythm does not need to find the phrase.

Compare it with how a human actually builds an argument. You start with something you believe, usually a half formed objection to a thing you read. You write it badly. You notice the second paragraph contradicts the first and have to decide which one you mean. That friction is not a flaw in the process. It is the process, and it is why the result contains an opinion that survived being questioned. We made a related point about code in our look at the "almost right" problem: output that is nearly correct is more dangerous than output that is obviously wrong, because it invites a skim where a read was needed. Prose has the same failure mode.

What AI Is Good For in a Content Team

None of this is an argument for banning the tools. A blanket ban would be unenforceable and wrong. The question is where in the workflow the model sits. Our position is that it belongs on the input side and the audit side, and stays out of the sentence producing step.

Notice the asymmetry. Every use in the left column makes a human better informed or catches a human error. Every item in the right column replaces a human judgment. That is the test we apply to any proposed use: does the tool inform a decision a person still makes, or make the decision for them?

One caution on the left column. Using a model to find sources is only safe if a person opens each source. Models can return a convincing title and URL for something that does not exist, a failure we dug into when we looked at hallucinated packages in the software supply chain. The same mechanism that invents a package name invents a citation. Treat model suggested links as leads, never as evidence, and click every one.

The Economics People Skip

The pitch for AI drafted content is a cost curve. If a post takes a writer days and a model takes minutes, the spreadsheet says publish far more. That arithmetic holds only if all posts are worth the same, and they are not. Content value is heavily skewed. A handful of pieces with a real point of view earn the links, the shares, and the inbound conversations. Most of the rest earn nothing. Multiplying the volume of the part that earns nothing does not move the total.

Meanwhile the cost you did not put in the spreadsheet is correlated across posts. If readers decide your blog is filler, the good pieces get discounted along with the bad. You paid to make the average worse and the best work harder to find. That echoes the pattern we described in AI productivity theater: when you measure output volume, you get volume, and the measurement hides whether anyone benefited.

A better ledger counts the cost of a wrong claim. A single invented statistic in a post that gets picked up can mean a public correction, a sales call that opens with an awkward question, or a prospect who quietly checks the other numbers on your site. We cannot give you a dollar figure for that, because it depends on your market and nobody has published a credible one. We can say the downside is lopsided: sourcing a claim costs minutes, and defending an unsourced one costs reputation.

The Counterarguments, Taken Seriously

A fair version of this argument has to deal with the best objections, so here are four.

First: "Plenty of human written content is also bad." True, and it is the strongest objection. Human prose has its own filler, its own stock phrases, and its own unsourced numbers; Trakkr's finding about company blogs was a finding about blogs that people have been writing for years. The point is not that humans are good and machines are bad. It is that a person who writes has to decide what they think and answer for it, and a policy that keeps that step intact gives you a floor the other approach lacks.

Second: "Non native English speakers rely on these tools." That is a real use and a good one, and it fits the left column. A writer who drafts in their own language or in rough English and uses a model to translate, tighten grammar, and flag unclear sentences is still the author of the argument. The arguments and the claims stay theirs. Fowler's concern is delegating the thinking, and this does not.

Third: "Readers cannot actually tell." Some cannot, some of the time. Our claim is narrower than the headline suggests: the readers you most want, people who evaluate vendors and read carefully, are also the ones with the most exposure to model output and the lowest tolerance for it. You do not need every reader to notice. You need the few who decide to notice, and they tell their colleagues.

Fourth: "We do not have the time." Then publish less. One post a month with a point of view beats eight that read like filler, on the evidence of what actually gets cited and shared.

What a Human Written Core Looks Like in Practice

"Write it yourself" is advice nobody can act on without detail. Here is how the workflow looks for a piece like this one.

It starts with an assertion on paper, in one or two sentences, a person can disagree with. "AI is changing content" fails the test. "Companies that publish AI drafted posts pay for it in trust, and the fix is a sourcing rule" passes. If you cannot write the claim in your own words before you open any tool, you are not ready to draft, and no tool will repair that for you.

Next, research. This is where a model earns its keep. Ask it to find reports, list counterarguments, and summarise long documents. Then read the sources yourself, at least the passages you plan to quote, and keep a note of exactly what each one says. Keep the rule simple: every number links to a page a person read. That makes for a short list of numbers, which is the point. A post with three numbers you can defend beats one with fifteen you cannot.

Then comes drafting, by a person, in their own voice, fast and badly. Specific beats general at every opportunity. Name the system, the date, the person, the thing that broke. If you find yourself writing a sentence that could appear in any other article, either add the detail only you have or delete the sentence. Resist the urge to start from an outline a model produced, since the outline is where the thinking lives.

After that, the audit pass, where AI is useful again. Hand the finished draft to a model with a narrow instruction: list every factual claim, quote, and number, and mark which ones have no link. Ask it to identify sentences that contradict each other. Ask it to argue against the thesis. You are using it as an adversarial proofreader, and you keep or discard each suggestion yourself.

Last, a human reads it aloud. It sounds like a gimmick and it works. Stock phrases are far easier to hear than to see. The place you run out of breath or wince is usually the place a sentence is doing nothing.

A Voice Test You Can Run in Ten Minutes

If you inherit a content library and want to know how exposed it is, you do not need a detector. Pull ten posts and answer five questions about each. Does it contain a number with a link? Does it name a person or a date that the author could be asked about? Does it take a position someone could reasonably oppose? Is there one detail that suggests the author did the thing being described? Could the same text appear under a competitor's logo unchanged?

Score honestly. Posts that fail four or five checks are the ones to rewrite or delete first. Deleting is underrated. A library of forty strong posts reads as expertise, and the same library with a hundred filler pieces mixed in reads as a content farm. Redirect the weak ones to the nearest good post and move on.

An Editorial Policy You Can Paste Into Your Team Doc

This is the practical part. Copy it, change the names, and argue about the parts you disagree with. A policy people have argued about is more useful than one they were handed.

Two notes on enforcement. The sourcing rules can be checked mechanically in review: a reviewer scans for numbers and clicks the links. The writing rules cannot, so they depend on trust inside the team, and that works only if nobody is punished for publishing less. If the calendar is sacred and the policy is aspirational, the calendar wins every time.

What to Do This Week

None of this needs a reorganisation. Start with five moves, in order, and expect most of them to take under an hour.

• Read Fowler's LLM Voice with your content team, then ask each person where in their current workflow a model writes sentences. Do not skip the ones who answer "nowhere." Check anyway.
• Run the ten post voice test above on your most visited pages. Pick the two worst and rewrite or redirect them before you publish anything new.
• Audit numbers on your top ten pages. Add a link to every statistic you can source, label the vendor figures, and remove the ones you cannot trace. Do it with a model's help if you like, but open every link yourself.
• Adopt the policy above, or your edit of it, in the same doc where your editorial calendar lives. Policies in a wiki nobody opens do not exist.
• Cut the next publishing quarter's target by a third and put the saved time into one piece with a strong opinion and a named author. Measure whether anyone replies to it.

If you are building tooling around any of this, the same split applies. An internal assistant that retrieves and cites sources for a writer is a research tool. One that produces a finished article is a different product with a different risk. Our recommendation is to treat the first as infrastructure and keep the second behind a human sign off, much like the discipline in our piece on performative AI strategy, where measured outcomes beat visible activity.

The reason to act now rather than next year is that the reputational ledger is lagging. Posts you publish today get read and cited for years, and readers' pattern recognition only improves. Content that passes as human this quarter may read as obvious filler in twelve months, after the reader has seen ten thousand more samples. A sourced, opinionated, human written post does not have that expiry problem.