A vendor in this industry published a guide on what it costs to launch a prop firm. It is live right now. One of its section headings reads “Evidence gap: Refresh or deepen the existing coverage” and then names a keyword the site wants on page one. Same font as every other heading. Sitting there like it belongs. That is not a heading. That is an instruction written for a writer, published as a heading, because nobody read the page before it went out. That page is a guide about hidden costs, and it is currently the best available evidence that the biggest hidden cost is the writer.
It has been live for weeks. Founders researching a six-figure decision have been reading a note to a content team instead of an answer.
Here is the useful part. Most advice about spotting AI writing points at the wrong things, so people burn their editing time deleting em dashes while the actual damage sits three paragraphs further down. The loud artefacts are not the common ones, and the common ones are the ones doing the harm.
What unedited AI content looks like when it goes live
Unedited AI content is content published with the model’s own scaffolding still inside it. Placeholder tokens, refusal messages, instructions from the prompt, and statements about the model’s own uncertainty all survive because the person who generated the text never read it.
The clearest public example is retail. Sellers automating product listings ended up with catalogue pages whose titles were the model refusing to write a title. The Washington Post documented a chest of drawers named after a refusal message, and found more listings of the same kind still live after the first batch was pulled.
Print is not immune. In November 2025 a Pakistani national daily ran AI editing artefacts in a physical edition and published an apology. Earlier that year the Chicago Sun-Times ran a summer reading list containing books that do not exist, which Poynter listed among the year’s worst newsroom AI failures. Press Gazette now maintains a running tracker of AI mistakes in journalism, which is a sentence that would have sounded like satire in 2022.
Which AI tells actually show up most often
Academic publishing is the only place anyone has counted this properly. A 2024 paper describing the Academ-AI repository analysed the first 500 collected cases of apparently undeclared AI text in published research and measured how often each artefact appeared.
| Artefact | Share of 500 papers | What it actually is |
|---|---|---|
| The model writing as “I” | 55.2% | Chatbot voice sitting inside a multi-author paper |
| A reference to its own training cutoff | 49.0% | The model warning you it might be out of date |
| A “certainly, here is” preamble | 31.8% | The reply to the prompt, pasted along with the answer |
| Pointing the reader to newer sources | 20.6% | The model handing the research back to you |
| Addressing the reader as “you” | 13.8% | Customer service voice in a methods section |
| The “regenerate response” button label | 11.6% | A piece of the interface copied off the screen |
| Identifying itself as a language model | 8.6% | The famous one |
| Stating it has no real-time access | 8.6% | The model declining the job it was given |
The artefact everyone quotes finishes last, which makes sense once you think about it. “As an AI language model” is the phrase every editor on earth now knows to search for, so it is the Nigerian prince email of AI tells. Famous enough that it barely catches anyone any more.
The two at the top are both the model talking about its own limits. Nearly half of those papers contained something like the phrase “as of my last knowledge update”, published, in peer-reviewed research. So the artefact most likely to be sitting on a page right now is not a placeholder or a refusal. It is the model hedging about how much it actually knows, left in by someone who never read the hedge.
The obvious catch: those percentages come from journal articles, not prop firm blogs. Nobody has run the same count on marketing content. Partly because nobody retracts a blog post.
Around 3% of the cases in that dataset were ever corrected after publication. The other 97% are still sitting there.
Which brings us back to the costs guide I opened with, because that heading was not the only thing on it.
I found the page because it cites me. It quotes my article on starting a prop firm twice, credits me by name, and links correctly. All fine. The rest of it is a content brief with the lid left off.
Once you notice the heading that tells a writer what to do, the rest of the page stops hiding. Two other headings end in a template slot label about the reader’s decision. The opening summary finishes with the target keyword itself, lowercase, dropped in as a fragment after a full stop. The meta description ends with a variable that was supposed to be swapped for the brand name and was left as lowercase filler instead. The page numbers its inline citations up to eight and lists four sources.
Then there are the disclaimers, which are exactly the artefact the table above puts at the top. Paragraph after paragraph tells the reader that the evidence is descriptive rather than quantitative, that the ranges are directional, that no dollar figure should be extrapolated. That is a model hedging about its own inputs, and nobody deleted it, on a page whose entire purpose is telling a founder what something costs.
On names: I am leaving the vendor out of this. The pattern is the point, and any operator who wants live examples can run the searches at the end of this article and find a dozen before lunch.
Cosmetic AI tells versus structural AI tells
Cosmetic AI tells are style habits: em dashes, curly quotes, “delve”, tidy groups of three, and the “it is not X, it is Y” construction. Structural AI tells are artefacts of the generation process itself. Only the second kind proves that no human read the page before it went live.
| Cosmetic tell | Structural artefact | |
|---|---|---|
| Example | Em dash, curly apostrophe, “delve”, “in today’s fast-paced world” | “[SOURCE NEEDED]”, “As an AI language model”, “Certainly, here is”, a content brief instruction published as a section heading, a citation numbered higher than the source list goes |
| What it proves | A model was probably involved somewhere | Nobody read the page between generation and publication |
| What a reader concludes | Nothing. Most readers do not notice. | This company publishes things nobody checks |
| What it costs you | Roughly zero | Trust, on a page a trader reads before handing over a card number |
| Worth your editing time | Only if house style says so | Every time, without exception |
The em dash panic is a waste of your editing budget
Punctuation is not evidence. Editors have used dashes for a century. Even Wikipedia’s community guide to signs of AI writing warns against this trap. It points out that curly quotation marks prove nothing, because any publication following the Chicago Manual of Style produces the same characters.
A reader who spots an em dash thinks nothing at all. A reader who finds a note to the content team where a heading should be learns that your company ships pages nobody reads. Then he starts wondering what else nobody reads. In this industry the answer is usually the payout terms.
Why prop firms publish AI content nobody read
Prop firms publish unedited AI content because nobody reads it, including the person who made it. Generating a draft feels like writing. The job feels finished. Publish is one click away.
The standard advice here is hire an editor. Ignore that. You do not need an editor, and telling a five-person firm to add headcount is how advice gets ignored.
Every artefact in this article would have been caught by the person who produced it, reading their own page once, before publishing. Not proofreading. Not a style pass. Reading it, the way a stranger would. That is the whole fix. It is free, it takes ten minutes, and it is apparently the hardest thing in this industry to arrange.
The reason it does not happen is that nobody thinks of it as a step. Look at a prop firm org chart. There is a CMO, a media buyer, an affiliate manager, a community lead, and one content writer on contract. Nobody has “read the blog” in their job description, including the person writing the blog.
The blog is also treated as a chore rather than a product surface. The dashboard gets QA. The rules page gets legal review. The blog gets a monthly target of twenty posts and a contractor who has quietly become a prompt operator. At twenty posts a month, reading them all is a day of work nobody has budgeted, so it does not happen, so the artefacts ship.
The tell behind the tell: when a firm ships an obvious artefact, the artefact is not the problem. The absent review step is the problem, and it is absent on every other page too.
The Academ-AI paper has the perfect illustration of this. One article was flagged because somebody had copied the “regenerate response” button label into the reference list. Copying a button is harmless. But it prompted somebody to actually read the bibliography, and 18 of the 76 references turned out not to exist. The button label did no damage at all. It just happened to be the only visible symptom of a page that nobody, through writing, peer review, copyediting and typesetting, had checked.
That is what an artefact is worth. It is not the injury. It is the bruise that tells you where to press.
You paid for cheap content and cheap content arrived on time
The second cause is pricing. A model does the writing now, so the writer should cost less. That is the logic, and it is not stupid. But the vendor who agrees to your new price has to make the maths work somewhere, and you do not get to choose where.
Generation got cheap. Review did not. Reading a draft, checking every rule against a help centre, clicking every link, and cutting the paragraphs that say nothing takes about the same hour it took in 2019. When you pay twenty dollars for an article, you have bought the half that got cheap and refused to pay for the half that did not. That is the equivalent of a restaurant printing “TODO: find out if this is chicken” on the menu, in the same font as the prices, and staying open.
Low-cost agencies solve this with templates and volume, which is how a section heading ends up with a slot label about the reader’s decision still attached to it. Somebody built a section template, ran forty keywords through it, and shipped. Blame the template if you like, but the template did its job. Nobody opened the output, and at that price nobody was ever going to.
The saving is also fake. You pay once for the article. You pay again for the audit that finds the problem, and again for the rewrite. Between those three invoices the page sits live in front of the operators you are trying to sell to, doing its work for free. Cheap content is the most expensive kind. It just arrives faster, which is how it keeps getting bought.
Spending more is not automatically the fix either, and the Academ-AI paper is oddly reassuring on this point. The journals that published undeclared AI text charged higher processing fees and had higher citation scores than their peers. You can pay three thousand dollars for editorial processing and still end up with a chatbot apologising in your methods section. What you are buying is not a price tier. It is whether one person opens the file.
What unedited AI content costs a prop firm
Unedited AI content costs a prop firm trust before it costs rankings. A trader choosing between firms reads the blog as evidence of operational care, and a page carrying a visible placeholder is evidence pointing the other way.
Traders in this category arrive suspicious already. They have watched firms close without warning and payouts stall, and they read everything you publish as a signal about whether you are a real operation. A page that admits its own numbers are unsourced estimates answers that question for them.
There is a search cost underneath the trust cost. No penalty exists for using a model, but an unread page is a weak page, which is the same standard that applies to every other part of prop firm SEO. Answer engines make it sharper. When a model builds an answer about drawdown rules or payout timing, it picks a source. A page that states its own figures are guesses is a source it will not pick, and getting cited depends on being the page that explains the messy thing properly.
The commercial cost is quieter. Affiliates, review sites, and B2B partners screenshot this material. A one-pager with a placeholder in it circulates in operator group chats for a week. That does more damage than any bad ad, because nothing in prop firm marketing kills trust faster than looking careless. Nobody screenshots your good pages.
How to catch AI artifacts before you publish
Catching AI artefacts takes about ninety seconds and needs no detection tool. You are not trying to work out whether a model wrote the draft. You are checking whether a person read it, which is a much easier question with a much more useful answer.
- Search for the strings that actually show up: “my last”, “training data”, “cutoff”, “up to date”, “real-time”, “certainly”, “regenerate response”, “I cannot”, “as an AI”. Start at the top of that list, not the bottom.
- Search the draft for square brackets. Placeholders almost always live inside them.
- Search for “TODO”, “TBD”, “insert”, “XX”, and “SOURCE”.
- Read the first sentence out loud. “Imagine this” and “in today’s fast-paced world” openers mean nobody touched the top of the page, and the top of the page is what gets read.
- Read your headings on their own. If one reads like an instruction to a writer, it was one.
- Click every link in the draft. Invented URLs are a common fabricated fact and the fastest to catch.
- Count your citations against your source list. Inline references pointing past the end of the list mean the sources were assembled by something that was not counting.
- Check every number has a source you can open. If the draft calls something an estimate, either source it or cut it.
- Check every rule against the firm’s own help centre. Models describe trailing drawdown and consistency rules confidently and wrongly.
- Read the page yourself before you publish it. Not skim. Read. This one catches everything above and you keep skipping it.
Detection tools do not help here. They return a probability score about authorship, which is not the thing you need to know, and they flag careful human writing often enough to be useless as a gate. The bracket search is more reliable than any of them.
What to do about the AI content already live on your site
Start with a site search rather than a full content audit. Run site:yourdomain.com in Google with each giveaway string in quotes, and you will find the worst pages in a few minutes without touching a crawler.
Check the formats nobody thinks about. Sales decks, PDF one-pagers, partner brochures, and pitch documents get less review than blogs, because they are never indexed, never re-read, and sent directly to the people whose opinion of you matters most. A blog post might get caught by a stranger. A deck goes straight to the buyer, who reads every slide, because he is deciding whether to give you money.
Fix first
Pages with traffic, pages a partner or affiliate would read, and anything containing rules, fees, or payout claims.
Fix or delete
Old volume posts with no traffic and no sources. Unpublishing is cheaper than rewriting and usually the better call.
Then fix the process, or the artefacts come back next month. The rule is one line long. Whoever writes the page reads the page, start to finish, before it goes live. If you are building a firm from scratch, put that in now, while your content backlog is still zero.
Good work is not cheap and cheap work is not good
Every artefact in this article is somebody trying to buy the reading hour at a discount. That hour has not got any cheaper since 2019, whatever a model does with the first draft, and a rate that assumes otherwise buys you a draft rather than a page. I do not compete on price and I have written up what this work actually costs, so you can decide what you are budgeting for before you brief anyone.
What prop firm SEO actually costs →FAQs about unedited AI content
What is the most common sign of unedited AI content?
The model referring to itself, either in the first person or by flagging its own training cutoff. Across 500 documented cases in academic publishing, first-person chatbot voice appeared in 55.2% and a reference to the knowledge cutoff in 49.0%. Self-identification as a language model appeared in 8.6%, making the famous tell the rarest one measured.
Is an em dash proof that AI wrote something?
An em dash is a punctuation choice that editors have used for over a century, and it proves nothing on its own. Placeholder text, refusal messages, and broken citations are the signals worth acting on, because those only appear when nobody read the page.
Do AI detectors work well enough to use as a publishing gate?
AI detectors return a probability about authorship, which is a different question from whether a page is accurate and reviewed. Searching a draft for square brackets, refusal strings, and dead links catches the damaging problems faster and with fewer false alarms.
Should a prop firm use AI for blog content at all?
Using a model to draft is normal and it is not the thing that damages a firm. Publishing without a named human reviewer is the thing that damages a firm, and that risk exists whether or not a model was involved.
Is cheap content the reason this keeps happening?
Cheap content is a large part of it. Generating a draft got cheap and reviewing one did not, so a budget set on the assumption that a model does the work buys the generation and skips the review. The visible artefacts are what that trade-off looks like once it is published.
How do I find this problem on my own site quickly?
Run a site: search on your domain with each giveaway phrase in quotes, then repeat the check on your sales decks and PDF one-pagers. Documents sent to partners are the least reviewed material most firms produce.
