How to Tell if Content Was Written by AI
There is no definitive indicator. All you can do is combine multiple minor indicators. These include factual texture, structural consistency, and the layers of supplementary documentation that accompany a draft, until the image is sufficiently precise to warrant action.
How to Tell if Content Was Written by AI: The Basics
Two different questions are being conflated, and they require distinct evidence. “Was this text generated?” is a question about the past: what occurred in the hours prior to the saving of the document. “Does this text resemble generated text?” is a question that relates to the current text being evaluated. Among the two questions, only the second one is accessible to a reader, and it is by far the weaker question.
A human author is more likely than a machine to cite specific instances. A draft written by a machine will probably describe a category. When I spend the morning writing a piece, people would be able to notice by the way I end prices (e.g. 95 cents), the way a message is worded (e.g. an error message), possibly the name of the person who called, and maybe even a date which would have been better to avoid. In contrast, machine written text is super fluent, but it is specific about next to nothing.
Next, consider form. Paragraphs are nearly identically sized. Each segment summarizes into three ideas. Transitions come from a limited rotating set. Sentences are constructed to the same level of complexity, so writing will never have variance with pace. A person drafting under a time constraint produces something much less uniform, and typically editors will not smooth something out when the lumps have meaning.
- Claims that sound checkable but resist checking: no source, no date, no number
- References and attribution to a person that cannot be traced back to the original stated citation
- Even paragraph lengths and an unbroken three-beat rhythm across every section
- Hedged, symmetrical treatment of a question the writer should have an opinion about
- Recently discussed topics using an out-of-date or stale approach, or not discussed at all
Why How to Tell if Content Was Written by AI Matters
It’s a costly affair when decisions go wrong, and that is true of calling the judgment when it spikes a commissioned article. An editor, a lecturer who opens a case of misconduct, or a client who refuses to pay, all cost someone real given the strength of the hunch about sentence rhythm. Each of those calling this judgment owes the writer something more concrete than “it felt smooth.”
This is the most common type of case. Very few examples of professional content are straight raw output from a prompt. Much more of it is machine-generated, rearranged, fact-checked, and partly rewritten by someone competent. This mixture almost completely conceals reading signals because the human pass restores the exact irregularity that you were looking for, while the underlying structure remains machine-generated.
A secondary, often ignored, cost occurs for the writer too. From the belief that the surface style of their writing will be judged, some writers begin to purposely destroy their own style even further through longer, messier sentences, unbalanced structures, or undesired vagueness. Punishing writers for clear prose results in prose that is even less clear. This is an unfortunate trade-off for any publication and a frustrating experience for any writer.
A reading signal is a reason to ask, not a reason to conclude
Every tell on this page has a perfectly innocent explanation. An example could be uniform paragraphs which could indicate a careful sub-editor. Vague language could indicate a brief that is far too brief. Treat a suspicious mind as the first sign to engage in a dialogue with the writer, rather than closing the book on the piece.
How Articled Approaches How to Tell if Content Was Written by AI
For ourselves, we don’t use any of these. We use provenance. That’s the dull answer and the only one that stands the test of time. Writers keep drafts, notes, source lists, and whatever they recorded along the way. If a client asks how a paragraph was created, there is a record of it, and the writer can explain the thought process behind their work.
Detection tools are always tool number two, never tool number one. Before going to clients, every draft goes through Originality.ai, GPTZero, Turnitin, Copyleaks, Winston AI, ZeroGPT, Sapling, Content at Scale, Writer.com AI Detector, Crossplag, Grammarly Authorship and Copyscape. The reports are attached to the email with the delivery. They show what the twelve tools said the day we ran them. It is useful to have that record, and it is not a certification.
There is one fundamental rule. We will ban anyone caught running text via an AI humanizer to make a detector behave. It will terminate your interaction with us on this platform. Humanizers will only create a surface level change, which is exactly what you have now been told you can not rely on. If you put an AI thinking process in place while keeping text reading generation AI, it would make our verification reports a lie.
- Before the piece is finalized, request the working drafts, with timestamps
- Request the writer clarify an edit: what was removed and why did it not survive
- Inquire about the claims they were unable to verify, and what actions they took as a result
- Request the source list, then verify two sources at random
- Ask what they learned that was unexpected during the research
How to Tell if Content Was Written by AI FAQs
Indicators are not proof. Only evidence about the way a draft was made can do that. We keep that evidence for all orders, since an AI-free promise is hollow if no one can substantiate it.
Indicators are not proof. Only evidence about the way a draft was made can do that. We keep that evidence for all orders, since an AI-free promise is hollow if no one can substantiate it.
From the text alone, no. A detector gives a probability, and a reader gives an impression; neither one is proof. Proof comes from provenance: version history, drafts, research notes, and a writer who can answer questions about their own reasoning. Ask for these first.
They decay quickly. A good editor reintroduces variety to sentences, adds detail, and removes hedge language. This removes most of the surface level indicators while leaving generated structure. Mixed drafts are the hardest to evaluate and the most common.
No, and don’t buy one. Every device on the usual list, from em dashes to the tidy three-part structure to an appreciation of “moreover,” is present in the writing of normal, everyday humans, especially after passing through a house style guide.
Rather than ask with a hostile undertone, ask process questions. Which source was challenging to pin down? What did you cut? Why did you cut it? What claim are you least confident about? Someone who did the work would be quick to respond and would give their answer in detail. Those who did not are more likely to answer in summary form.
This means one classifier did not flag it. This does not mean accurate, original, well-sourced or worth reading. Fact-checking, plagiarism screening, and an editor’s judgment are separate tasks, and a clean AI score does not substitute any of these.
If you are commissioning writing rather than policing it, the practical step is to make provenance part of the deal at the start. Agree that drafts and notes are part of the delivery. That single line in a brief removes almost every argument you would otherwise be having about sentence rhythm six weeks later.
Content a person actually wrote
$10 per 100 words, and the writer keeps all of it. Our 1% sits on top, 0.5% goes to trees, and no generated text appears anywhere in the process.