How to Tell AI Writing From Human Writing

Most advice on spotting AI writing is a list of words. Delve, tapestry, in the realm of, a testament to. Ban those in one line of a prompt and the list is useless, which is roughly what has already happened.

The words were never the tell. They were the easy tell, and easy tells get patched. What is left is harder to fake and far more reliable, because it is not about how the writing sounds. It is about what the writer actually knew.

The Surface Tells Have Already Stopped Working

The famous ones are all about vocabulary. Certain words models overuse. The three-item list where two would do. The perfectly even paragraphs. The “in conclusion” that concludes nothing. Punctuation habits.

These are real patterns. They are also the first thing anyone editing AI output learns to strip, and the first thing a decent prompt suppresses. A tell that vanishes the moment the writer is told about it is not a tell you can lean on. The word-list approach was already a season out of date by the time it went viral.

Treat surface features as a weak prior, not proof. A page full of the tell-tale words is worth a second look. A page with none of them can still be entirely machine-made. Whether an AI detector can settle it runs into the same problem from the other direction, and the answer is less reassuring than most people want.

The Tells That Survive a Good Prompt

There are three, and they are all absences rather than features. You are not looking for a word so much as for something that should be there and is not.

Specificity. Firsthand experience. A stance the writer is willing to defend.

A model can imitate the shape of all three. It can write a sentence that looks specific, an anecdote that looks lived, an opinion that looks held. What it cannot do reliably is supply the underlying thing, because the underlying thing is knowledge it does not have. That is the difference the AI versus human writing question turns on, and it is why the tells worth learning are the boring structural ones.

It Never Names Anything

The clearest tell is category language sitting where an instance belongs.

AI writing reaches for the general noun. “A range of factors.” “Various tools.” “Several considerations.” “Industry best practices.” It hovers one level of abstraction above the actual thing, because the actual thing is a detail it would have to know. A person who has run the process names the tool, the month it broke, the number it landed on.

This is the same property that makes writing rank or fail. A page built from category language adds nothing to what already ranks for the query, which is exactly what information gain measures.

Read a suspect paragraph and ask one question. Could I check any claim in this? If every sentence is technically true in a way that is impossible to verify, because it never commits to a specific, you are probably reading a machine, or a person writing like one.

Nothing in It Could Only Be Known by Doing It

The second tell is the absence of firsthand experience, and it is the hardest of the three to fake.

Experience leaves residue. The detail too specific to have been guessed. The aside about the thing that went wrong the second time. The preference that only makes sense once you have hit the edge case. None of it is decorative. Each is the byproduct of having actually done the thing, and a model has done none of the things it writes about.

So AI writing describes competently and remembers nothing. It can tell you what a process involves. It cannot tell you the part of the process nobody warns you about, because that part is not in the summary it learned from. This is the same signal search now weights heavily under experience, the first E in E-E-A-T and named authors, and the one that is hardest to manufacture at scale.

The Hedge and the Sameness

Two smaller tells that travel together.

The hedge. AI writing avoids commitment. It gives you both sides, notes that it depends, weighs pros against cons, and lands on a conclusion so balanced it says nothing. This is not caution. It is the absence of a view. A writer with a real opinion states it early and spends the rest of the piece earning it. A model optimizing to be defensible never gets there.

The sameness. Put ten AI articles on one topic side by side and they converge on the same structure, the same subheadings, the same order, the same even spread of attention across every point. A person writes lopsided, because some parts of a topic are more interesting than others and they know which. That flatness is why this kind of writing plateaus in search rather than failing outright. It is competent, uniform, and interchangeable, which is precisely the profile that no longer moves.

The AI Tell	The Human Signal
“Various tools”	Names the tool
A vague range	A specific number
An anecdote that fits anyone	A detail that could only be firsthand
Both sides, no conclusion	A stance stated early and defended
Even, symmetrical sections	Lopsided attention to what matters
Describes the process	Remembers the part that went wrong

How to Read for It in Practice

A short procedure that works faster than any detector.

• Pick the three most specific-sounding claims and try to verify one. If none commit to a checkable detail, that is your answer.
• Look for a single sentence that could only have been written by someone who did the thing. If the whole piece could have been written from the search results, it probably was.
• Find the stance. If the piece refuses to prefer anything to anything, note it.
• Check the shape. Even paragraphs, symmetrical sections, and a tidy three-item rhythm all the way through are structural tells, not stylistic ones.
• Ignore the vocabulary. Whether it says "delve" tells you almost nothing now.

None of these is conclusive alone. Together they are far more reliable than a word list or a detector score, because they test for knowledge rather than phrasing, and knowledge is the thing that cannot be prompted in.

Frequently Asked Questions

Can you tell AI writing from human writing just by reading it? Often, but not from the words. The reliable signals are structural. Whether the piece commits to specifics, whether it contains anything that could only be known firsthand, and whether it takes a position. A polished piece with none of those is the giveaway, regardless of vocabulary.

Are words like “delve” a reliable sign of AI? No. They were a weak signal a year ago and they are close to useless now, because a single line in a prompt removes them. Treat overused vocabulary as a reason to look closer, never as proof.

What is the single most reliable tell? Specificity, or the lack of it. AI writing names categories where a person names instances. If a piece never commits to a checkable detail, a specific tool, a specific number, a specific month, it reads as machine-made even when every sentence is grammatically perfect.

Can AI writing be edited to pass as human? The surface can. A good editor can strip the vocabulary and smooth the rhythm. What editing cannot add is firsthand experience the writer does not have. That is why the strongest human signal is a detail or a judgment that had to come from doing the thing.

Does it matter if the writing is AI, as long as it reads well? It matters where the piece has to rank or carry authority, because the same absences that reveal the source also cap the performance. Writing that names nothing and knows nothing tends to plateau. Whether it is good enough comes down to the job it has to do.

How Postdex Handles This

Before an article lists, a named editor verifies it against the things a reader uses to tell the difference. Real specificity rather than category language. A point of view rather than a hedge. Claims that hold up rather than sentences that merely read cleanly.

That is the practical meaning of human written here. A person made the choices a template would not, the example that had to be lived, the number that had to be looked up, the stance that had to be argued. It is also why the writing tends to keep working, the same reason human written content still wins for SEO.

The catalog lists finished pieces you can read before you buy, so you can apply every tell above to the actual text. If you need something specific, the commissioning desk takes the brief. Either way, what ships is writing that names things, knows things, and is willing to take a side.