Human or AI? Test Whether You Can Actually Tell Who Wrote It

Key Pointers

• Many people think they are better at the task than they really are. Research studies have shown that untrained human accuracy is at a coin flip level, and often people who are the most confident in their judgment are not the most accurate.
• The clues that are useful do not come from words alone, but from structure. The rhythm and structure of the text are much more informative than the specific words used in the text.
• An em dash is not a clue; neither is the word “delve”. These are simply stylistic choices.
• Any section flagged as AI in this article was created specifically for that article, and you will see the prompt that led to its creation.
• Reading a piece of writing and testing it is very different. One tells you how good the piece feels, while the other gives you a probability.

What “human or AI” actually means in 2026

How binary the question sounds. Yet it turns out that hardly anything you read is actually on one side or the other. A paragraph may be written by a human and edited by a machine. It can also be generated from scratch but then polished till the seams are invisible. It can also be crafted solely by human hands of a person, who pays enough attention to AI outputs to have their writing style shaped by them. All of this is common, and none of these is “A.I. written” in the sense in which the question states. So the trick is not to label the extract correctly. The trick is to figure out what specific component makes the text produce the wrong feelings and how much weight each of the components carries.

That is the purpose of the sections that follow.

Round one: human or AI?

Read this passage. Decide before you scroll.

The transition to remote work has fundamentally reshaped how organizations approach collaboration. Teams that once relied on spontaneous hallway conversations now depend on structured check-ins and asynchronous updates. This shift has brought both significant advantages and notable challenges. On one hand, employees report greater flexibility and reduced commuting stress. On the other, many describe a persistent sense of disconnection from colleagues. Understanding this balance is essential for leaders navigating the modern workplace.

It’s AI. I generated it for this article with the prompt “Write a short paragraph about remote work and collaboration in a neutral professional register.”

Here’s what gave it away, in order of reliability.

Six sentences, and five of them run between 14 and 19 words. Human paragraphs are lumpier than that. Someone writing naturally produces a four-word sentence and then a thirty-word one, because they’re thinking, not balancing.

The argument is perfectly symmetrical. “On one hand” is answered by “On the other” with equal weight given to each. Real writers have opinions and the proportions show it.

And the last sentence does the thing models almost always do: it restates the topic as a conclusion without adding anything. “Understanding this balance is essential” tells you nothing you didn’t know six sentences ago.

Notice what isn’t on that list. No odd word choices. No factual errors. Nothing you could point at and call a mistake.

The five tells that still hold up

Word lists are the most common source of AI indicators; they are the first to vanish alongside the model. However, these five examples are distinctive, which is why they have lived for four years since the launch of the model.

No variations in sentence length

If you scan the first column, you’ll see how many sentences there are. A machine generates the same number. Human writings have sentences of various lengths.

This is the most reliable of the indicators since it remains even after editing. You need to change not simply the wording, but also the rhythm of the sentences to correct this.

Uniform paragraphs

Three sentences in one paragraph but three in another. In reality, a paragraph might contain one up to six lines or just a phrase, depending on how it fits the schedule.

Conclusions restate rather than summarise

A machine model gives the same answer. In other words, a machine model will restate the initial question, while someone will make a concrete point at last.

Balance instead of opinion

A machine goes for the middle ground since it is the safest option. The phrases like “there are advantages and disadvantages” are rarely found in human texts.

Imprecise wording

AI text refers to categories like “many companies”, “studies show”, or “experts believe”. If you want these broken down further with examples, we’ve covered the full set in our guide to how to tell if something was written by AI.

Round two: a harder human or AI test

Same instruction. Decide before you read on.

I kept the receipt for eleven months. Not deliberately, it just migrated between coat pockets until it turned up in a jacket I hadn’t worn since March, soft as tissue and barely legible. £4.20. A coffee and something I can’t read anymore. I don’t know why I didn’t throw it out. I think I wanted proof that the afternoon had happened.

Also AI. Prompt: “Write a short first-person paragraph about finding an old receipt, in a literary register, with varied sentence lengths and a specific detail.”

This one is harder, and the reason is worth understanding. Every structural tell from round one has been deliberately prompted away. Sentence lengths vary from four words to twenty-six. There’s a specific number. There’s a fragment. The ending lands rather than summarises.

What’s left is thinner and harder to articulate. The emotional beat arrives exactly where the structure wants it. The specific detail, £4.20, is doing conspicuous work. It’s placed to prove authenticity rather than because it mattered. Nothing in the paragraph resists the point it’s making.

That’s the honest state of this in 2026. The tells work on default output and degrade fast against a prompt written by someone who knows what they’re doing.

How to score your human or AI guesses

Two rounds is insufficient data to give you any information, so consider this as calibration, not result.

Did you get both right? You are reading structurally, which is good. Try this with some longer pieces that will have more rhythm to feel different about.

Did you get round one but not round two? That is what happens most of the time. Round two is adversarial.

Did you miss both? Needs to be said about it. Many people overestimate their ability here, and the gap between their confidence and the actual accuracy is the spot where difficulties start.

Why people get this wrong so often

Three major types of failure account for most of it. The first is that people have a tendency to trust vocabulary. There is em-dash theory (“delve”), the theory that emoji or neat parallelism provides evidence that a text was produced by a machine. This is a matter of training data habits, not fingerprints; the same applies to many careful human authors. Editors, lawyers and academics use em-dashes all the time. The next issue is reading cleanliness as evidence.

Good style is polished, formal, and low-variance writing, which is what good academic writing is supposed to be. However, this is what the models produce, too. This is the largest driver of false accusations and it hits people who write in foreign languages the hardest. The third failure mode is confirmation bias.

Once someone suspects a text is suspicious, every characteristic of the text is read as supporting that suspicion. The structure is too nice; the vocabulary is too sophisticated. There is no way to disprove the suspicion that was formed before the text was read. The problem is not purely philosophical in nature. A study led by Stanford in 2023 showed that detectors wrongly flagged more than 61% of authentic TOEFL essays written by foreigners, and a replication study conducted in February 2026 showed that this figure dropped to 23.1%, but that there was no improvement in terms of false positives. We’ve written about that pattern and who it lands on in detector bias against ESL students.

Test it on something real

Two paragraphs composed for an article convention serve the purpose of demonstration, not of measurement. The practical side of this task utilizes material present in any case.

Something written during the last week shall be chosen. It can be either a long email, LinkedIn post, and blog entry, or a post from the discussion board. Application of the five rules in order consists of the following: variability of sentence structure, symmetry of paragraphs, concluding sentences, view taken, evidence and proof.

Finally, a conclusion and a certainty have to be written, and then a tool has to be applied for checking the result instead. Run the same passage through our AI content detector and compare the score to your own verdict.

When the two disagree, that disagreement is the interesting part. A detector reads statistical predictability. You read rhythm, intent and whether the writing has a point of view. Neither has access to who actually typed it.

What a verdict is worth

Here’s the part that matters more than the score.

Nothing in this article establishes authorship, and neither does any detector. A score is a probability about how text reads, not evidence about who produced it. We say this plainly even though we sell one, because the alternative is people treating a percentage as proof and accusing someone on it.

If something genuinely needs resolving, whether that’s a graded assignment, a hiring decision or a published piece, the reliable evidence has never been the text itself. It’s the process around it. Draft history, version history, the ability to talk through an argument you supposedly made.

Read the signals. Hold the verdict loosely. And when it counts, ask for the drafts.

Check any passage with Quetext’s AI Detector

Frequently asked questions

Can humans tell the difference between human and AI writing?

No, certainly not. Reports indicate that untrained performance is equivalent to chance, and the ability to make confident judgments is not a good predictor of accuracy. Studies show that individuals perform better on output that shows a certain pattern than on text that has been adapted. The message here is that your own insights should prompt you to further investigate the story.

• Untrained performance can be likened to a game of heads or tails
• Original text isn’t too hard to recognize
• Confidence isn’t a good indicator of one’s success

What are the clearest signs that something was written by AI?

Structural signals are more lasting than word signals. You should check whether the sentences are of the same length, grammatically accurate, and whether the conclusion of the paragraph only summarizes instead of providing a conclusion to it. It is impossible to prove anything by these alone.

• Little difference in sentence length
• Conclusion does not summarize the paragraph
• Categorical instead of real items

Does an em dash mean something was written by AI?

While em dashes, “delve” and carefully structured sentences may be distinct writing styles, they should not be seen as purposeful components of a text. Plenty of professional authors use these elegantly. Authors who make judgments based on punctuation may find themselves labeled incorrectly as not being careful writers.

• Punctuation is an aspect of writing rather than a sign of authorship
• Stylish habits indicate that people tend to be focused
• No particular word or punctuation mark can identify authorship

Are AI detectors more accurate than a person reading carefully?

They evaluate a different aspect instead of a better aspect. A detector assesses the statistical guessability of the full passage, something that is difficult to do visually. A careful reader assesses rhythm, purpose and point of view that the detector does not assess. They frequently disagree from each other and the disagreement gives some useful information. In both cases of either detectors or readers, it does not help in determining who has actually written the text.

• Detectors and readers evaluate different aspects
• Disagreement between them gives useful information
• Neither serves as evidence of authorship

Can AI writing be made undetectable?

Yes, the research exists and it’s publicly available. The use of requests for different sentence lengths and additional detail has largely eliminated structural indicators for writing. The second round of writing this particular article was indeed done in a way mentioned above, which is why process evidence has proved to be more effective than writing analysis today.

• Requests help eliminate structural indicators
• Paraphrasing lowers detection rate
• Process evidence outweighs textual analysis