How to Tell If Something Was Written by AI (2026 Signs)

July 1, 2026 · FiftyGPT Editorial Team

Spotting AI writing used to be easy. Early chatbot text was stiff, repetitive, and obvious. That is no longer true. Newer models produce text that is harder to separate from human writing, and research now shows that the average person is close to guessing when asked to tell the difference. One 2025 study found that people performed no better than random chance at distinguishing AI text from human text.

That does not mean it is hopeless. AI writing still follows patterns, and once you know what to look for, you can spot the tells far more often than chance. This guide walks through the practical signs for a US audience of teachers, editors, students, and anyone who reads a lot of online writing, along with the honest caveat that no single sign proves anything on its own.

The short answer

You can often tell text was written by AI by looking for a cluster of signs: unusually uniform sentence rhythm, overused transitions, stock opening phrases, a lack of contractions, abstract rather than concrete examples, repetitive phrasing, and hollow, detail-free passages. The strongest single signal is uniform sentence length and structure. No tell is proof by itself. AI authorship becomes likely only when several signs appear together.

Why this is getting harder

Two things make detection tougher every year. Models keep improving, so the obvious tics fade. And human writing is starting to absorb AI patterns, because so many people now read and edit with these tools, which blurs the line from the other direction.

There is a useful nuance in the research. People who barely use language models do only slightly better than chance at spotting AI text. People who use them heavily, every day, can correctly flag AI writing far more often, in one preprint around 90 percent of the time. Recognizing AI writing is a learnable skill, and the more machine text you read closely, the better your eye gets.

The signs that text was written by AI

No single item below is a verdict. Treat them as a checklist, and grow suspicious as more of them stack up.

Uniform sentence rhythm

This is the most reliable manual signal. Humans write in bursts, mixing a short, punchy sentence with a long, winding one. AI text tends to keep sentences at a similar length and a similar complexity throughout, which creates a smooth, slightly monotonous rhythm. If every sentence feels about the same size and shape, that evenness is a strong tell.

Overused transitions and connectors

AI loves to stitch ideas together with formal connective words, and it often stacks them at the start of consecutive sentences. Watch for a steady drumbeat of words like "consequently," "subsequently," and "thus" opening sentence after sentence. Human writers use transitions too, but rarely with that mechanical regularity.

Stock opening phrases

Machine drafts often lead with a generic, throat-clearing setup. Openers built on a participial phrase, the kind that begins "In an era of rapid change" or similar, are a common AI fingerprint. They sound polished but say almost nothing, and a careful human editor usually trims them.

The rule of three

AI has a habit of listing things in tidy groups of three. "It is clear, concise, and catchy." A single set of threes means nothing, since humans write that way too. But when nearly every list and emphasis arrives as a perfectly balanced trio, the pattern starts to show.

Missing contractions and an overly formal tone

AI defaults to complete forms ("you are," "cannot," "do not") unless told otherwise, which makes the writing feel stiff. Humans, especially in casual or conversational contexts, lean on contractions. Writing that reads like a press release in a setting that calls for a quick note is worth a second look.

Abstract examples instead of concrete ones

When AI gives an example, it often stays vague and hypothetical: "a marketing team might use this to improve engagement." Human writers tend to reach for the specific: a real product, a named place, an actual number, a moment that happened. Generic, placeholder examples are a common machine tell.

Repetitive phrasing

AI frequently circles the same idea two or three times in slightly different words. You might read that a home "is charming," then in the next sentence that "buyers will be charmed by this home." That kind of redundancy is unusual in a confident human writer and common in machine drafts.

Inflated, recurring vocabulary

Studies have shown that certain words started appearing far more often in text written after chatbots became widespread. Recurring favorites include words like "underscore," "showcasing," "intricate," "multifaceted," and "testament." One word means nothing. A passage stuffed with several of them, used heavily, points toward AI involvement.

Punctuation tells

Two small signals show up often. AI models, ChatGPT in particular, reach for the em dash at a much higher rate than typical writers, sometimes squeezing one into nearly every other sentence. Many chatbots also default to curly quotation marks and curly apostrophes rather than straight ones, and sometimes mix the two inconsistently. Neither is proof on its own, since plenty of human writers love the em dash and published work uses curly quotes, but heavy, mechanical use is a clue.

Hollow emotion and missing sensory detail

AI can describe a feeling without ever making you feel it. A passage may discuss grief with no shaky hands, no specific memory, no taste or smell, just a smooth summary of the emotion. Writing that celebrates a win yet feels oddly weightless, as if no one actually sweated for it, often started as a prompt.

Fabricated facts and citations

This one is both a tell and a warning. AI sometimes invents oddly specific data, fake citations, or links that lead nowhere. If you see a strangely precise statistic with no source trail, a study that does not exist, or a broken reference, treat it as a red flag and verify before trusting any of it.

AI tell	Human signal
Uniform sentence length	Varied, bursty rhythm
Stacked formal transitions	Natural, occasional transitions
Vague hypothetical examples	Concrete, specific examples
Repetitive restated ideas	Confident, non-redundant phrasing
Hollow, detail-free emotion	Sensory, lived detail
Fabricated or unsourced facts	Verifiable, sourced claims

Why no single sign is proof

Here is the part many people skip, and it matters. Every one of these signals appears in genuine human writing too. Some people naturally write in clean, even sentences. Some love the em dash. Some are taught to use formal transitions and lists of three. A second-language writer may use simpler, more predictable phrasing for reasons that have nothing to do with AI.

That is why a single tell should never lead to an accusation. The signs are meaningful only in combination, and even then they point to a likelihood, not a fact. This is the same reason detection tools produce false positives, a problem we cover in our guide on why detectors flag human writing.

Different models leave different fingerprints

Not all AI text reads the same, and knowing the differences sharpens your eye. The vocabulary that gives away machine writing has shifted as models changed. Earlier ChatGPT output leaned on one set of favorite words, while later versions drifted toward a different set, with words like "emphasizing," "enhance," "highlighting," and "showcasing" climbing in frequency. Other models have their own quirks. Some lean heavily on pseudo-scientific words like "causal," "empirical," and "correlate," and on the word "underscore," far more than a typical writer would.

Because these are linguistic habits rather than topic choices, they show up regardless of subject, which is what gives a lot of AI text an identifiable "voice." The practical takeaway is to calibrate to the model you suspect. A passage that feels machine-made will often carry a recognizable cluster of that model's pet words, and once you have read enough output from a given tool, you start to hear its default cadence almost on sight.

A two-minute test you can run

You do not need software for a fast first read. When a passage feels off, try this:

Read it aloud and listen to the rhythm. If every sentence lands at about the same length and beat, that evenness is your first flag.
Scan the connectors and lists. Count the formal transitions and the groups of three. A steady, mechanical pattern is telling.
Pressure-test one example. Pick the most specific-sounding claim and ask whether it is genuinely concrete or just a vague hypothetical dressed up as detail.
Verify one fact. Take a single statistic or citation and check it. Fabricated or unsourced specifics are among the most reliable tells.
Then add a tool. If your read raises suspicion, confirm with a detector rather than stopping at a hunch.

This takes about two minutes and catches far more than a snap judgment, because it forces you to look at structure and substance instead of reacting to surface polish.

Manual reading versus detection tools

Your own reading catches things software misses: a hollow anecdote, an example that is suspiciously generic, an argument with no real stake. A detection tool catches statistical patterns your eye cannot measure, like overall perplexity and burstiness across a long document. Where they agree, you can be more confident. Where they disagree, the honest conclusion is uncertainty, not a verdict.

How to check more reliably

If it matters, do not rely on a gut call.

Look for a cluster, not one clue. Several tells together mean far more than any single one.
Cross-check with a detector. Run the text through an AI checker like FiftyGPT to add a statistical read to your manual one, and remember that a score is a signal, not proof.
Verify the facts. Fabricated citations and broken links are some of the most reliable tells, and checking them takes minutes.
Consider the writer and context. A polished, even style is normal for some people and some genres. Account for that before drawing a conclusion.

For teachers and editors

If you evaluate other people's writing, two principles keep this fair. First, use signs and tools to start a conversation, not to deliver a sentence. Second, weigh process over output: drafts, notes, version history, and a short talk reveal far more than any single tell or score. The goal is to understand the writing, not to police it.

Keep reading

FAQs

Can you reliably tell if text was written by AI?

Not from a single sign, and not with certainty. You can raise your odds well above chance by looking for several tells together, especially uniform sentence rhythm, but every signal also appears in human writing, so treat results as likelihood rather than proof.

What is the most reliable sign of AI writing?

Uniform sentence length and structure. Human writing naturally varies its rhythm, while AI tends to keep sentences a similar size and complexity, creating a smooth, even pattern.

Are em dashes a sign of AI?

Only in combination with other signs, and only when used heavily. AI models do overuse the em dash, but many skilled human writers love it too, so frequency alone is never proof.

Why is it getting harder to spot AI writing?

Models keep improving and producing more varied text, and human writing is absorbing AI patterns from constant use. Studies show most people now perform near chance, though heavy AI users spot it far more often.

Do AI detectors work better than reading it myself?

They measure things you cannot, like statistical predictability across a document, but they also produce false positives. The strongest approach combines your reading with a detector and trusts neither alone.

Can AI writing have fake facts?

Yes. AI sometimes invents precise-sounding statistics, fake citations, or dead links. Unsourced, oddly specific claims are a strong tell, and verifying them is one of the most reliable checks you can run.

Is it fair to accuse someone based on these signs?

No. These signs indicate likelihood, not proof, and all of them appear in genuine human writing. Use them to start a conversation and to look at the writer's process, not as grounds for an accusation on their own.

Can heavily edited AI text still be spotted?

It gets much harder. Once a person rewrites AI output in their own words, adds specifics, and varies the rhythm, most surface tells fade and both manual reading and detection tools become far less reliable. Blended human-and-AI writing is the toughest case to call, which is another reason to lean on process and conversation rather than a single judgment. ---