The ‘Show Don’t Tell’ Edit That Removes the Biggest AI Tell in Essays

Most “AI-sounding” essays don’t sound robotic because of one weird word. They sound that way because they keep *telling* you things instead of *showing* you anything. The fix is older than any detector: trade a vague claim for a specific scene, and the whole paragraph wakes up.

This is the single highest-leverage edit you can make to a draft that came out of ChatGPT, Claude, or your own tired brain at 1 a.m. It happens to also be the edit that moves AI-detection scores the most, and the reason why is worth understanding.

Key takeaways

• “Telling” prose summarizes and generalizes. That flat, abstract texture is the most common reason an essay reads as machine-written.
• “Showing” swaps a claim for a specific the reader can picture – a number, a name, a moment, an example.
• You don’t have to rewrite everything. Find the three or four sentences that only assert things, and ground each one.
• Detectors measure statistical predictability, not honesty. Real specifics are genuinely harder to predict than generic summary, so this edit lowers scores as a side effect.
• Keep your meaning. Show don’t tell adds evidence and texture; it shouldn’t change what you argue.

Why “telling” is the biggest AI tell

When a language model writes, it reaches for the most probable next word. Average that out over a few hundred words and you get prose that’s grammatically clean, evenly paced, and oddly weightless. It states that something is “significant” or “multifaceted” or “plays a crucial role” without ever pinning down what actually happened.

That’s telling. And telling is exactly what AI models are best at, because summary is generic by nature – there are a hundred fluent ways to say “this had a big impact,” and the model picks the smoothest one.

Detectors notice this. Tools like GPTZero and Turnitin score text partly on how predictable and how uniform it is – concepts often called perplexity and burstiness. Low perplexity means the words were easy to guess. Low burstiness means every sentence has roughly the same shape. Generic summary scores low on both. We break down the statistics behind that in our explainer on perplexity and burstiness, but the short version is: vague writing is predictable writing, and predictable writing is what gets flagged.

Here’s the part people miss. You can’t reliably fix predictability by shuffling synonyms or adding a few em dashes. The predictability lives in the *content*, not the vocabulary. A vague sentence stays vague no matter how you dress it. The only durable fix is to put something specific in the sentence – something the model couldn’t have guessed because it came from your actual source, your actual experience, or your actual data.

What “showing” looks like in practice

Showing doesn’t mean writing flowery descriptions. In an essay, it usually means evidence and precision. Three quick moves cover most of it:

1. Replace adjectives with the thing the adjective describes

“Telling” leans on evaluation words: *important, effective, significant, various, numerous.* Each one is a placeholder for evidence you haven’t supplied yet.

• Tells: “The Clean Air Act had a significant effect on urban pollution.”
• Shows: “After lead was phased out of gasoline under the Clean Air Act, the average blood-lead level in U.S. children aged 1 to 5 fell about 77% between the 1976-80 and 1988-91 federal health surveys.”

The second sentence makes the *same argument*. It just hands the reader the fact instead of the verdict and lets them reach the verdict themselves. That’s more persuasive to a human grader and harder for a model to predict.

2. Swap the general for the particular

Models love categories. Writers convince with instances.

• Tells: “Many companies struggled to adapt to remote work.”
• Shows: “In February 2013, Yahoo’s CEO Marissa Mayer banned working from home and told remote staff to report to an office by that June – seven years before a pandemic shoved nearly every office worker in the country back to their kitchen tables.”

One named example does more than three sentences of “many companies.”

3. Turn a claim into a moment

When you can, drop the reader into a scene. This works even in formal writing if you keep it tight.

• Tells: “Patients often experience anxiety before surgery.”
• Shows: “The nurse said the patient kept asking what time it was, then asking again four minutes later.”

You’re still making a point about pre-surgical anxiety. You’re just proving it with behavior instead of asserting it.

A mini-scenario: editing a flagged paragraph

Say a student runs a paragraph through a detector and gets an 89% “likely AI” score. Here’s the original:

> “Social media has had a profound impact on modern communication. It has changed the way people interact and share information. While it offers many benefits, it also presents various challenges. Overall, its influence on society is significant and continues to grow.”

Read it again. Every sentence tells. Nothing is pictured. There isn’t a single fact, name, or number – which is precisely why it’s so easy for a model to produce and so easy for a detector to flag. Four sentences, four verdicts, zero evidence.

Now the show-don’t-tell pass. We keep the meaning and add specifics:

> “When my aunt wanted to find a kidney donor in 2019, she posted once on Facebook and had three offers by the weekend – something the local transplant registry hadn’t managed in two years of waiting. That same platform also fed her a stream of vaccine conspiracy videos she half-believed by spring. Social media didn’t just change how she communicated; it changed what she trusted, and how fast.”

Shorter on hedging, longer on detail. There’s a date, a person, a concrete outcome, and a turn. The sentence lengths now vary instead of marching in step. The argument is intact – social media’s impact is profound and double-edged – but now you can *see* it.

That variation in rhythm and the unpredictability of “three offers by the weekend” are the things detectors register as human. Not because the tool understands the story. Because the words got harder to guess.

Why this works on detectors (and why that’s not the point)

It’s tempting to treat this as a trick. It isn’t, and the framing matters.

A detector doesn’t know whether you used AI. It estimates a probability based on statistical patterns, and it gets things wrong in both directions. OpenAI quietly retired its own AI Text Classifier in July 2023 because it was too inaccurate to trust – by its own account, the tool correctly flagged only about a quarter of AI-written text. A 2023 Stanford study led by Weixin Liang found these tools disproportionately flag non-native English writers, whose more formulaic phrasing reads as “machine-like” to the math. So a high score is a signal, not a sentence.

Show don’t tell lowers that signal because it genuinely makes your writing less generic – the same reason it makes the writing *better*. You’re not gaming a number. You’re fixing the underlying flatness the number was pointing at. That’s the white-hat version of this work: you understand what the detector reacts to, and you respond by writing more specifically, not by hiding anything.

If you want to confirm the edit landed, you can run your draft through a detector and clean up whatever statistical patterns remain – but lead with the craft. A tool smooths residual patterns; it can’t invent the specific detail that was missing from your draft. Only you have that.

How to run the edit on your own draft

A repeatable pass takes about ten minutes:

1. Highlight every evaluation word with no evidence attached – *significant, important, various, effective, numerous, crucial, profound.* Those mark your telling sentences.
2. Pick the three or four that carry your argument. You don’t need to fix all of them; over-showing every line gets exhausting.
3. For each, ask “what’s my evidence?” A number, a name, a date, a quote, a specific example, a small scene. Put it in the sentence.
4. Read for rhythm. You’ll notice the sentence lengths naturally vary now – short summary, longer detail. Keep that mix.
5. Check the meaning didn’t drift. Showing adds support; it shouldn’t change your claim.

If a sentence genuinely has no available evidence, that’s useful information too – it might be a claim you can’t actually back up, and the honest move is to soften it or cut it.

Frequently asked questions

Does “show don’t tell” only matter for creative writing?

No. It started as fiction advice, but it applies to essays, lab reports, case studies, and discussion posts too. In academic writing, “showing” usually means evidence: a specific study, a number, a named example, or a short scenario instead of a broad claim. “The policy had major effects” tells; “enrollment dropped 12% the year after the policy passed” shows.

Will adding concrete details actually lower an AI-detection score?

Often, yes, though no edit is a guarantee. Detectors flag text that’s statistically too smooth. Generic summary is exactly that. Replacing it with real specifics introduces words the model didn’t expect, which raises perplexity and sentence variety. Treat a lower score as a side effect of clearer writing, not the goal.

How do I know which sentences are “telling”?

Read each one and ask: could I picture this, or is it just an abstract claim? Watch for evaluative words with no evidence attached. Fix the three or four that carry the most weight in your argument.

Can I overdo show don’t tell?

Yes. If every sentence is a dense scene, the writing gets exhausting and loses its thread. Good prose alternates: tell to move fast, show to prove. The mix is part of what reads as human.

Does this replace using an AI humanizer?

They solve different parts of the problem. The craft edit fixes substance and is worth learning yourself. A humanizer smooths residual sentence-level patterns. Do the edit first, then check and polish.

The short version

The biggest AI tell isn’t a banned phrase. It’s the habit of stating conclusions instead of showing the reader how you got there. Fix that in three or four sentences and your essay gets clearer, more convincing, and – as a bonus – less likely to trip a detector.

Want to verify it worked? Run your edited draft through PaperBleach and check it with a detector and humanize what’s left. You can also browse the rest of our writing-craft posts or see plans if you write at volume.