Is It Human
Or Is It a Machine?

How To Spot AI Fiction — And Why It Matters

Large language models produce predictable prose. Not bad prose — predictable prose. They've been trained on billions of words of text and their default output lands in the statistical middle of everything they've seen. That makes AI fiction feel like fiction you've read before, even when the plot is brand new.

Fiction AI tells are different from nonfiction AI tells. Nonfiction AI leans on however, furthermore, it's important to note, and at the end of the day. Fiction AI leans on em-dashes, body-language stacking, something shifted vagueness, and stock emotional metaphors. A general AI detector will miss fiction-specific patterns because it's looking for the wrong things.

This detector was built from studying hundreds of AI-generated novels and comparing them to hundreds of human-written ones. The patterns are consistent across ChatGPT, Claude, Gemini, and DeepSeek — because they're all trained on similar data. Once you know what to look for, you can't unsee it.

The Five Big Fiction AI Tells

1. Em-dash overuse. AI fiction uses 15 to 25 em-dashes per chapter. Human fiction uses 1 to 4. Em-dashes are the AI comfort food — they let the model interrupt itself mid-thought, which mimics the natural flow of human writing without actually committing to a structure. If you see three em-dashes in a paragraph, you're probably looking at AI.

2. Body-language stacking. "Her jaw clenched. Her fists tightened. Her shoulders went rigid." All in one sentence. Or worse: "Her jaw clenched, fists tightening, shoulders going rigid as her breath caught in her throat." One physical reaction per beat is human. Three or more is AI desperately trying to show, don't tell and overshooting the target.

3. "Something [verb]ed" vagueness. "Something shifted inside her." "Something broke." "Something ancient stirred." "Something primal took over." AI avoids naming specific emotions because specific emotions are risky. A human writer who knows their character names it: "Shame. Hot and immediate. The kind she hadn't felt since she was nine."

4. Filter words everywhere. I felt, I noticed, I saw, I heard, I realized, I could see, seemed to, appeared to. These are distance creators. They push the reader one step back from the experience. Human close-POV writing cuts them: "I heard the door slam" becomes "The door slammed."

5. Stock metaphor library. A tapestry of emotions. The weight of the moment. Like a shroud. The raw edges of her grief. Something primal/ancient/feral/visceral. These phrases don't come from the specific scene — they come from the AI's training set. A human writer pulls metaphors from what the character is physically experiencing right now, in this specific place, with this specific history.

How To Use This Detector

Step 1: Paste a representative sample

Free up to 2,500 words a run. Don't paste the whole novel — paste a chapter opening, an emotional beat, a dialogue scene. The detector needs enough material to find patterns but not so much that it averages them out.

Step 2: Read the report

You'll get a score, a fiction-or-nonfiction classification, and a specific list of tells with short quoted examples. The score is probabilistic — a 70% score doesn't mean the text is AI. It means 70% of its patterns match AI writing. A human can write text that scores 70% if they happen to lean on the same patterns.

Step 3: Fix the tells

Whether the text came from a machine or a tired human, the tells are things worth fixing. Em-dash overuse makes prose feel samey. Body-language stacking makes emotional beats feel performed. "Something shifted" makes your writing feel interchangeable with every other AI-assisted draft on the market. Use the Humanizer to rewrite the flagged passages, or run your whole draft through the AI Slop Scanner for a broader pass.

Why Human Writing Sometimes Gets Flagged

Here's the uncomfortable truth: if your writing gets flagged as AI, the detector isn't wrong. It's telling you that your patterns match the patterns of generic machine output. That's a craft problem either way.

Human writers can absolutely write text that scores high on AI detection:

• If you overuse em-dashes (common among writers who love the "interrupted thought" effect).
• If you stack body language to signal strong emotion (common in YA and romance).
• If you default to "something" phrases to describe interior states (common among writers who learned to show, don't tell but haven't moved to the next level: name it specifically).
• If you filter everything through the narrator's perception ("I saw," "I heard," "I felt").

Getting flagged is useful diagnostic information. It tells you where your voice is sliding into the default template everyone else is writing in. Fix the patterns, and both your AI detection score AND your prose get better at the same time.

AI Detector FAQ

Is this detector really free?

Yes. Free up to 2,500 words a run. No account, no email. Bring your whole manuscript to the paid tools.

Can I use this in academic settings?

You can, but be careful. AI detection is probabilistic. Use the results to identify passages worth discussing — not to accuse. A high score is a signal, not a verdict.

Does it work on all AI models?

It catches patterns common across GPT-4, Claude, Gemini, DeepSeek, and Llama. The patterns are surprisingly consistent because the models are trained on overlapping data.

What about AI-assisted writing where a human heavily edited?

If the human edited enough to break the patterns (mixed sentence lengths, cut the em-dashes, named the emotions, killed the stock phrases), the score drops. If they just ran it through a grammar checker and called it edited, the patterns are still there.

Next Steps

• The Humanizer — rewrites flagged passages to sound like a human wrote them.
• AI Slop Scanner — a broader, pattern-specific scan with 21+ categories.
• Grammar Checker — for the mechanical errors AI detection doesn't cover.
• AI Beta Reader — developmental critique for bigger structural issues.