What ‘Mixed’ or ‘Partially AI’ Verdicts Actually Mean Under the Hood

You ran your draft through a detector and it didn’t say “human” or “AI.” It said something in between: “mixed,” “partially AI,” “likely contains AI-generated content,” or just a number stranded in the middle like 47%. That ambiguous result feels worse than a clear verdict, because now you don’t know whether to relax, panic, or start rewriting.

Here’s the thing most people miss: a mixed verdict isn’t a single state. Under the hood it can mean two almost-opposite things, and they call for opposite responses. Once you can tell which one you’re looking at, the scary middle number turns into useful information.

Key takeaways

• A “mixed” or “partially AI” label can mean two different things: some passages scored high while others scored low, *or* the entire text landed in the detector’s uncertain middle zone. They need different fixes.
• Detectors classify chunks of text, then roll those scores into one verdict. “Mixed” is often just the honest output of that rollup when the chunks disagree.
• A “partially AI” percentage is not a measure of how much of your text is AI. It’s a probability, and mixed scores tend to sit right where the detector is least confident.
• Heavily edited AI text, ESL phrasing, and genuinely co-written drafts all produce mixed verdicts for completely different reasons.
• Open the sentence-level view before reacting. Whether the flags are clustered or spread out tells you which kind of “mixed” you’ve got.

A detector doesn’t grade in one pass

To understand the middle, you have to understand how the number gets built. A detector doesn’t read your whole document and form one gut feeling. It breaks the text into chunks — usually sentences, sometimes short passages — and scores each one for how statistically predictable it looks. Predictable, evenly paced phrasing reads as machine-made; surprising, uneven phrasing reads as human. (If you want the math behind *why* predictability is the tell, we walk through the statistics that drive these scores.)

Then it aggregates. All those per-chunk scores get combined into the single headline verdict you actually see. The exact recipe is proprietary and varies by tool, but the key point is that the headline is *derived* from the chunks — it is not a simple average you can eyeball.

That aggregation step is where “mixed” is born. When the chunks mostly agree, you get a confident label. When they disagree, the rollup has to represent that disagreement somehow, and “mixed” or a mid-range percentage is the result.

The two flavors of “mixed”

Flavor one: a genuinely split document

This is the intuitive case. Some passages scored clearly AI-like and others scored clearly human, so the document lands somewhere in between.

Picture a student who drafted an essay themselves, then asked ChatGPT to “polish the introduction and conclusion.” The body paragraphs are messy, specific, full of their own examples — low scores, reads human. The intro and conclusion are smooth, generic, perfectly balanced — high scores, reads AI. Aggregate those and you get a mixed verdict that’s actually telling you something true: part of this really is machine-shaped.

In this flavor, the heatmap is concentrated. Open the sentence-level view and you’ll see hot spots in obvious places, with cool stretches around them.

Flavor two: a uniformly uncertain document

This one trips people up. Here, no individual passage screams AI. Instead, the *whole* document sits in the detector’s gray zone — every chunk scores around “maybe,” so the aggregate lands in the middle too.

This happens constantly with clean, plain, careful writing. Technical explanations. Definitions. A well-organized but unadventurous paragraph. None of it is predictable enough to flag hard, none of it is surprising enough to clear as obviously human. The detector hedges across the board.

In this flavor, the heatmap is spread out. There’s no smoking gun — just a wash of pale-orange uncertainty.

Same “mixed” label. Completely different situations. And you can’t tell them apart from the headline number alone.

Why the middle is where detectors are least confident

Every detector has a decision threshold — a cutoff where it flips from calling text human to calling it AI. A mid-range score means your text is sitting near that line, and the line is exactly where the tool’s confidence is lowest.

Think of it like a thermostat that’s uncalibrated. Far from the threshold, the readings are reliable: deeply predictable text is almost certainly going to flag, wildly idiosyncratic text almost certainly won’t. But right at the boundary, tiny differences in phrasing flip the call. That’s why the same passage can come back “partially AI” on one tool and “human” on another. Each tool draws its threshold in a slightly different place and trains on different data, so they hedge in different spots. Neither is “right.” They’re both estimating.

This is also why mixed verdicts are so common with non-native English writing. A 2023 study by Liang and colleagues, published in *Patterns*, found that detectors were biased against non-native English writers — more than half of the non-native-authored TOEFL essays they tested were flagged as “AI-generated.” The researchers tie this to low perplexity: conventional, smaller-vocabulary phrasing produces little statistical surprise, which the math reads as machine-like. A whole essay of that phrasing doesn’t flag as confidently AI; it sits in the uncertain middle, which is its own kind of false alarm.

A mini-scenario: same 50%, two different problems

Two writers each get a 50% “partially AI” verdict on a 400-word piece.

Writer A drafted everything by hand but writes in a clipped, uniform style — short declarative sentences, common word choices, steady rhythm. Their heatmap is an even haze. Every sentence scores around the middle. The 50% is the detector shrugging at the whole thing. The fix isn’t to “remove the AI” — there is none. It’s to add variation: let one sentence run long, follow it with a fragment, drop in a specific detail only they would know.

Writer B wrote three paragraphs themselves and pasted in two AI-generated ones. Their heatmap has two dark red blocks and a sea of green. The 50% is an average of “very human” and “very AI.” The fix is surgical: rewrite the two pasted paragraphs in their own voice and the score will likely fall hard, without touching the rest.

Identical number. The right move for one writer is the wrong move for the other. The heatmap is the only thing that tells them apart.

How to actually read a mixed verdict

A repeatable sequence beats panic every time:

1. Ignore the headline for a second. The percentage tells you *that* there’s ambiguity. It doesn’t tell you *where* or *why*.
2. Open the sentence-level view. This is non-negotiable for mixed results, because the whole question is whether the score is concentrated or spread.
3. Check the shape. A few dark clusters mean a genuinely split document (Flavor one). An even wash means uniform uncertainty (Flavor two).
4. Match the fix to the shape. Concentrated? Rewrite the hot passages in your own concrete voice. Spread out? Work on variation and specificity throughout — vary sentence length, swap abstractions for examples, break the steady rhythm.
5. Re-check and watch what moves. If the number drops after targeted edits, you confirmed the drivers. If it barely budges, the problem is more diffuse than you thought.

What a mixed verdict is *not*

It’s not a percentage of your text that’s AI. “40% AI” does not mean 40% of your sentences came from a machine; it’s a probability on a scale the tool calibrated for itself.

It’s not proof of anything. A mixed result is, by definition, the *least* confident output a detector can produce, since it’s sitting in the uncertainty zone. If anyone treats a partial score as evidence, remember that even OpenAI retired its own AI classifier in July 2023 for a low rate of accuracy — by its own published evaluation, the tool correctly flagged only about 26% of AI-written text while misflagging 9% of human writing. The company that built the model couldn’t reliably detect its output. A mid-range number from a third-party tool deserves even more caution.

And it’s not a reason to torch your whole draft. Most mixed verdicts come down to a handful of passages or one stylistic habit. Blind rewriting usually makes things worse by sanding off the uneven, specific bits that read as human in the first place.

Frequently asked questions

Does a “partially AI” verdict mean exactly that percentage of my text was AI-written?

No, and this is the most common misread. A detector doesn’t count words and report a fraction. It assigns each chunk a probability and combines those into one number, which is itself a probability on a scale the tool calibrated on its own data. A “40% AI” result doesn’t mean 40% of your sentences came from a machine. It usually means the detector is uncertain, or that a few passages scored high enough to pull an otherwise-human document into the middle. The sentence-level breakdown is the only way to see what’s really happening.

Why does the same text get “mixed” on one detector and “human” on another?

Because each tool draws its decision threshold in a different place, trains on different data, and aggregates per-sentence scores with its own undisclosed method. A passage that lands just above one tool’s “lean AI” cutoff can sit just below another’s. Mixed verdicts cluster around these thresholds, which is exactly where small differences between tools produce different labels. It’s not that one detector is right and the other broken. They’re estimating, and they hedge in different spots.

I wrote every word myself. Why am I getting “partially AI”?

Predictable, evenly paced writing scores like AI even when a person wrote it. If part of your draft is plain and uniform — a tidy intro, a definition, a list — and part is loose and specific, the detector can read the smooth sections as machine-like and split the difference into a mixed verdict. The 2023 Stanford-affiliated *Patterns* study by Liang and colleagues found detectors disproportionately misflag non-native English writing for the same underlying reason: conventional, lower-variability phrasing produces low statistical surprise, which detectors read as generated.

If I got a mixed verdict, which sentences should I fix?

Open the heatmap and find the hottest passages, not every flagged word. If the score is concentrated in two or three spots, rewrite those in your own concrete voice and re-check. If the flags are spread thinly across the whole document, the issue is uniformity rather than a few bad lines, and you’ll need to add variation and specifics throughout. The shape of the heatmap tells you which problem you have.

Should I trust a “mixed” verdict as evidence of anything?

Treat it as a weak signal, never proof. Detectors output probabilities, and a mixed result is the lowest-confidence output a detector can give, because it’s sitting near its own uncertainty zone. OpenAI retired its own classifier in July 2023 for low accuracy. If a teacher or editor is treating a partial score as a verdict, the mechanics of how that number is produced are your strongest argument for why it can’t carry that weight.

The short version

“Mixed” and “partially AI” aren’t a confidence dial set halfway. They’re what a detector says when its chunks disagree or when your whole document sits on its uncertainty line. A split document and a uniformly bland one can produce the identical number, and the only way to tell them apart — and to know what to fix — is the sentence-level view, not the headline.

Want to see which kind of “mixed” you’re actually dealing with? Run a draft through the free checker and read the heatmap: clustered flags mean a few passages to rewrite, an even wash means you need more variation throughout. You can also browse the rest of our guides on how detection works, or compare plans if you want to check longer pieces. Either way, use the score to improve the writing for real readers — not to chase a number.