FAQ: Why Did My Detection Score Change When I Pasted the Same Text Twice?

You ran your essay through an AI detector, got a 28%, felt fine, and walked away. An hour later you pasted the exact same paragraph back in and it said 34%. Same words. Different number. So what gives, and which one do you believe?

Short answer: both, sort of. A detection score isn’t a measurement like your height. It’s an estimate, and estimates wobble. Let’s walk through why that happens and what a shifting number actually tells you.

Key takeaways

• AI detectors output a probability, not a fixed reading, so the same text can score a little differently between runs.
• The usual culprits: the tool updated its model, your pasted text wasn’t byte-for-byte identical, or the detector sampled or chunked it differently.
• A few points of drift is noise. A jump from 10% to 90% on truly identical input means something changed that you didn’t see.
• No score is proof of anything. It’s one signal, and a shaky one in isolation.
• If you care about consistency, lock the conditions and watch the trend across drafts instead of fixating on a single number.

A score is a guess with a decimal point

Here’s the thing most people miss. When a detector hands you “62% likely AI,” it isn’t reporting a fact it discovered inside your text. It’s reporting how confident its statistical model is that text like yours tends to come from a machine. That confidence is built on probabilities, and probabilities aren’t carved in stone.

Many detectors lean on two ideas: perplexity, which measures how predictable your word choices are, and burstiness, which measures how much your sentence rhythm varies. We dig into the math in perplexity and burstiness, the statistics behind AI detectors, but the short version is that these are continuous numbers being squeezed into a single percentage. Small changes in how the text gets read can nudge those numbers, and a nudge in the inputs becomes a nudge in the output.

So the real question isn’t “why did the score change.” It’s “what changed between the two runs.” There’s almost always something, even if you can’t see it.

The five things that usually moved

1. The detector updated its model

This is the biggest one, and it’s invisible to you. Detection companies retrain their models on their own schedule, and almost never with a public changelog. The classifier you hit on Monday may not be the same one you hit on Friday. If you paste the same text a week apart and the score shifts, the most likely explanation is that the tool changed underneath you, not your writing.

This is also why a screenshot of an old score is weak evidence in any dispute. The number was true for that model, on that day. The model has moved on.

2. Your text wasn’t actually identical

This trips up more people than anything else. You think you pasted the same thing, but you didn’t. Copy from Microsoft Word and you often drag along curly quotes, non-breaking spaces, and stray formatting characters. Copy the “same” text from a browser and those get flattened to plain versions. To you it looks identical. To a tokenizer, “don’t” with a curly apostrophe and “don’t” with a straight one are different sequences, and they can move the perplexity math a hair.

A trailing space, a double line break, an em dash that became two hyphens, all of it counts as different input.

3. The tool sampled or chunked differently

Some detectors don’t read every token the same way every time. Longer documents get split into chunks, and the tool may score each chunk and combine them. If the chunk boundaries land in slightly different places, or if the tool samples a subset of a long text, the combined score can shift between runs even with the same paste.

4. Surrounding context changed

If you score a paragraph by itself, then score it again inside the full essay, you’ll often get different numbers. The detector may read neighboring sentences as context. A clean, plain paragraph that looks human on its own can read differently when it’s sandwiched between two very uniform AI-sounding ones.

5. Short text amplifies everything

The less text you give a detector, the more any single word matters and the jumpier the score. A 40-word snippet can swing wildly. A 600-word piece is far steadier, because the statistics have more to average over. If your “two scores” came from a short blurb, expect them to disagree.

A mini-scenario: Priya’s discussion post

Priya writes a 150-word discussion-board reply in Google Docs. She pastes it into a detector and gets 22%. Reassured, she tweaks one sentence, copies it again, and this time gets 41%.

She panics. Did her edit make it look more AI?

Not really. Two things happened. First, her edit was real, so the input genuinely changed, and on a short 150-word post a single sentence carries a lot of weight. Second, her first paste came straight from Docs and carried curly quotes and a non-breaking space; her second paste went through a plain text box first and lost them. Different characters, shorter text, more sensitivity. The 19-point swing looks alarming but it’s the expected behavior of a probability estimate on a small sample.

The fix wasn’t to chase the number down. It was to recognize that her post sits somewhere in the 20s-to-40s for that tool, which is a borderline, inconclusive range, and to make her writing concrete and varied rather than optimizing for one reading.

So which number should you trust?

None of them, exactly. Trust the range.

If you want a stable read, do this:

• Paste the same text three or four times and note the spread. If you get 27, 31, and 29, your text lives in the low 30s for that detector. That band is your signal.
• Keep the conditions fixed. Same tool, same text source, same session. Don’t compare a Word paste today against a browser paste tomorrow.
• Use a meaningful length. Score the whole piece, not a sentence, if you want a number you can rely on.
• Watch the trend across drafts, not the absolute value. If revision moves you from the 60s into the 20s, that’s real progress, regardless of the exact digits.

And remember what the number is for. It’s a flashlight, not a verdict. You’re using it to find the parts of your writing that read as flat and predictable, then fixing those, not to win a fight with a decimal point.

The bigger point: scores were never proof

It’s worth saying plainly. Every reputable detector reports a likelihood, not a finding. OpenAI pulled its own AI text classifier in July 2023, telling users it was taken down because of its low rate of accuracy. A 2023 Stanford study led by Weixin Liang, published in the journal Patterns, found that several detectors disproportionately flagged writing by non-native English speakers as AI, misclassifying over half of the non-native TOEFL essays it tested. These tools are useful as signals and genuinely bad as evidence.

So if a single score changing between runs unsettles you, that instinct is correct, and it should make you skeptical of anyone who treats one screenshot as the final word. A measurement that won’t sit still isn’t proof of anything. It’s a probability doing what probabilities do. You can read more in the rest of our writing on detection if you want the full picture.

Frequently asked questions

Is it normal for an AI detector to give two different scores for the same text? Small differences are normal. Detectors output a probability, and most have some built-in variability from how they sample or chunk text. A swing of a few points between runs is nothing to worry about. A dramatic flip on truly identical input is unusual and almost always points to a model update on the detector’s side or a hidden difference in what you pasted, like a trailing space or a smart quote that got converted.

Does copying from Word versus a browser change my AI score? It can, slightly. Word and Google Docs often insert curly quotes, non-breaking spaces, and other invisible characters. A detector tokenizes those differently than the plain versions, which can nudge the perplexity math a little. The text reads the same to you, but it isn’t byte-for-byte identical, so the score can move.

If the score keeps changing, which number is the real one? None of them is a hard fact. The honest way to read a detector is as a range, not a point. If you paste a draft three times and get 28%, 31%, and 27%, your text sits around the low-30s for that tool. Treat that band as the signal and ignore the last digit.

Why did my score change a week later with no edits? Because the detector probably changed. These tools retrain and update their models on their own schedule, often without announcing it. The same paragraph can score differently this month than last month purely because the underlying classifier moved. This is also why a screenshot of a score is weak evidence in any dispute.

Can I make my score stable by just running it more times? You can average several runs to get a steadier read, and that’s a reasonable habit. But you can’t make the underlying tool deterministic, and you shouldn’t try to optimize toward one exact number. Focus on whether your writing varies in rhythm and stays specific. That improves both the score and how a real person reads it.

Bottom line

A score that shifts between identical pastes isn’t a glitch you need to solve. It’s a reminder of what the number actually is: a probability that moves with the model, the characters, and the length you feed it. Read the range, watch the trend, and put your energy into writing that’s specific and varied rather than chasing a decimal. Want to see how your own draft reads, sentence by sentence? Run your draft through PaperBleach and look at where it flags, not just the final percentage.