Why the Same Text Can Score 90% AI Today and 10% Tomorrow

The same text scores differently from one month to the next because the detector underneath is constantly being replaced — retrained models, moved thresholds, and swapped components ship silently, which is why AI detection scores change over time even when not a word of your draft has. Nothing spooks writers more than this drift — a clean scan turning red weeks later reads like an accusation materializing out of nowhere. It isn’t. It’s version churn, and once you see the moving parts, both the panic and the false reassurance evaporate.

Key takeaways

• A hosted detector is a service, not a fixed program — vendors retrain and recalibrate continuously, almost never with announcements.
• The score is a probability squeezed through a threshold; retraining moves the probability, recalibration moves the threshold, and borderline text flips either way.
• Smaller mechanics stack on top: chunking boundaries, file-versus-paste formatting, integration differences between a vendor’s own products.
• Drift is *expected* behavior in an arms race — each new AI model generation forces retraining that reshuffles scores on old text.
• The practical defenses are dated evidence and revision that lowers scores across tools, not faith in any single snapshot.

The number you see is two moving parts

To understand why AI detection scores change over time, split the score into its components. Underneath, a classifier outputs a continuous probability — “this text resembles machine output this much.” On top, a presentation layer converts that probability into what you see: a percentage, a verdict, a color. Both layers move independently.

The classifier moves because it must. Detectors are trained on corpora of machine and human text, and every new model release makes yesterday’s corpus stale — the treadmill that defines this industry. When a vendor retrains, the probability assigned to *your unchanged essay* shifts, because the model now has different learned boundaries. Benchmarks like RAID show how sensitive detector outputs are to exactly these version and condition changes.

The threshold moves for product reasons. After launch backlash in 2023, Turnitin didn’t just tweak models — it changed how scores display, adding asterisks to low ranges and warning against acting on them. GPTZero’s public changelog records a steady cadence of model updates. When a vendor decides false positives are hurting them, they raise the bar for flagging; borderline documents that scored “78% AI” quietly become “23% AI,” or the reverse. Your prose never changed.

The extreme version of this is a vendor deleting the ruler entirely: OpenAI’s classifier gave scores in June 2023 and refused to exist in August. Every hosted detector does a smaller version of that constantly.

The smaller gears that add noise

Version churn explains the big swings. A cluster of mechanical quirks explains the smaller, weirder ones.

Chunking. Detectors analyze text in windows — sentences, or fixed-size segments. Add one paragraph and every downstream window boundary shifts, changing per-chunk scores and the aggregate. The same document as a file upload versus a paste can tokenize differently (smart quotes, line breaks, headers), nudging results.

Length effects. Scores on short text are noisy in every tool, so trimming or expanding a document changes not just what’s scored but how confidently. A 200-word excerpt of your 2,000-word clean essay can flag on its own.

Same vendor, different pipes. A detector embedded in an LMS integration, an API, and a consumer web page can run different versions or configurations simultaneously. This is a recurring source of student-teacher disputes — we’ve broken down a common case in why the same essay scores differently in Canvas and Turnitin.

Genuine nondeterminism. Less common, but some pipelines include components that introduce run-to-run wobble on borderline text. Most of what users call “randomness,” though, is one of the causes above wearing a mask.

We keep a shorter answer to the everyday version of this question in our FAQ on why a detection score changes on the same text; this piece is the long version — the *over time* dimension, which is almost entirely about versions.

Why AI detection scores change over time by design

Here’s the uncomfortable structural truth: score stability and detection accuracy are in direct tension. A detector that never changed would give you the same score forever — and grow uselessly stale as new models shipped. A detector that keeps up *must* retrain, and retraining *must* reshuffle scores on existing text, including flipping some verdicts. Vendors chose accuracy (and the subscription revenue that funds the chase). Stability was the price paid, and it was paid silently because publicizing “your old scores may no longer hold” invites exactly the disputes institutions use these tools to avoid.

That has a sharp implication for high-stakes use: a score is evidence about a text at a moment, produced by an instrument that no longer exists. Re-scanning a semester-old essay with today’s detector and treating the new number as truer is like re-grading last year’s exam with this year’s answer key. Institutions rarely grasp this; writers should.

Living with a moving ruler

The defenses follow directly from the mechanics.

Date your evidence. If a gatekeeper will judge your work by a score, capture your clean result when you submit — screenshot with visible date, tool name, and score. Drift can’t retroactively erase a documented snapshot.

Keep process proof. Version history and drafts outlast every model update; they verify authorship in a way no probability can, in either direction.

Revise for robustness, not for one tool. A passage that scores machine-flat everywhere — even rhythm, generic claims, statistically safe phrasing — will keep flagging across versions. Rewriting it with specifics and your own cadence lowers scores across tools and across time, which is the only stability on offer. A sentence-level heatmap shows exactly where to work: you can try it on your own text free.

Re-check before resubmission. If weeks passed between writing and submission, scan again — the ruler may have moved. At volume, see what each plan handles.

For the rest of the machinery behind these tools, browse the rest of our writing on detection.

Frequently asked questions

Why does the same text get different AI scores on different days? Because the detector is not the same detector. Vendors retrain models on fresh AI output, recalibrate thresholds, and swap ensemble components — usually silently. Your text didn’t change; the ruler did. Smaller effects stack on top: different text chunking between scans, formatting differences from copy-paste, and in some tools genuine run-to-run variation.

Are AI detectors deterministic — should identical input give identical output? Within a single model version, most are close to deterministic: same text, same version, same score. But you almost never control the version. Hosted detectors update continuously, so the practical answer is no — across days or weeks, identical input frequently produces materially different scores, and vendors rarely announce the changes that caused it.

My essay scored low at submission and high when my teacher checked it. What happened? Most likely a version or configuration gap: the tool retrained or recalibrated between scans, or the teacher used a different tool entirely — and integrations can differ from the same vendor’s web checker. Chunk boundaries can also shift when text is uploaded as a file versus pasted. This is exactly why timestamped screenshots of your pre-submission checks, plus draft history, are worth keeping.

Does score drift mean the newer score is more accurate? Not necessarily for your text. Retraining improves accuracy on average, mostly against newer AI models — but recalibrated thresholds move borderline cases in both directions, and a human-written passage can drift into flagged territory after an update. Newer is better in aggregate; on any single document, drift is just drift.

How should I protect myself against shifting scores? Treat any single reading as a snapshot, not a property of your text. Check drafts with a detector that shows sentence-level detail, keep dated evidence of clean results, preserve version history, and revise the passages that read machine-flat — that lowers scores across tools and versions, which is the only stability that exists in this system.

The bottom line

A detection score is not a fact about your text; it’s a reading from an instrument that is perpetually being rebuilt mid-flight. Retraining moves the probabilities, recalibration moves the thresholds, chunking and formatting add noise — and none of it is announced. So collect dated evidence when scores matter, keep the drafts that prove your process, and put your revision effort into writing that reads human across every version of every tool. The ruler will keep moving. Your paper trail and your voice are what don’t.