Fall 2026 AI Detector Model Changes: Why Before-and-After Scores No Longer Compare

Table of Contents

• Why did my AI score change when I didn't change the text?

• What changed, and when? The fall 2026 changelog

• Why scans from before and after a model change can't be compared

• How do you retest fairly?

• FAQ

Why did my AI score change when I didn't change the text?

Most likely because the detector changed underneath you. Between September 18 and October 19, 2026, GPTZero, Copyleaks, Pangram and Turnitin each moved to a new detection model, and GPTZero and Pangram each made a second change inside the same stretch. A scan from before one of those dates and a scan from after it were read by different models, and the difference between the two numbers can't be pinned on your edits.

That is not a satisfying answer if you spent an evening reworking a draft and watched the number go the wrong way. It is still the honest one. A score is one model's reading of one piece of text on one day. This fall, the model behind most of the scores people check changed at least once, and in several cases the result on your screen gives no sign of it.

What changed, and when? The fall 2026 changelog

Every entry below comes from the vendor's own page, opened on October 27, 2026. Where a page gives only a month, the table gives only a month.

GPTZero: 4o, released September 18, default from September 20

GPTZero's release notes head the entry "September 13, 2026 (2026-09-13-base, Model 4.10b aka 4o) | released on September 18, 2026." The two changes it lists are short: "Decreased false positive rate on a variety of domains" and "Improved performance on third party benchmarks." The entry is dated the 13th, but the release day it gives is the 18th, and that's the one to use.

The announcement, "Introducing GPTZero 4o," adds the date that matters to most people: "As of 20 September 2026, the updated model is available to all users and is the default model offered on GPTZero." It also changes what gets read. GPTZero says it now masks headers, "meaning our detector does not consider headings as a feature," and that a new gray highlight shows "which pieces of text are excluded by our AI detector." So if your newer GPTZero result has gray headings and your older one doesn't, that alone tells you the two weren't produced the same way. The post also reports one false positive out of 10,402 essays on the PERSUADE student-writing set. That is GPTZero's own claim, and it's mentioned here only because it's the stated goal of the update.

Then GPTZero changed again. The same release notes list "October 7, 2026 (2026-10-07-base, Model 4.11b aka 4.1o) | released on October 7, 2026," with "Improved recall on sub-classes AI paraphrased, mixed and polished" and "Improved document chunking and aggregation." That's a second break in GPTZero's timeline less than three weeks after the first, and it touches exactly the kind of text people tend to recheck: drafts that have been edited, polished or mixed with their own writing.

Copyleaks: AI Text Detection Model V11.0, September 22

Copyleaks' API documentation has a "What's New" page, and its entry dated September 22, 2026 reads: "We've released version 11.0 of the AI Text Detection model. This release improves detection accuracy and extends coverage to recently released large language models." It goes on: "No request changes are needed - the modelVersion field in the response now returns v11.0." Copyleaks' testing-methodology page lists the V11 test date as September 10, 2026 and the publish date as September 28, 2026.

What the note doesn't say

This is documentation for developers, written for the software other products are built on. It says the new model is live on the API. It doesn't say when Copyleaks' own website, or the Copyleaks report built into a school's learning platform, made the same switch. Those dates may match. The note doesn't promise that they do, so it's safer to treat September 22 as the API date and nothing more.

There's a second wrinkle that's easy to miss. On October 20, 2026, Copyleaks moved German, Spanish, French, Italian and Portuguese detection to "new, dedicated models" and said the modelVersion field "stays v11.0." English and the other supported languages are "unchanged," by Copyleaks' account. If you write in one of those five languages, the version number on two results can match even though different models read them.

Pangram: Pangram 4 due to become the default after September 30, then Pangram 4.1 on October 19

Pangram's dates are tidy on paper and a little slippery in practice. Pangram 4 was announced on July 29, 2026, and the Pangram 4 model card says it was "released by Pangram Labs on July 29, 2026." But being released wasn't the same as being the default. The migration guide, dated August 6, says "Pangram 3.3.2 is still the default and will remain live until September 30, 2026," and its timeline reads "July 29, 2026: Pangram 4 released through the API" and "Through September 30, 2026: Pangram 3.3.2 remains live and continues to be the default model." The launch post says the same thing more plainly: "Default requests that don't specify a model will continue routing to Pangram 3... until Pangram 3 is deprecated."

So, by Pangram's own account, software that called Pangram without naming a model got 3.3.2 through the end of September. The guide calls September 30 "the end of the announced Pangram 3.3.2 transition period." None of the Pangram pages we opened says the switch actually happened that day, and the API reference describes a default setting that "follows Pangram's current default model" without naming which model that is. These pages are also about the API. None of them says which model Pangram's own website dashboard used between July 29 and September 30.

Then came Pangram 4.1: "an AI detection model released by Pangram Labs on October 19, 2026. It is the successor to Pangram 4." Two of the listed changes bear directly on rechecking. Pangram says predictions are "more stable under small edits," with unintended verdict flips across 29,103 paired edits falling from 1.08% in Pangram 4 to 0.15% in 4.1. That's Pangram's internal figure. It also says the "Dashboard UI has changed to reflect less certainty for text under 150 words." A short passage scanned before October 19 and again after it may look different on the screen even where the underlying judgment hasn't moved much.

One visible sign of a model change

The migration guide says Pangram 4 merges the old "Lightly AI-Assisted" and "Moderately AI-Assisted" segment labels into a single "AI-Assisted" label, so two results with different label sets came from different models.

Turnitin: an October 2026 English model; old reports aren't rescored

Turnitin's guide page for its AI writing detection model has an October 2026 entry: "We have improved AI detection in our English AI writing model while minimizing false positives." It adds that there are no settings to change and "no visual changes to the AI writing report," which means nothing on the report itself will tell you a new model ran. Then comes the line that matters most here: "This release will not retroactively update previously-generated AI writing reports. To view the AI writing score generated by this update, existing submissions must be resubmitted."

The page gives a month and no day, so this post doesn't give a day either. If you're comparing two Turnitin reports from October, you can't place either one before or after the change with any confidence from Turnitin's note alone. (For an earlier round of this, see our post on Turnitin's previous AI update.)

Why scans from before and after a model change can't be compared

A score is a reading by one model on one day. Change the model and you've changed the instrument, much as you would if you weighed yourself on two different scales and blamed the difference on lunch.

That cuts both ways. A drop after September 18 isn't proof that a rewrite worked, and a rise isn't proof that it failed. As HumanPen puts it, "The two numbers mix your edits with a model change, and nothing in either result separates them."

Old results also stay the way they were. Turnitin says so directly. GPTZero, Copyleaks and Pangram don't say they rescore old results either. None of their notes say so, which isn't quite the same as saying they never do, but it means a saved result is best treated as a snapshot of whatever model ran that day.

Laid end to end, this fall's break points look like this: September 18 to 20 for GPTZero 4o, September 22 for Copyleaks V11.0, September 30 for the scheduled end of Pangram 3.3.2 as the API default, October 7 for GPTZero 4.1o, an unspecified day in October for Turnitin's English model, October 19 for Pangram 4.1, and October 20 for Copyleaks' five non-English languages. If a pair of scans straddles any of those, the comparison is already compromised. And different detectors disagreeing about the same file is a separate problem with its own causes, covered in which detector is scoring your text.

How do you retest fairly?

Scan the old draft and the new draft on the same day, with the same tool, in the same mode or setting. Save dated screenshots that show any model name or version the result displays, such as Copyleaks' modelVersion or Pangram's version, keeping in mind that Copyleaks' October 20 change kept the same version number. Compare only pairs that sit on the same side of a change date.

If you run drafts through StealthGPT and then check them on a detector, that same-day pair is the only comparison that tells you anything about the rewrite. For the full check across several detectors, follow the GPT humanizer workflow. And if the question underneath all this is whether any of these scores can be trusted, that's a longer conversation, and we've written about how accurate AI detectors are.

FAQ

What is GPTZero 4o?

It's GPTZero model 4.10b. GPTZero's release notes say it was released September 18, 2026, and GPTZero's announcement says it became the default for all users on September 20. GPTZero then released model 4.11b, which it calls "4.1o," on October 7.

Why did my GPTZero score change?

If your two scans fall on either side of September 18 to 20 or October 7, 2026, a different GPTZero model may have read each one. Rescan both versions on the same day to compare them.

Does Turnitin update old AI reports when the model changes?

No. Turnitin's October 2026 note says the release won't update earlier AI writing reports. A submission has to be resubmitted to get a score from the new model.

When did Pangram switch to Pangram 4?

Pangram released Pangram 4 through its API on July 29, 2026, but kept 3.3.2 as the default through September 30, according to its migration guide. Pangram 4.1 followed on October 19.

Which Copyleaks model is checking my text?

Copyleaks released V11.0 on September 22, 2026, and its API returns the model version with each result. Whether your school's version shows that number depends on the system it runs in, and for German, Spanish, French, Italian and Portuguese, the October 20 update changed the model without changing the number.