Can humanised AI text still be detected in 2026?

Three questions keep turning up in AI answer engines: can AI still be detected after using a humaniser, can humanised AI be detected, and is humanised AI text detectable at all. They are the same question, and the honest answer sits between the marketing claims on both sides. Humanisers and paraphrasers do lower detection rates in published research. Detector vendors are also retraining against them. The result is a moving target where the outcome depends on which detector, which tool, how long the text is and how much a person actually rewrote it.

This guide pulls together what the published studies and vendor statements say. It reports no new testing; every figure below links to the page it came from.

Can AI still be detected after using a humaniser?

Yes, for three reasons that have nothing to do with any single tool.

First, humanisers leave their own fingerprints. A detector does not look for one tell, such as an em dash; it reads statistical patterns across a passage. When a humaniser rewrites text, it swaps one pattern for another. Turnitin's chief product officer told Plagiarism Today in August 2025 that humanisers "leave a statistical trace that can be learned and detected", and that Turnitin had trained its model on the patterns of leading tools. That is a vendor claim, but it describes how the arms race works.

Second, partial edits leave AI-shaped seams. Most people rewrite the opening and skim the rest. Turnitin's own data, published in a May 2023 update, found 54% of its falsely flagged sentences sat right next to actual AI writing. Mixed documents are where highlighting clusters, so an untouched middle section still draws attention.

Third, some defences do not read style at all. The 2023 DIPPER study proposed a retrieval defence, where a model provider searches its own log of past outputs for a close match. Using a database of 15 million generations, it detected 80% to 97% of paraphrased generations while classifying only 1% of human text as AI (Krishna et al., 2023). That method needs the provider's cooperation and is not what a university checker runs today, but it shows paraphrasing is not a permanent shield.

For a plain-English primer on perplexity and burstiness, the two signals most detectors start from, see our guide to how AI detection really works.

What research shows about paraphrasing and detection

The research record since 2023 is consistent on direction and inconsistent on size. Paraphrasing usually lowers detection, sometimes dramatically, and occasionally does nothing or even makes detection easier.

DIPPER (2023). Krishna and colleagues built an 11 billion parameter paraphraser and ran its output past several detectors at a fixed 1% false positive rate. For GPT-2 XL text, watermark detection fell from 100% to 57.2% and DetectGPT from 70.3% to 4.6%; for GPT-3.5 text, OpenAI's classifier fell from 30.0% to 15.6% (arXiv 2303.13408).

Recursive paraphrasing (2023). Sadasivan and colleagues paraphrased the same passage several times over. On passages of about 300 tokens, watermark detection dropped from 99.3% to 9.7% at a 1% false positive rate, and a zero-shot detector's AUROC fell from 96.5% to 25.2% (arXiv 2303.11156). Even the retrieval defence above fell from 100% to below 60% accuracy against this attack.

Academic integrity testing (2023). Weber-Wulff and colleagues tested 12 public tools plus Turnitin and PlagiarismCheck. For ChatGPT text run through the Quillbot paraphraser, overall accuracy was 26%, meaning most paraphrased AI text went undetected; for manually edited text, roughly half went undetected (arXiv 2306.15666). The authors called the tools "neither accurate nor reliable".

RAID benchmark (2024). The largest shared test so far covers over 6 million generations from 11 models, 8 domains and 11 adversarial attacks, scored against 12 detectors at a 5% false positive rate (Dugan et al., arXiv 2405.07940). Its finding is the most useful one for this question: attacks hit detectors unevenly. The chart below shows five detectors with no attack and after DIPPER paraphrasing.

Read that chart carefully. In RAID's 2024 results, paraphrasing raised Originality's accuracy from 85.0% to 96.7%, while GLTR fell from 62.6% to 47.2%. The same rewrite that slips past one checker can make another more confident. The authors suggest detectors sometimes improve when an attack nudges text closer to their training data. A note of caution on age: every study above tested 2023 and 2024 models and detectors, so treat the numbers as evidence of how detection behaves, not as a forecast for any specific tool today.

What Turnitin says it now detects

Turnitin is the checker behind many university AI policies, including the Bristol guidance cited below. Here is what it has stated publicly, with the date of each claim.

• Bypasser detection, 27 August 2025. Turnitin's press release says its AI writing detection now includes "AI bypasser detection", aimed at text that may have been intentionally modified by humaniser tools. It sits inside the Turnitin Originality add-on and the AI add-on for iThenticate 2.0.
• English only. A footnote in the same release says the bypasser feature was trained and tested for English only.
• No new screen. The University of Bristol's September 2025 staff note explains the change sits inside the existing AI writing report, needs no assignment setting changes, and makes no visual change to highlighting. Reports generated before 27 August 2025 had to be resubmitted to use it.
• Earlier paraphrase detection. According to Plagiarism Today (August 2025), Turnitin added AI paraphrasing detection in July 2024, a year before the bypasser update.

What Turnitin has not published is an accuracy figure for bypasser detection. Its press release contains no detection or false positive rates for the feature, and Campus Technology's September 2025 report relies on the company's own statements. Plagiarism Today noted outside testing would be needed to judge it. So the fair summary is: Turnitin says it now looks for humaniser edits, but neither the announcement nor its press coverage cites an independent evaluation. For how its scoring compares with a popular rival, read our Turnitin vs GPTZero comparison, and for the 2026 state of play see Turnitin AI detection in 2026.

Why results vary by detector and by text length

Different detectors, different weak spots. RAID found that swapping a small share of words for synonyms cut Binoculars' accuracy by 36.1 points, while GPTZero lost only 0.3 points to a homoglyph attack that cost five other detectors an average of 40.6 points (RAID, 2024). A humaniser tuned against one checker tells you little about the next. If you use GPTZero, our guide to reading a GPTZero report explains what its sentence highlights mean.

Length changes the maths. Short passages give a detector little evidence either way. Turnitin raised its minimum from 150 to 300 words in 2023 because, in its words, accuracy "increases with a little more text" (Turnitin, May 2023). The same update said false positives were more common in documents scored under 20% AI, which now show an asterisk instead of a number, and in a document's first and last few sentences.

The writer matters too. Detection is not only about AI text slipping through; human text also gets caught. A 2023 Stanford study found seven detectors misclassified 61.22% of 91 TOEFL essays by non-native English writers as AI-generated, against near-perfect accuracy on US eighth-grade essays (Liang et al., arXiv 2304.02819). Simpler, more predictable phrasing reads as "machine-like" to perplexity-based tools. Our piece on AI detection and non-native English writers covers this in more depth.

Decision table: how detectable humanised text is in your situation

The table below summarises the evidence above by scenario. It describes likely detector behaviour, not guarantees.

Caveat: no detector score should be the only evidence of misconduct. The University of Bristol's guidance calls the AI score "an indicator rather than evidence of AI use".

Responsible use: edit your own draft and follow your institution's policy

A humaniser is a writing tool, not a way around academic rules. The legitimate uses are the ones where the words are already yours: smoothing a stiff draft you wrote, tidying prose after heavy self-editing, or fixing the flat rhythm that creeps in when you write tired. Used that way, the question of detection becomes less important, because you can explain every sentence and show how it got there.

Three habits keep you on the right side of most policies:

1. Read the rules first. Module handbooks differ. Some allow AI for brainstorming but not drafting; some require a short statement of how you used it. If the policy says disclose, disclose, whatever any tool claims.
2. Keep a paper trail. Draft in a tool with version history, save your notes and keep your reading list. If a score is ever questioned, Bristol's guidance says a referral would typically need "additional evidence"; your drafts are the counterweight.
3. Rewrite for meaning, not for a score. Add your own argument, your own examples and your own sources. That is what makes writing yours, and it is also what no detector or humaniser can supply.

If you want help with rhythm and readability on text you wrote, the free humaniser is built for that final polish. For a step-by-step editing method that works without any tool, see how to make AI-assisted writing sound like you.

Frequently asked questions

Can Turnitin detect humanised AI text?

Turnitin says it can detect text changed by leading humanisers, since 27 August 2025. Its press release describes "AI bypasser detection" built into the AI writing report, trained and tested for English only. Turnitin has not published accuracy figures for the feature, and no independent test has been published alongside it, so how well it works on any given tool remains unverified.

Does paraphrasing AI text stop it being detected?

Paraphrasing often lowers detection, but not always. A 2023 study cut DetectGPT's detection rate from 70.3% to 4.6% with a paraphrasing model. Yet the 2024 RAID benchmark found paraphrasing raised one commercial detector's accuracy from 85.0% to 96.7%. Results depend on the detector, the paraphraser and the text.

Can human-written text be flagged as AI?

Yes, false positives are documented. Turnitin put its sentence-level false positive rate at about 4% in May 2023, and a 2023 Stanford study found seven detectors misclassified 61.22% of non-native English essays as AI-generated. That is why universities such as Bristol treat an AI score as an indicator, not proof.

Is a short text easier or harder to detect?

Short texts are harder to judge accurately in either direction. Turnitin raised its minimum for AI scoring from 150 to 300 words in 2023, stating that accuracy increases with more text. On short passages, any detector's verdict should carry less weight.

Is it allowed to use a humaniser on my essay?

It depends entirely on your institution's policy for that assessment. Some courses permit AI use with disclosure, while others forbid it for specific tasks. Polishing text you wrote yourself is different from disguising AI output. Check your module handbook, disclose where required, and keep your drafts.