What changed in Turnitin's April 2026 update

Turnitin pushed an AI-detection update in early April 2026. Three real changes:

Paragraph-level scoring

Previously the AI score was computed by averaging across sentences. The new version classifies whole paragraphs as a unit, which means a single AI-sounding paragraph in an otherwise human essay can flag the entire document.

Practically: if you wrote 90% of an essay and pasted one ChatGPT paragraph for the conclusion, the old detector might smear that signal across the document and produce a 25% AI score. The new one scores that single paragraph at ~95% AI and surfaces it directly to your teacher.

Wider model coverage

The classifier now recognizes outputs from:

• GPT-5 (released March 2026)
• Claude 4 (mid-March)
• Gemini 3 (February)
• DeepSeek-R2 (January)

Older detector versions were 6–9 months behind their target models. This release closes most of that gap. If you've been using a humanizer that hasn't shipped an update in Q1 2026, assume it's flagging now.

False-positive guard

A new "human signal" pass looks for patterns that AI rarely produces:

• Typos in the middle of well-spelled sentences
• Half-finished thoughts that loop back
• Inline asides like "honestly, this is the part I'm least sure about"
• Personal references that don't fit the formal register

Documents containing these score lower regardless of other signals. Turnitin's claim is that this reduces false positives on rushed but human work.

What this means for students

You can't just sprinkle asides into a polished AI draft and clear it. The detector models the whole paragraph as a unit, so the AI fingerprint of the surrounding sentences still wins.

The fix is structural. Sentence variance, real punctuation breaks, and intentional rule-bending across the entire piece, not just decorated at the start.

Humanixio's V2 fine-tune was retrained against this update on April 10, 2026. We hold a 4–7% AI score on the new Turnitin classifier in 18 of 20 test essays (n=20, 500–2,500 words, mixed disciplines). The two outliers are heavily citation-dense academic samples. Citations get masked but their surrounding sentences sometimes inherit AI fingerprint.

Working on it.