Turnitin AI detection accuracy, updated test data

Turnitin updates their AI detection model regularly, so accuracy numbers from 2024 or even early 2025 are outdated. I ran fresh tests in April 2026 to get current numbers.

Test Methodology

40 essays total: 15 human-written by current university students, 10 ChatGPT-4o generated, 5 Claude 3.5 generated, 5 Gemini generated, and 5 mixed (human-written with AI-assisted sections).

Results by Category

Human-written (15 essays): 13 correctly identified as human. 2 false positives (13% false positive rate). Both false positives were from ESL students with very formal writing styles. Scores were 18% and 22% AI respectively.

ChatGPT-4o (10 essays): 9 correctly identified as AI. 1 missed (10% false negative). The missed essay was a very short, simple topic that didn’t have enough statistical signal.

Claude 3.5 (5 essays): 4 correctly identified. 1 missed. Claude’s output tends to be slightly more varied than ChatGPT’s, which may explain the miss.

Gemini (5 essays): 4 correctly identified. 1 missed. Similar pattern to Claude.

Mixed content (5 essays): This is where Turnitin struggles most. Scores ranged from 12% to 58% AI with no consistent pattern. The tool seems to average across the document rather than precisely identifying which sections are AI.

Year-Over-Year Improvement

Compared to my tests from 2025, Turnitin’s overall accuracy has improved by about 5-8 percentage points. The biggest improvement is in detecting Claude and Gemini output, which they previously struggled with more than ChatGPT.

The ESL Problem Persists

Despite improvements, the false positive rate for ESL writers remains higher than average. Turnitin has acknowledged this issue publicly but hasn’t fully resolved it. This is something every school should consider when interpreting results for international students.

Bottom Line

Turnitin is the most accurate AI detector available to institutions in 2026, but it’s not infallible. An 87% overall accuracy rate means roughly 1 in 8 results may be wrong. Schools should always combine detector results with human judgment.