Comparisons

Academi.cx vs GPTZero Accuracy Test

We compared each detector’s overall AI scores and sentence-level highlights with Turnitin’s results across five essays.

Brayden · 2 min read

GPTZero markets itself as the most accurate AI detector, claiming 99% accuracy in distinguishing AI-generated text from human writing. Aimed at educators, students, and writers, it promises a reliable way to identify AI use in a document. But how closely do its results match Turnitin? In this comparison, we put GPTZero and Academi.cx to the test to find out.

Watch the full comparison on YouTube

How the test worked

Five essays were sourced from IvyPanda, a free essay database with no affiliation to Academi.cx. Selected paragraphs in each essay were rewritten using ChatGPT to introduce AI-generated content. The same version of each document was then submitted to Academi.cx, GPTZero, and Turnitin. We compared how closely Academi.cx and GPTZero matched Turnitin’s overall AI scores and sentence-level highlights.

The results from each detector

Reported results across five documents. Academi.cx scores: 96%, 50%, 32%, 38%, 56%. Predicted Turnitin ranges: 88–100%, 43–54%, 23–36%, 30–43%, 47–60%. Turnitin scores: 100%, 45%, 31%, 37%, 41%. GPTZero displayed AI probabilities: 2%, 25%, 0%, 0%, 9%. Separate lanes use blue Academi.cx circles and ranges, gray GPTZero squares, and navy Turnitin diamonds. GPTZero probability is a different metric from percentage of text flagged.
  • Document 1: Academi.cx detected 96%, predicting 88 to 100%. Turnitin returned 100%, inside the range. GPTZero’s probabilities were 2% AI, 5% mixed, and 93% human.
  • Document 2: Academi.cx detected 50%, predicting 43 to 54%. Turnitin returned 45%, inside the range. GPTZero’s probabilities were 25% AI, 9% mixed, and 66% human.
  • Document 3: Academi.cx detected 32%, predicting 23 to 36%. Turnitin returned 31%, inside the range. GPTZero’s probabilities were 0% AI, 7% mixed, and 93% human.
  • Document 4: Academi.cx detected 38%, predicting 30 to 43%. Turnitin returned 37%, inside the range. GPTZero’s probabilities were 0% AI, 2% mixed, and 98% human.
  • Document 5: Academi.cx detected 56%, predicting 47 to 60%. Turnitin returned 41%, 6 points below the range. GPTZero’s probabilities were 9% AI, 14% mixed, and 77% human.

Where the sentence highlights differed

Academi.cx's sentence highlights generally stayed close to Turnitin across the first four essays, with the main differences coming from additional sentences we flagged or lower AI likelihoods assigned to text Turnitin highlighted. GPTZero picked up some of the same passages, particularly in the second and fifth documents, but also marked whole paragraphs as human where Turnitin had flagged AI.

Average numerical deviation of displayed scores from Turnitin across five documents: Academi.cx 5.2 percentage points; GPTZero AI probability 43.6 percentage points. Blue and gray rounded bars share the same scale and zero baseline. GPTZero reports AI probability, so this measures displayed-score agreement rather than its detection error.

The fifth document fell outside the range

The final essay was the only one where Turnitin's score fell outside Academi.cx's predicted range, with additional highlighted sentences pushing our estimate above the actual result. Several passages toward the end were identified by both Academi.cx and Turnitin, while GPTZero caught most of one paragraph near the conclusion but left out other passages Turnitin flagged.

The final scorecard

Academi.cx gave a closer preview of Turnitin across the five essays, with all but the final result falling within its predicted range and its sentence highlights following more of the same text. GPTZero gave human writing the highest probability on every document despite the AI-generated sections we had added, and its highlights left out paragraphs that appeared in Turnitin's results.

Related reading