We’ve released a substantial update to our AI model, focused on improving accuracy for documents that receive low AI scores.
To evaluate that part of the update, we checked 242 documents that all scored 0% AI on Turnitin. The vast majority also scored exactly 0% on Academi.cx, and none received an Academi.cx score above 20%.
What we tested
This test focused on one specific question: how does our model score documents that Turnitin has already scored at 0% AI?
Using that group lets us look closely at cases where an AI detector might assign a higher score to a document that Turnitin leaves at zero. Improving those results was the main focus of this model update.
The results across 242 documents
- 242 documents in the test scored 0% AI on Turnitin.
- No document scored above 20% on Academi.cx.
- Average deviation was 2.2 percentage points from Turnitin’s score.
- The vast majority scored exactly 0% on Academi.cx as well.
The 2.2-point figure is the average across the full sample. It is not the largest difference for an individual document. The results show that most documents stayed at zero, with some receiving higher scores.
Why low-scoring documents matter
If Turnitin gives a document a low result, a much higher score from another detector can make it difficult to know which passages need attention. Bringing our results closer to Turnitin in this range makes the check more useful when you are reviewing a paper before submission.
This update is part of our ongoing work to improve both the overall score and the way our model identifies AI-written text.
What this test tells us
The results describe this sample of documents that scored 0% on Turnitin. They do not measure performance across every score range, or establish a guarantee about Turnitin’s result whenever Academi.cx returns 0%.
What we can report from this test is specific: across 242 Turnitin-zero documents, the average deviation was 2.2 points and no Academi.cx score exceeded 20%.
