Thanks Audrey Turnitin Team for the question. I think we're all aligned on the need to step back from the score as a definitive indicator for judgement, and instead, see the technology as a tool to open a dialogue with students. We've mentioned it in prior posts, but we intentionally optimized our detection tool to minimize false positives as a guardrail to protect students from being falsely accused of AI. 

These are opportunities for all of us - educators, students and companies like Turnitin - to learn what using AI responsibly means, and how it can translate back to the learning process. We understand the importance of maintaining trust between educators and students. With that lens, we are exploring how we might bring more insight and interpretability to the AI detection report, enabling educators to have more fruitful conversations with students, and diffuse the potential for negative dialogue and mistrust.  

We look forward to working with you as we continue to evolve our tools to better support your needs!

Hi Tamer Abdullah

Thank you for sharing this experience. While we are continuously monitoring our tool's efficacy to keep our False Positive Rate (FPR) below 1%, the tool can make some mistakes. Some tools that generate AI content are better than others at overlapping human writing. When we train our model to learn from these tools, there is the opportunity that we'll incorrectly identify that overlapping text. When content is less complex, very uniform in structure and has perfectly grammatical, if not robotic, sentences, these signals can indicate potential AI content.

It's also possible that the cited text was short enough for us to find some signals, but they may be weaker than others in the paper because of the length of the sentences. All predictions come with a spectrum of confidence, and we are researching ways we can improve our detection report to make it easier to interpret how we are detecting the content. We're excited to deliver more of these usability updates in the coming months!

Another area we're hearing more about is when students over-edit to try and "perfect" their writing, they can unintentionally reduce or eliminate their unique, natural voice, making the text read as more synthetic. This is admirable, but risks the student's ability to come across as original.

Finally, even generalized tools that are acceptable to use in academia, like grammar and citation checkers are more and more being powered by AI, leaving their fingerprints behind for detectors like ours to pick them up.

All these scenarios reinforce for me that no detector should be used as a sole decision point. They are entry points to conversations about how a student embarked on their writing process. Did the student use any tools to help them format the citation?  If not, were they editing and polishing their document excessively? Our Chief Product Officer, Annie Chechitelli, recently posted an article on LinkedIn that resonates so well with me. She mentions tools like health screening, metal detectors and weather predictions as guides to help me decide my next course of action. This is how we see our detection tool being a valuable guide for educators. Hopefully you find the article helpful as well!