Turnitin Turns Off Timnit

Following an unfortunate incident, an aggrieved student reached out to the director of the Distributed Artificial Intelligence Research Institute (DAIR) and former Google ethicist, Timnit Gebru, narrating the ordeal of being graded zero in an assignment. At the root of the entire blowout was evaluation platform Turnitin that claimed that 67% of the paper was written by AI, which the student denied. 

“How many people’s lives are being ruined like this. This should be unacceptable,” tweeted Gebru. 

She said that the student currently stands at a 4.0 GPA and is a member of the President’s Honor Roll at the University. “I have tried understanding Turnitin’s methods of distinguishing AI and I still do not see how it fits within my writing,” the student wrote to Gebru. 

However, none of this is new. In another incident, a professor at the University of Texas failed the entire class because he thought the students had cheated from ChatGPT. The professor attempted to use ChatGPT to check if his students had plagiarised their assigned essays or if they had submitted original work. However, ChatGPT made an erroneous identification, flagging the students' essays as being generated by a computer program.

Gebru had previously argued that AI systems like ChatGPT and Google Bard lack the capacity to comprehend the meaning or significance of the words they process, no matter how convincing their language is. Her former colleague Alex Hanna, also, called these platforms “Bulls*** generators.” 

False Positives 

Timnit seem to have taken the conversation to a whole new level, questioning the working of Turnitin’s AI detection tool and its impact on students. 

Turnitin, known for checking plagiarism for years, is currently being used in 10,700 secondary and higher-educational institutions, assigning “generated by AI” scores and sentence-by-sentence analysis to student work. Turnitin’s AI writing detection model is trained to detect content from the GPT-3 and GPT-3.5 language models, which includes ChatGPT. Turnitin says that text generated by ChatGPT follows a pattern and is predictable while on the other hand human writing tends to be unique and unpredictable.

However, this method cannot be trusted blindly as it merely predicts if the text is generated by AI or not. In a blog post, Turnitin accepted that their model is not foolproof and tends to make mistakes. These mistakes can have serious consequences for students. They can damage their academic standing, which, in turn, might impact their future career prospects. Besides, such incidents take an emotional toll on students, causing them stress, anxiety and mental trauma as they are suddenly thrown into defending the authenticity of their work.

Turnitin has two matrices to determine if text is AI generated. First, on the document level and second, on the sentence level. On the document level, Turnitin in has 1% chance of generating false positives and on the sentence-level, false positive rate is about 4%. This means that the specific sentence it highlights may be human-written 4 times for every 100 highlighted sentences. Even though a 4%, or for that matter 1% error rate might seem insignificant, but every wrongful accusation of cheating can result in disastrous consequences for a student.

According to Turnitin chief product officer Annie Chechitelli, in cases where it detects less than 20% of AI writing in a document, there is a higher case of false positives. 

Meanwhile, the Washington Post carried out a test on 16 samples of real, AI-fabricated and mixed-source essays to run past Turnitin’s detector. Turnitin identified only six out of 16 samples, completely failed on three, and showed 8% plagiarism on an original essay. 

Masking their inability to assess accurately, Turnitin said that the false positive rate for AI writing detection is not zero and the instructor will need to apply their professional judgment, knowledge of  students, and the specific context surrounding the assignment. 

Turnitin's FAQ page only mentions that it is trained on GPT-3 and 3.5 models without providing further details about the specific parameters or models used. In contrast to Grammarly, which offers links to the original sources of copied content, detecting AI-generated content is more complex since it doesn't exist elsewhere but is predicted to be generated by AI. The swift implementation of Turnitin's AI detection software in schools raises concerns about the accuracy and fairness of its testing. 

What’s the solution? 

Besides Turnitin, GPTZero is known for quickly and efficiently detecting whether a text is ChatGPT-written or human-written. There are a slew of tools that detect AI generated content. Some of them include GPTZeroX, Detect GPT, Originality.ai and others. Again, the fairness and effectiveness of these models still remain dark and scary. 

It is crucial to monitor Turnitin's software to ensure it offers precise and impartial evaluations while minimising any potential harm to innocent students who might get caught in its system. 
It also becomes imperative for teachers and educators to explore new methods to check AI generated content, or eliminate it entirely. “I was talking to my cousin in high school about ChatGPT,” said the founder of healthcare startup Rupa Health, Tara Viswanathan, adding that they must now submit homework via Google Docs so the teacher can view the history to see if they really wrote it or not.