Why current AI plagiarism detectors aren't reliable

Looking at revisions of a document is the reliable way to determine originality

TL;DR Establishing the authenticity of academic work is crucial for attracting and nurturing outstanding talent and maintaining the integrity of education. Current methods for detecting originality, such as Turnitin and GPTZero, are fallible and susceptible to circumvention, which can result in unwarranted plagiarism allegations. To create a more dependable and precise detection system, it is essential to capture every draft involved in the creation of the final document.

Importance of originality in academia

It is essential for academic institutions to evaluate the originality of students' work to attract and cultivate exceptional talent. Learners who exhibit innovative thinking in their assignments are more likely to perform well academically and acquire a profound comprehension of their subjects. This, in turn, enables schools to better equip these students for their professional lives, solidifying an institution's status as a distinguished center of learning.

Three groups of stakeholders benefit from accurate evaluation of students work:

  1. Companies: Organizations can gain substantially from evaluating the uniqueness of students' work, as this enables them to identify and recruit top-performing students from various academic backgrounds. Original thinkers possess problem-solving and creativity skills that help them better find ways to solve an organization’s specific problems. 
  2. Students: For learners, demonstrating critical thinking in their projects not only allows them to assess their grasp of a subject, but also distinguishes them from their classmates, earning them merited recognition
  3. Educators: For educators, being able to correctly evaluate students' grasp of a subject, helps them evaluate their instructional techniques and fine-tune their methods to better accommodate the needs of their students. 

Consequences of inaccurate originality detection

The ramifications of misjudging the authenticity of ideas within an educational setting can be significant. The decisions can be between awarding a student an A+ grade or attributing allegations of plagiarism to a student's permanent academic record. Overlooking cases of dishonesty not only compromises the validity of the scholastic process, but it also disheartens genuine learners who diligently produce authentic work. Falsely categorizing honest students as cheaters can lead to serious consequences in their academic and professional journeys, diminishing their self-assurance and potentially damaging their reputations forever. 

Current tools for determining originality

There a number of different services used to determine originality, below are few examples:

All these services leverage ML models that predict how likely it is that a piece of text was generated by AI. These models are trained with thousands of text samples that have been pre-identified as either AI-composed or human-authored, enabling the models to discern patterns in which generative AI assembles words to create sentences. After training, the models can make predictions about a new text piece concerning its likelihood of being crafted by AI or a human.

Why current tools are not reliable

Let’s look at claims made by some companies that detect plagiarism. 

"Turnitin’s ChatGPT and AI writing detection capabilities go live with 98% confidence rating".

Now let’s a look at a quick video that show how easy it is to bypass detectors

It’s safe to say the 98% confidence levels claimed are not accurate. In fact, there have been a number of scientific articles that demonstrate how flawed existing detectors can be. For example a recent study from University of Maryland finds that:

“Empirically, we show that paraphrasing attacks, where a light paraphraser is applied on top of the generative text model, can break a whole range of detectors, including the ones using the watermarking schemes as well as neural network-based detectors and zero-shot classifiers. We then provide a theoretical impossibility result indicating that for a sufficiently good language model, even the best-possible detector can only perform marginally better than a random classifier.” 

OpenAI also mentions that 

“Our classifier is not fully reliable. It should not be used as a primary decision-making tool, but instead as a complement to other methods of determining the source of a piece of text.”

Here are few other video links that show how a document can be manipulated to bypass detectors

Current workarounds

Given that false accusations of plagiarism can carry significant ramifications on a student's (and institutions) reputations, some instructors are taking additional steps to verify any claims of plagiarism made by originality detection tools. Specifically, students are show their draft history to prove their work. For example, recently one student in UC Davis was accused of plagiarism  

"To appeal his professor's accusations to university officials, the student shared a Google document history of his exam writing that showed proof he didn't use AI and a slew of research on the fallibility of GPTZero and other AI detection tools, according to school records."

This is actually very similar to what instructors in other parts of the country are doing to determine originality of submitted papers.

The reliable way to determine originality

How do you get more accurate results from a Machine Learning Model, ask any data-scientist and they will say, you need two things:

  1. Better Machine Learning model
  2. Better data 

Current solutions might claim that they yield more precise outcomes, when in reality, they primarily concentrate on enhancing their models, which entails employing a more appropriate machine learning model (e.g. GPT) and additional training data for improved fine-tuning. However, these solutions do not attempt to utilize more data from the document composition process to generate more accurate and dependable predictions. We believe that the key lies not only in refining the model, but more significantly, in capturing all the drafts employed to produce the final version.