AI Detection Tools are Flagging Student Assignments, Even If They Aren't Generated

When Lavina Kashyap (name changed), a 24-year-old mass communication student in Bengaluru, submitted her research on building parasocial relationships on social media that she had worked on for over six months, she was aghast to find that it had been flagged ‘AI-generated’.

“I had submitted my research paper like any other assignment. A few days later, my professor told me it was flagged at around 70–75% AI. I was shocked because I had written it myself,” she tells AIM.

She felt the pressure to prove to her university that she hadn’t, in fact, cheated. “When a report already says ‘75% AI,’ it’s hard to change that perception,” she laments.

Kashyap then had to work around the detection tools by simplifying her writing and deliberately introducing errors in her paper. “I started wondering, should I write more simply, or change my style? Am I being penalised for writing well?”

Kashyap and a number of other students are finding themselves at the short end of academic reviews that rely heavily on detection tools to scan assignments, dissertations, and even PhD-level research papers for AI-generated content. Institutions found it necessary to place these mechanisms after students turned to ChatGPT and other AI chatbots en masse as a shortcut to generate entire assignments and draft essays.

However, as the adoption of AI detection tools grows, so do concerns about their reliability. 

A growing body of research suggests that these systems operate less like definitive detectors and more like probability engines. A study published on the peer-reviewed AJIRB platform found that while AI detection tools can identify patterns associated with machine-generated text, their accuracy drops significantly when evaluating structured academic writing.

The Role of AI Writing Tools

At the same time, AI-powered writing assistants are becoming increasingly common in academic workflows.

Platforms like QuillBot are widely used to improve clarity, grammar and structure, especially by non-native English speakers. Rather than replacing human effort, these tools often layer in refinements onto original writing.

“Most users aren’t looking to outsource their work—they’re trying to get their ideas across clearly,” asserts Eric Wang, Vice President of Product and AI at QuillBot.

But this introduces another layer of complexity. Even QuillBot, which boasts 75 million users worldwide, acknowledges that paraphrasing and editing tools can influence how AI detectors interpret text. Because detection systems rely on statistical patterns, any change in phrasing can shift the outcome of a detection result.

An independent analysis by Anangsha Alammyan, who completed her PhD from IIT Guwahati, found that AI detection tools like Quetext often produce inconsistent results when content has been edited or partially generated by AI, even after ‘humanising’ them using tools like GPTHuman. She found that AI detection tools falter when the context is nuanced.

“AI detection is all about probability. AI detectors check for statistical patterns like sentence structure, repetition, and commonly used phrases. That’s why you get percentages instead of a simple yes-or-no answer,” she wrote on Medium.

Despite these limitations, AI detection systems remain central to how institutions monitor academic integrity.

Turnitin, a widely used plagiarism detection and AI writing verification platform in India serving over 16,000 academic institutions, publishers, and corporations, compares submissions against a vast dataset that includes over 2.4 billion student papers, 54 billion web pages, and extensive academic publications. Its system is designed to identify similarities and flag potential AI-generated text. However, it acknowledges that the AI detector is an imperfect tool.

“Our tools are designed to support and not replace human judgement,” says Chaitali Moitra, Regional Director for South Asia at Turnitin, adding that the final judgement must always rest with the educator, who should combine the platform’s data with their own professional assessment of the student's work.

This reflects a broader reality. AI detection tools can highlight patterns and similarities, but cannot fully interpret context, intent, or the nuances of human writing.

Emerging Tensions in Academia

Academic writing, particularly research papers, often follows a formal, standardised style. Ironically, these are the same characteristics that AI detectors sometimes associate with machine-generated content.

A study by the University of Chicago Booth School of Business questions whether such tools are reliable enough for high-stakes academic decisions. The analysis highlights that AI detectors frequently struggle with nuanced, edited, or hybrid content, leading to outcomes that are difficult to interpret with confidence.

In practical terms, this means the same document could receive different classifications depending on the tool used.

S Sudarshan, Chair Professor of the Department of Computer Science and Engineering at IIT Bombay, echoes these concerns. He notes that while AI tools are now widely used by students, they may not always be misused. “Many researchers use these tools to improve their language and phrasing, especially in countries like India where not everyone is equally comfortable with English,” he explains.

However, he also warns of the over-reliance on detection systems. “Flagging for AI is a dangerous check. It can be extremely unfair and often unreliable. At best, these tools should act as a filter, not a final judgement,” he asserts.

He suggests that institutions should rethink their evaluation methods rather than relying heavily on detection tools. 

The tools also undermine the effort and time invested to furnish research. 

Anita Paul (name changed), a 2013 graduate from Bangalore University, recalls the effort required to write research papers without AI assistance. “I didn’t have any shortcuts. I had to sit with the material whether I liked it or not. I spent hours at the library, took notes by hand and even kept rewriting again and again,” she says.

She adds that while the process was exhausting, it led to a deeper understanding. At the same time, she acknowledges that tools are not inherently unfair. “Every generation gets better tools. But if you depend on them too much, you miss out on actually learning,” she explains. 

QuillBot, for instance, frames its role as a collaborator in the writing process and emphasises adherence to academic integrity policies. But even responsible use can affect how detection tools interpret text.

Paul’s experience also highlights a shift in how academic integrity is evaluated. “Back then, it was more about whether your professor believed in what you wrote. Now, it feels system-driven.

 Why Human Judgement Still Matters

Across research, industry and platform-level insights, one conclusion stands out. AI detection tools are not definitive. They can flag patterns, identify similarities, and raise potential concerns, but they cannot fully understand intent or context. 

At the same time, AI tools are hugely beneficial to students and researchers, and some academic institutions are beginning to understand that.

The Indian Institute of Science Education and Research recently allowed the default use of generative artificial intelligence in academic work, stipulating that if AI use is to be disallowed for any assignments or research papers, it must be specified beforehand. It formed guidelines after consultations with students and staff that apply to all documented output at the university. No attribution to AI is required.

Rather, to check student integrity, it will rely on viva-voce verification, shifting focus from detection to learning outcomes and the role of humans.

In a system where both writing and detection are increasingly shaped by artificial intelligence, the role of human judgement is not diminishing. It is becoming more important than ever.