Can AI Writing Really be Detected? The Problem with AI Checkers in Schools
- Nina Ranieri
- 4 hours ago
- 3 min read
Artificial Intelligence has quickly entered schools globally, providing students with new ways to research, brainstorm, and complete assignments. Because of this, many teachers have turned to AI detection software to prevent cheating and maintain academic integrity. However, an increasing amount of research shows that these tools are unreliable, responding with both false negatives and false positives, which can harm students academically. This raises serious concerns about how schools attempt to regulate AI use. AI detection tools are unreliable, easy to bypass, and risk unfairly punishing students.
AI detectors do not understand writing or its meaning. Instead, they look for patterns that indicate it was made by AI. The patterns AI checkers look for are predictability and variation in sentence structure. They analyze the structure of the text instead of comparing it with existing texts. The problem with AI checkers relying on these patterns is that human text can be predictable and structured, especially in formal, academic, or legal writing, causing it to be marked as AI. These detection systems are based on the probability of a text being written by Artificial Intelligence, not proof that it was.
Research shows that AI detectors often fail to correctly identify both AI-generated and human text. Many independent studies have tested popular detection software used in schools, such as Turnitin and GPTZero, to measure their real-world accuracy. Some studies show that the overall detection accuracy for AI text can be as low as 27.9%, with the best tool reaching only 50% (Springer, 2023). This means that in many cases, AI-generated writing is not detected at all, allowing it to pass as human work. Although companies often advertise accuracy rates between 80 and 90 percent, testing frequently shows much lower results, often between 50 and 70 percent. Popular AI detectors such as GPTZero have also produced high numbers of false positives, incorrectly labeling human writing as AI-generates, making them unreliable in academic settings.
Even when AI writing is initially detected, small revisions can quickly fool the software. In one investigation reported by MIT Technology Review, AI-generated text was correctly detected only 74% of the time. However, after small edits, accuracy fell to 42%. This shows that even small revisions, such as rearranging sentences or changing wording, can cause detectors to fail. Students often edit or paraphrase AI-generated text before submitting, which further reduces detection accuracy. Additionally, paraphrasing tools and humanizers are designed to make AI text appear more human, making detection even harder. As a result, detection tools struggle most with mixed writing that many students submit.
Detection tools also create serious risks by incorrectly labeling human writing as AI-generated. In some cases, detectors even misidentify well-known historical documents as AI-written. A biblical text sample was labeled as 88.2% AI-generated by a detector, and parts of the U.S. Constitution were also flagged as AI. These mistakes show how unreliable these detection tools can be, especially when accusations of academic dishonesty may result in repercussions. Because many students face serious consequences based on these tools, many researchers warn that AI detection software should not be used as the only evidence of AI use or academic dishonesty.
Given that widespread AI use is relatively new, schools are still adapting and creating policies for it. As of early 2026, about 92% of university students have reported using AI. Its adoption was extremely fast, without allowing schools to properly develop AI policies. As a matter of fact, nearly half of schools lack AI policies, and only 40% of US states are providing guidance to schools. Because AI use is widespread, students need to be taught how to use it responsibly rather than simply telling them not to use it. Schools are shifting towards guidance on AI instead of prohibiting it.
Overall, AI writing cannot be detected with 100% certainty right now, using detection software presents more problems than solutions. Studies indicate that these products are spotty, can be bypassed, and may produce false positives that negatively impact students. Nonetheless, it is already too late to ban AI, as it is a part of everyday academic life. After all, rather than relying on flawed detection systems, schools may need to emphasize teaching responsible use of the technology and redesign assignments. The problem for education, as AI continues to develop, is not how to eliminate AI but how to transform teaching and how student work is assessed so that learning, creativity, and critical thinking remain the center of education.
Sources:
Comments