The Challenge of AI in Education
Teachers at all levels of education are currently grappling with an unprecedented challenge: the pervasive influence of artificial intelligence (AI). This cutting-edge technology holds the potential to transform learning models and, in many instances, facilitate the understanding of complex subjects. However, there exists a fine line between employing tools such as ChatGPT as supportive resources and completely delegating task completion to automated systems. The widespread adoption of generative AI raises numerous questions within the educational landscape. One of the most pressing inquiries is: How can educators ascertain whether they are evaluating work that has been entirely or partially produced by AI?
In the United Kingdom, a recent survey conducted by the Higher Education Policy Institute in Oxford, involving 1,054 undergraduate students, revealed that approximately 94% of respondents utilised generative AI tools to assist with assessed work, while 12% admitted to directly incorporating AI-generated text into their submissions.
In the United States, a series of surveys targeting over 95,000 students across 20 universities estimated that 9% had employed AI in their coursework, fully aware that this was against institutional regulations. Similarly, in Mexico, over 60% of higher education students and faculty frequently use generative AI systems, albeit acknowledging that their use occurs within a context of inadequate training and a lack of clear regulatory frameworks, raising concerns about ethical and critical applications. Meanwhile, the question remains: how can one detect AI-facilitated academic dishonesty?
Innovative Detection Strategies
Beyond the capabilities of platforms like Claude, ChatGPT, or Gemini to label AI-generated content, educators are developing techniques to identify chatbot-generated tasks while also integrating this technology into their teaching methodologies. In a recent informative article for Nature, Katarina Zimmer highlights some of these innovative approaches.
Nicholas Mattei, an associate professor of Computer Science at Tulane University, allows his students to utilise AI in his History of Technology course. However, students are required to identify sources, apply three models of fuzzy logic to summarise them, and critically analyse the results before creating infographics and essays based on their findings. Furthermore, Mattei awards extra credit to students who successfully identify errors made by AI, such as non-existent references or infographics that misrepresent the sources.
Simulations Enhancing Understanding
In Canada, sociologist Daniel Silver from the University of Toronto Scarborough implements AI agents mimicking merchants in his classes to enhance students’ comprehension of concepts from Adam Smith’s seminal work, “The Wealth of Nations,” published in 1776. Previously, students were tasked with writing an essay on this text. By allowing the use of AI, Silver explains that the technology encourages students to explore various potential scenarios, such as the implications of individuals becoming unreliable. Students analyse the transcripts generated by the agents and write reflections on them. Consequently, Silver notes that it becomes easy to discern if students have requested AI to perform all steps simultaneously, as the system may simply fabricate information rather than accurately cite the transcripts.
Fostering Critical Evaluation Skills
Ruben Verborgh, a computer scientist at Ghent University in Belgium, permits the use of AI in his Web Development course, believing that this technology should be included in his students’ education. However, he assesses their skills through questions such as, “Why isn’t this webpage working?” or “Why is it slow?”—queries that AI systems do not always answer effectively. Verborgh’s experience suggests that while chatbots can respond, they often lack a solid and precise methodology.
Similarly, Etienne Roesch, a statistician and cognitive scientist at the University of Reading, UK, employs neuroactivity analysis snapshots and asks students to interpret them. He has found that when an AI resolves the request, “the response often bears no relation to what is shown in the screenshot, and that is noticeable,” allowing him to determine if the answer was generated artificially.
Process Over Product
Some educators focus on examining how the work was constructed rather than merely reviewing the final product. Yanjun Shen, a researcher in Ecological Geology and Engineering at Chang’an University in Xi’an, China, requests records of students’ interactions with AI tools to assess whether these technologies have expanded their knowledge or merely replaced their critical thinking. Similarly, Nikita Bezrukov, a professor of Linguistics and Communication at Columbia University, employs revision history platforms like Google Docs to monitor the evolution of texts. He permits the use of AI for grammar corrections or to enhance writing but does not allow it to generate complete works.
Oral Exams and Knowledge Verification
Oral examinations are used to verify that students comprehend and can defend their written submissions. Gianmarc Grazioli, a computational chemist at San José State University, allocates 70% of the grade to Python code and 30% to an oral exam where students must explain specific lines. This assessment method shifts the focus from the final product to the student’s ability to articulate their decisions and demonstrate mastery of the subject matter. Therefore, a response generated with the aid of AI does not necessarily pose a problem if the student can understand, question, and defend it.
AI Detectors and Alternative Methods
AI detectors continue to be utilised, yet some educators view them as merely supplementary tools, alongside erroneous references, abrupt changes between drafts, and other indicators. Ollie Thomas, a biologist at the University of Melbourne, warns of a sort of “arms race,” as AI tools evolve and can evade detection systems. Consequently, identifying artificially generated content does not solely rely on automated tools. Comparing different versions of a document, reviewing the sources used, and having direct interactions with students can provide additional insights into how a piece of work was produced.
Creative Approaches to Assessment
Will Teague, a historian at Angelo State University in San Angelo, Texas, employed what he termed the “Trojan Horse method,” which involves concealing an instruction interpretable by AI but nearly invisible to the student. By requesting that essays be written from a Marxist perspective, he identified 47 out of 130 papers as clearly produced by a chatbot. However, he deemed the experience unsatisfactory and subsequently turned to educational and discussion-based strategies.
This case also highlights the limitations of relying solely on surveillance mechanisms. While a technique may reveal the use of a chatbot, it does not necessarily enhance learning outcomes. As a result, some educators have opted to redesign activities to incorporate AI into the educational process without replacing the intellectual engagement of students.
Encouraging Personalised and Critical Thinking Tasks
Ultimately, some educators design assignments that require personalisation, reflection, and critical analysis. Risa Morimoto from SOAS University of London allows students to choose topics that interest them and write texts addressed to family or friends, which has enabled her to better discern their individuality in the submissions. Jonathan Vallano employs open-ended questions and activities in which students must identify research and articles from clues, read them, and summarise them in their own words.
Together, these experiences illustrate that the integration of AI within classrooms does not necessarily entail prohibition or unreserved acceptance of work produced with these tools. The strategies employed by educators to detect potential AI-related dishonesty aim to transform assessment methods: by analysing the creation process, demanding explanations, encouraging discussions, presenting challenging problems that cannot be resolved automatically, and requesting demonstrations of independent thought. While many of these tactics have become commonplace, their effectiveness remains uncertain in light of the rapid technological advancements.