Robot caught in a flashlight. Artifical intelligence plagiarism, cheating and ai detection concept.

AI detectors: An ethical minefield

Best Practices, Resources, Teaching, Technology Trends, Tips, Web Tools

code of conduct team people work together on paper document on laptop screen - vector illustrationGenerative AI use has been on the rise among faculty and students. Tools like ChatGPT, Gemini, Adobe Firefly, and Claude, among others, have transformed how students approach academic work. In response, some faculty have clamored for AI detectors to help them identify content they believe is AI-generated, with the goal of upholding academic integrity. However, these tools are far from perfect, and they can lead to unintended consequences for students. AI detectors’ false positive rates (and the consequent serious ramifications of false accusations) and the equity issues that arise from their use deserve careful scrutiny. Instead of relying on this flawed technology, faculty and institutions should use alternative approaches to navigating the challenges posed by generative AI in education. Ultimately, any approach should prioritize fairness, understanding, and promotion of the responsible use of AI.

What AI Detectors Claim To Do

Robot carrying a magnifying glassAI detection tools are marketed as solutions for identifying AI-generated content, particularly in educational settings, with the purported goal of upholding academic integrity. These programs rely on algorithms to differentiate between human and AI-generated text. Accuracy varies significantly between tools and depends heavily on the complexity of the text and the methods used to disguise AI-generated content. It’s important to point out that AI detectors are themselves a type of artifical intelligence.

False Accusations; Serious Ramifications

AI detection companies advertise high accuracy levels: Copyleaks boasts accuracy of 99.12%, Turnitin claims it’s 98% accurate, Originality.AI contends their accuracy is 98.2%, Trace GPT professes 93.8%, Winston AI brags 99.98%, and GPTZero claims 99% accuracy.Person sweating and looking worried with eyes in the shadow behind them

While companies claim false positive rates in AI detectors are low, any false positive comes with potentially serious consequences for accused students. In a Bloomberg test of two AI detectors (GPTZero and CopyLeaks), false positive rates were 1-2% (they could actually be higher) when a sample of 500 essays (written before generative AI’s release) were run through the checkers. Incidentally, detectors can also miss AI-generated writing, marking it as human-generated. With the rise of generative AI and AI detection tools, another category of technology has emerged: “AI humanizers,” designed to make AI-generated writing appear more human.

Returning to false positives, even a small error rate can add up. If a typical first-year student writes 10 essays, and there are 2.235 million first-time degree-seeking college students in the U.S., that would add up to 22.35 million essays. If the false positive rate were 1%, then 223,500 essays could be falsely flagged as AI-generated (assuming all were written by humans). That’s a lot of false accusations. In addition to potential psychological impacts on students (e.g., stress, anxiety), there are also material consequences, including academic penalties, loss of scholarships, and damage to future opportunities.

Equity Issues Abound

According to research, there are numerous ethical issues with AI detectors. For example, AI detectors disproportionately target non-native English writers. Additionally, black students are more likely to be accused of AI plagiarism by their teachers, according to a Common Sense Media report. Neurodiverse students are also more likely to be falsely flagged for AI-generated writing, and research has revealed biases against linguistic patterns and dialects in AI detectors. Woman with anxious face looking at laptop screen. The disproportionate rate of false positives against already marginalized student groups could have chilling effects in education and beyond. For instance, it could foster an atmosphere of distrust between faculty and students, discourage academic participation and engagement, and undermine the perception of fairness in assessment and disciplinary processes. Additionally, false accusations can increase educational inequities: students who are neurodiverse, Black, or non-native English speakers already face systemic barriers in education, and biased AI detection can exacerbate these inequities. Further, marginalized students may lack the resources to contest these claims, which would further disadvantage them and reinforce structural inequities. Other unintended consequences could include students using AI to get around detectors, AI detectors replacing educators’ critical judgment, and legal issues surrounding intellectual property and privacy (e.g., FERPA violations) and discrimination (e.g., Title VI, the ADA).

What You Can Do Instead

While it may feel “righteous” to use an AI detector because you feel like you have “proof” of student cheating, as we can see in the data, that’s a false sense of comfort. These tools are ubiquitous because they purport to solve a problem educators have. So, how do we combat student use of generative AI (and, yes, they are using generative AI)? Here are a few ideas:

  • Increase your own AI literacy: AI literacy means understanding how AI tools work, what their biases are, and what their ethical implications are. Stay informed about AI news, and actually use different geneative AI products so you know how they work, what they can do, and what they cannot do (yet).
  • Promote AI literacy among students: Help students understand how AI tools work, including their capabilities and limitations, how AI is being used today and how it might be used in the workplaces of the future. Engage them in discussing these issues and help them think critically about AI’s impact on their learning and future careers. Help them see how they can use AI to their benefit without allowing AI to replace them.
  • Teach ethical and responsible AI use to students: Make students aware of AI’s ethical issues (e.g., bias and discrimination, environmental impacts, intellectual property concerns, etc.). Try to use real-world examples, case studies, and data to show the ethical challenges of AI and promote responsible AI use to students.
  • Use critical judgment over outside tools: If you do choose to use an AI detector, don’t let it replace your own judgment. Evaluate your students’ work holistically. Get to know their writing style throughout the semester. Engage with students about their writing process before making accusations about their writing output.
  • Diverse human hands holding speech bubblesCreate open conversations about AI: Discuss AI among fellow faculty—share strategies and ethical considerations—and engage students in creating guidelines for AI use in your classes, which will help promote transparency, trust, and accountability.
  • Rethink assessment: Instead of wasting time trying to detect AI writing, rethink how you assess student learning. Star by designing meaningful assessments that students find intrinsically valuable, ideally grounded in real-world contexts to enhance relevance and engagement. Incorporate alternative formats like one-on-one conferences, self-reflections, or multimodal projects that resist AI use. Break larger assignments into smaller, more manageable components through scaffolding. This approach allows for incremental feedback and reduces students’ feelings of overwhelm when tackling substantial projects. Additionally, consider ethically integrating AI into assignments to teach students how to use it critically, including proper citation of AI-generated content. To further emphasize learning over grades, consider alternative grading methods that prioritize learning and growth. By deploying meaningful, well-supported assessments, educators can foster deeper learning, promote student engagement, and prepare students to navigate AI use in real-world contexts.

A.I. and human balance

AI detectors are often marketed as solutions for maintaining academic integrity, but their significant drawbacks often outweigh any perceived benefits. False positives can cause emotional and psychological harm, unwarranted academic penalties, and long-term consequences for students. These harms are disproportionately borne by marginalized groups, exacerbating existing inequities in education.

To address these challenges, educators should prioritize equity and fairness first by increasing our own AI literacy and then promoting responsible, thoughtful, and critical AI use among students. Rather than adopting a punitive approach, develop a balanced and informed perspective by designing meaningful assignments, facilitating open dialogue with colleagues and students, and emphasizing learning and growth over punishment. By taking these first steps, we can begin to meet the challenges of generative AI while fostering an inclusive and enriching learning environment for all students.

Resources for Faculty

Resources for Students

Print Friendly, PDF & Email

3 comments

  1. I think there is sometimes an odd way in which people think that you come quite close to elucidating but not directly identifying and describing. That is this mentality of that the rate is low so it doesn’t matter. If someone were about to be murdered, I don’t think the last thought in their mind would be that the murder rate is low therefore everything’s fine.

    I have out of curiosity tested one or two readily available tools out there on texts I have written which are entirely written by me and without any plagiarism. In most cases the results are reasonably accurate but it has not taken many tries to find one service that falsely reported a work as 100% AI generated plus another that reported plagiarism. At least with the latter it is possible to check the matches to observe that they are completely ridiculous. With the former there is a lack of any mechanism in all cases that I am aware of to properly scrutinise results.

    Using AI to test for AI generated results comes with a number of potential problems that might not be arrested with common methodologies such as testing with multiple tools. I did a test with AI not long back and found that it can be biased in ways that you might not expect. It can be quite nasty. In one text I asked it to classify it as either A or B with the latter being effectively trash content of low value. It was high value content.

    The text included a legitimate criticism in it of government policy which impacted the ability to distribute the work done. With this AI was more likely to assign it to B than A and without the reverse. This would apply to politics in a one sided matter. Any legitimate criticism of academia also causes AI to be more likely to classify a text as plagiarised, fraudulent, spam, sometimes AI generated believe it or not, toxic, dangerous, conspiracy theories, mentally unstable and so on.

    This can be tested if you take a text and create alternate versions that are the same text but change merely the entity being discussed. Take a criticism that applies to two bodies. The UN and then a body that have opposition values to it. With criticism of the UN it is more likely to pass a negative judgement of it compared to an opposing body.

    A similar problem occurs with spam filters. Certain groups are more likely to use it maliciously which then impacts its training. The result is that if you take particularly texts simply on two sides of a particular debate it is not uncommon to sees a distinct difference in the rates of spam identification between the two even if none are spam. This applies to many filters and it is likely to become a real problem with AI detection tools as well which tend to operate as a black box.

    At this point I am not even certain it’s the biggest problem. One of the studies you cite has an enormous flaw in it that apparently passed muster. In some respects that is much worse than plagiarism or using AI generation as a tool. If the objective is to seek accurate results rather than to be evaluated on your own ability to operate independently then that’s a phenomenal failure. It would be better that it were instead AI generated or plagiarised than that it were incorrect. The same concept for if you need a plane to fly. Given the choice of what could go wrong once you’re in it and up in the air you would rather it be plagiarised than not flight worthy.

Comments are closed.