Can AI detectors reliably identify AI-generated student work?

No. AI detectors give different scores each time you submit the same text. They measure similarity, not authorship, and can't distinguish between AI use, grammar checkers, and human writing. At the moment, they seem particularly bad at identifying machine-paraphrased text as written by AI.

What happens if a student is falsely accused by an AI detector?

Disciplinary actions can mean loss of scholarships, visas, internships, and future admissions. Some universities now prohibit using detectors as sole evidence for cheating due to unreliability.

Is using AI for grammar checking considered academic dishonesty?

It depends on the institution's policy. Traditional tools like spell check and thesauruses have always been acceptable. AI assistance exists on a spectrum, where clear policies should define where the line is drawn.

What should universities do instead of using AI detectors?

Define clear, fair AI policies and communicate them effectively. Use human judgment, assess student engagement and learning outcomes, and treat detectors as supplementary evidence.

AI Detectors Can't Determine Student Cheating (And Schools Keep Using Them Anyway)

AI is a tool that can be used for good or for harm. It's not "coming for us." The people wielding it are responsible for what happens based on its output. There are people who benefit from you and me focusing blame on AI itself rather than them as humans using a tool that has led to deaths, job loss, unauthorized hacking, and data destruction.

AI is inserted into everything

The rest of us have to figure out how to deal with AI use when it intrudes in our lives. For example, schools have to respond in some fashion to students using AI in a myriad of ways as it's shoved into phones, computers, games, social media, transportation, ads, movies, tv, and shopping. And refrigerators. There are AI tutors and AI writing helpers that students and instructors alike use. I see AI generated images all over my son's school for events and promotions. Meanwhile, that same school has many talented students who could've created at least some of them.

Schools and teachers using AI generated text and images when students are told not to isn't the only contradiction. Using AI to detect AI, instead of your own human judgement is a contradiction. A goal that high schools and universities share is student learning. The idea is we want them to work for it and do their own writing and creating. If you want to actually learn something, it takes engagement and participation. Writing is one really good way to learn.

Academia's contradiction

There's another interesting contradiction in academia about using AI detector tools at the university level. In academia, you want reproducible, verifiable results. The stats are based on probability and correlation, but you need someone else to be able to repeat what you found and find the same thing. It requires a human being who has trained on the literature and findings of past scholars. With the AI detectors, you're relying on a tool that's incapable of producing the same answer twice. Take out the human factor and you've got nothing useful.

Even when we generate random numbers in research, we use what's called a "seed" so you can be more likely to generate the same random numbers as someone else. LLMs generate language that's been coded into data structures. They're numbers at the core. You will get a different result each time you prompt it.

In the case of AI detectors used on papers and homework in schools and colleges, you get a probability score of whether text was generated by AI or a human based on a black box that will provide different scores each time you ask. The AI-generated "fingerprint" is not a simple formula, it's a correlation that academia should treat with the same caution they do the findings of their own statistical formulas.

When is it ok and not ok?

Universities and schools are reaching for the AI tools to try to combat something real that they're wrestling with. Some students do abuse AI in their learning process. I've seen it. However, universities' response seems to be to allow an accusation based on an AI tool, relying on unreliable detectors. I'd lump high schools in this too. The punishment can affect someone's future. This shifts accountability from clear policy communication and enforcement to speculation based on a black box.

You might think of a student fully generating an essay without any thought or even glancing at it, and that does happen. But there are students who have to go through a punishing accusation and haven't used AI. Others likely fall into a grey area. If the accusation holds, the resulting disciplinary action can mean loss of a scholarship, visa, or a mark that bars admission to any school. It can cost money in the short term as well as negatively affect the long term.

The grey area involves answering questions like, is it ok to use AI as your spell check? LLMs can point out typos that me and my "dumb" spell checker missed. From looking online, people say you can "accept grammar changes" and reject changing phrasing for some AI tools. Where does that fit in? The same type of AI can also point out real words that the spell checker thinks are fine. Did the dog say "rough" or "ruff?"

How much wording change is academic dishonesty? When I was in school and would use lists of "better words to use" or "precise words" or rephrase based on my Bedford Handbook that teaches style, was that just as bad as AI changing a student's wording for the exact same thing? Do schools frown on students for using AI for grammar checks or spell check? That's a tough determination when someone uses writing help or common phrasing from somewhere else.

AI companies say don't use it as the sole determination

The companies who create these AI detectors say that they shouldn't be used as the sole determination of whether text was AI-generated, human-generated, or a mix. They can't even tell you if something is plagiarism. They can tell you a score of how similar text is to something else (which can include the student's same, prior work). The instructors who say, "you need to understand the underlying math" need to understand the underlying mechanisms in these AI tools. They're not a guaranteed result that's clearly correct.

I do want to note that some universities have caught on and their policies state that they don't allow these tools as sole evidence for cheating.

AI is a tool that should be used with other evidence

It's up to universities, their programs, and their professors to define a clear, fair AI policy that makes sense in our world. It should be communicated effectively to students, staff, and faculty. But relying solely on AI to detect AI is weak evidence.

Human responsibility

Individual humans are using the AI tools to determine if "academic dishonesty" has taken place. This may be with or without policy guidance from their institution or peers, like other faculty or teachers. The point is to question when you hear "AI is coming for us."

People are responsible for the consequences of acting on information supplied by tools like AI.

The AI's programming, training data, and guardrails are set by humans and sourced from our work. Humans deploy AI models. Humans provide tasks to the AI and train the AI on tasks completed by humans. If someone makes a mistake because they used a tool incorrectly, it's on them for using something they don't understand. It's human responsibility all the way through.

References:

P.S. I can tell you this is definitely not AI generated text. It took days to write something this short.