Student AI cheating just got trickier

... and why we're still fighting the wrong battle.

Ever since ChatGPT came on the scene, educators have been worried about its impact on academic integrity.

How do I know? Because whenever I do workshops and presentations with educators, it’s one of the main questions …

… and my “AI vs. Cheating, Plagiarism, and Academic Integrity” session is usually one of my most-attended sessions. (Check out the slides here.)

(Spoiler alert: If you come to that session, it’s not about beating students at the cheating game — but rather how to think differently about the entire “academic integrity” situation.)

Since the beginning, the “AI cheating” game has been a cat-and-mouse game … a game of “whack-a-mole” that is impossible to win.

And even as new tools emerge, it’s still going to be impossible to win.

Here’s a new slide in my aforementioned conference session on AI cheating:

Back and forth. Back and forth.

Accusations. Threats. Technology measures and countermeasures.

Let’s address the latest step in this cat-and-mouse game — humanizers and adversarial prompting. (I share this not as a solution, but to keep you informed on what people are talking about so you’re prepared.)

Then, let’s talk about some real solutions.

(This post published originally in AI for Admins, my FREE weekly email newsletter. Subscribe here.)

😼 🐭 Text humanizers

In the natural progression of the cat-and-mouse game of “AI cheating,” the most recent step makes logical sense …

If teachers are going to use AI detectors (which are heinous and inaccurate … more on that in a moment), then students who want to use AI to avoid classwork should try to beat the AI detectors.

That’s where text humanizers come in.

They take AI-created text and make it sound less like AI — and less likely to be detected by AI detectors.

Some examples (without links, because I don’t want to give these sites extra web traffic or eyeballs):

• BypassAI
• Humbot
• Undetectable AI
• Humanizer . org

And the list goes on and on and on.

😼 🐭 Adversarial prompting

You don’t even need a text humanizer to beat AI detectors, though. Some extra prompting in ChatGPT (or your AI assistant of choice) will do it.

This paper, written by university faculty in Vietnam and Singapore, highlights AI prompting techniques that can avoid detection.

They include:

• Rewriting with intentional errors
• Varying sentence length
• Increasing text complexity
• Downgrading text complexity
• Rewriting as a Non-Native English Speaker (NNES) with IELTS Band Level 6
• Paraphrasing

The paper’s conclusion offers three implications:

• The major implication of these findings for educators and administrators is that the use of AI text detectors should not be implemented uncritically; users must consider the impacts and limitations of AI text detection technologies before using them for assessment or educational practices.
• Secondly, the results of such technologies should not be used for punitive actions or in accusations against students without a high degree of certainty, and those in the position to evaluate the results from such technologies need to consider the ease with which detectors can be evaded, and their potential to inequitably impact certain student populations.
• Finally, the findings imply that educators must radically reconsider their assessment structures and practices in light of new technology, given that current efforts to detect GenAI content are unlikely to be successful.

😼 🐭 AI detectors

We’ve touched on this a LOT in this newsletter, but in case you’ve missed it …

AI text detectors are terrible at their jobs. If you go to their websites, they won’t tell you that. They’ll boast some percentage accuracy rate that’s not true (or realistic).

Published academic research has backed this up …

• “… the tools exhibited inconsistencies, producing false positives and uncertain classifications.”
• “… available detection tools are neither accurate nor reliable …”
• “These error rates raise doubts about relying on AI detectors to enforce strict policies around AI text generation …”
• AI detectors are biased against non-native English speakers.

🤦🏻♂️ We’re on a race to nowhere

Have you noticed something about this back-and-forth battle about AI detection?

The longer it goes, the less it focuses less and less on the learning.

And, to some extent, it’s a monster of our own creation.

When we agree to play this game against our students — and try to beat them at it — we only encourage them to take further steps in their own game.

It’s a no-win situation.

The real problem is that we are actually using these AI detectors in the first place.

Lots of teachers are being lured in by the siren’s song of “everything can go back to the way it was before” and “you won’t have to change how you teach.” They’re buying a bill of goods that doesn’t produce results — accurate results, that is.

🤷♂️ So … what do we do?

We’re in a messy time of transition. Much like we did with previous innovative disruptions to the classroom (calculators, encyclopedias, search engines, 1:1 computing, YouTube, etc.), we need to evolve.

That doesn’t mean that we have to throw out all of our traditional classwork and create brand new assignments right away.

We’ll have to ask ourselves a couple crucial questions …

1. Can we save our original classwork / assignments?

I think the answer is — maybe, maybe not. Here are some things to consider:

• When AI cheapens the final product, focus on the process.
• Collaborative work can encourage original human thought.
• Incorporate creative final products to demonstrate learning.
• Have students reflect on their work (how they’d do it different, etc.).
• Consider an element of AI/student collaboration to existing classwork.
• Scaffold writing activities -- or chunk them into smaller assignments.

2. How can we reimagine our classwork?

This is the big question lots of us are trying to figure out right now.

Here are a few things we’ll really need to think about …

• Thinking and skill development: Can AI support these instead of removing them? How can AI augment rather than replace?
• Goals and objectives: Instead of saying “this is how we’ve taught and learned before,” we’ll have to go back to the basics. What are we trying to accomplish?
• Experimentation: Let’s try incorporating AI into the learning process in some ways and afterward, ask ourselves: “What worked? What didn’t? How could this be different next time?”
• Focus on the future: With any changes — or decisions to stay the same — we must ask: How does this prepare students for the future they will face?

🤔 What do you think?

How do we get this right?

What are some positive steps you’re seeing in your realm of education?

What else do we need to consider?

Let me know what you think in a comment — I’d love to engage in the conversation!