Detecting AI Is Broken—So Let’s Build Smarter Assignments Instead

The explosion of generative AI tools like ChatGPT is shaking up classrooms around the world—and not always in a good way. While students now have AI at their fingertips to draft essays, summarize notes, and brainstorm arguments, educators are left scratching their heads: how do you keep learning real when machines can ace your assignments?

The go-to response for many institutions has been to combat AI use with AI detectors. But here’s the uncomfortable truth: these tools don’t really work. They misclassify human work as AI and vice versa. Worse, with a little paraphrasing or translation, almost anyone can easily sneak past them.

So what’s the fix? According to a new study by Muhammad Sajjad Akbar, we’ve been asking the wrong question. Instead of obsessing over detection, we should be focusing on design. Specifically, designing AI-resilient assessments that actually challenge students to think critically, solve problems creatively, and produce work that generative AI can’t replicate.

Let’s break down the key ideas from this groundbreaking research—and what it means for students, educators, and the future of learning.

Why AI and Education Are at a Crossroads

Generative AI isn’t just a tech trend—it’s mashed its way into education like a wrecking ball. Tools like ChatGPT can spin up essays, explain concepts, and even mimic student voices in a matter of seconds. For students, that can feel like an academic shortcut. For professors, it’s a potential academic integrity time bomb.

But beyond questions of cheating, there’s something more serious at stake: Are we trading away deeper learning for convenience?

Educators are noticing that when students offload too much thinking to AI, they’re missing out on developing key cognitive skills like problem solving, analysis, and creativity. In modern education, especially at the postgraduate level, these higher-order thinking skills are essential. They help graduates adapt, think independently, and contribute meaningfully in any industry.

Current AI Detection Tools: Not the Silver Bullet

Most colleges and universities have turned to AI detectors in a desperate attempt to catch machine-written content. But this study found that most of these tools are, frankly, not very good:

• Poor accuracy: Even popular commercial tools like Turnitin and GPTZero often misidentify both AI and human-written text. Some had false positive rates as high as 50%.

• Easy to trick: Students can confuse detectors just by paraphrasing, translating, or modifying AI-generated content.

• Lack of transparency: Many tools give vague reports (“This text is 70% likely written by AI”) without solid evidence.

The result? Educators may wrongly accuse students or let AI-authored content slip through—not because of laziness or malice, but because the tools they’re using simply aren’t up to the job.

Introducing a Better Approach: AI-Resilient Assignment Design

Instead of trying to catch AI after it’s been used, the researchers suggest beating AI at its own game—by creating assessments that AI tools struggle to complete effectively.

To do this, they built a smart, web-based tool that helps educators analyze how “AI-solvable” their assignments are. It measures whether a task falls into the easy-for-AI category (think: summarizing, defining terms) or if it requires complex thinking like analyzing data, evaluating arguments, or creating something original.

This system works by combining:

• Bloom’s Taxonomy – a classic model that categorizes learning objectives from basic to advanced.

• Natural Language Processing (NLP) – using GPT-3.5 Turbo, BERT semantic analysis, and TF-IDF to gauge how complex and nuanced the assignment questions really are.

• Python automation – to quickly scan PDFs of assignments and return an “AI-solvability score” with real-time feedback.

The lower the score, the harder it is for AI to generate a satisfactory answer. And that’s exactly what we want!

What the Research Found: Most Assignments Are AI-Too-Friendly

When the team tested 50 real-world assignments from various computer science courses, the results were sobering:

• 22 assignments had AI-solvability scores between 65–74% — meaning AI could probably handle them with ease.

• Only 4 assignments scored below 50%, indicating true resistance to AI-generated responses.

These high-solvability assignments were mostly definitions, factual explanations, or basic applications—the kinds of things ChatGPT gobbles up. On the other hand, the low-solvability ones required designing unique solutions, interpreting specific scenarios, and thinking critically, creatively, and contextually.

Put simply: the more thinking an assignment requires, the harder it is for AI to fake it.

How Students Use AI: Help or Hindrance?

The team also surveyed postgraduate students at the University of Sydney to understand how AI is actually showing up in their learning. The results paint a mixed picture:

• 70% of students said AI tools helped them understand material better. (Nice!)

• But over 60% admitted to making only minimal edits to the AI text before submitting. (Uh-oh.)

• 75% believed the AI answers were more accurate than theirs.

• 60% said AI tools actually made them less creative.

This suggests a potential over-reliance on AI—not just for support, but as a crutch. While the tools can boost confidence and clarify tough concepts, they can also kill creativity and reduce critical engagement when used passively.

The Real Impact: Smarter Assessment Design Can Fix This

By using the AI-solvability tool before finalizing an assignment design, educators can ask themselves key questions:

• Does this task really challenge students to think independently?
• Can ChatGPT realistically generate a viable response here?
• Am I valuing originality and synthesis over regurgitation?

Armed with this insight, instructors can rethink their prompts—pulling back on the easy-to-automate stuff and leaning into unique, personalized, and cognitively demanding formats. Think open-ended projects, scenario-based analysis, or unfamiliar case studies—things AI can’t fake.

Even better, the tool gives clear feedback so assignments can be revised before they’re given out. That means assessments not only become fairer, but more effective at encouraging actual learning.

Real-World Applications: Where This Can Make a Difference

Here’s how this AI-aware assignment design approach can be practically used:

• Instructors can automatically check if their assignment is too easy for ChatGPT and tweak it on the spot.

• Education departments can integrate the tool into Learning Management Systems to flag at-risk assessments early.

• Faculties can train professors to apply Bloom’s Taxonomy more deliberately when crafting course tasks.

• Students benefit from assignments that reward insight and creativity rather than formulaic responses, making room for real skill-building.

Key Takeaways

✅ AI detectors are unreliable. They produce too many false positives/negatives and are easy to circumvent.

✅ Stop trying to catch AI—start designing around it. Move from detection to prevention by crafting better, higher-order assessments.

✅ Use Bloom’s Taxonomy as your secret weapon. The more your task pushes into analysis, evaluation, and creation, the harder it is for AI to fake it.

✅ Students should be taught AI literacy. Knowing how to use AI tools responsibly—as support, not a substitute—builds better thinkers.

✅ Proactive assessment design helps everyone. Educators reduce cheating, and students engage more deeply with their learning.

✅ This is scalable and ready now. A Python-powered tool exists today that gives measurable, actionable feedback on how AI-susceptible your assignments really are.

In a world where machines can crank out essays in seconds, it’s tempting to react with suspicion, lockdowns, or blame. But this research shows us a smarter path forward: design assessments that AI can’t do—but students must.

Let’s raise the bar not just for what AI can do, but for what students need to learn.

Want to rethink your assignments or understand how AI-solvable your tasks are? Keep this question handy next time you write a course prompt:

“Could ChatGPT answer this well without understanding the course or the context?”

If the answer is “yeah, probably”… it might be time for a redesign.

Author Note: Based on research by Muhammad Sajjad Akbar, Beyond Detection: Designing AI-Resilient Assessments with Automated Feedback Tool to Foster Critical Thinking.

If you are looking to improve your prompting skills and haven’t already, check out our free Advanced Prompt Engineering course.

This blog post is based on the research article “Beyond Detection: Designing AI-Resilient Assessments with Automated Feedback Tool to Foster Critical Thinking” by Authors: Muhammad Sajjad Akbar. You can find the original article here.