Welcome to Teacher’s AIed: the newsletter about AI in the K12 Classroom.

How AI will affect K12 Classrooms is complex. Each week, we curate knowledge for educators about the strengths, weaknesses, opportunities, and threats of AI and K12 education.

In this edition, we finish our series focused on addressing the "student AI-plagiarism problem.” This week focuses on dissuading AI plagiarism on assessments where students can use AI.

We have been exploring in this series one element of the education process that AI will profoundly impact: assessments. We are facing a problem of authenticity - how do we know that what a student turns in accurately represents their knowledge - not ChatGPT’s ability to piece together a response?

Our operating framework is that the more inconvenient it is for students to use AI to complete an assignment, the less likely they will cheat. Therefore, we’ve addressed the power of a positive classroom culture in disincentivizing cheating. Last week, we considered simply taking the technology off the table. 

But, if you’re reading this, you’d likely agree that neither strategy is sufficient on its own or best for students in the long run.

In this post, we will share our thoughts on designing (more) AI-immune assessments by changing the assessment format. 

Accessibility: The Crucial Component

It is important to note that the ease of plagiarism for any particular assessment format correlates to the accessibility of AI tools. For example, ChatGPT has turned education on its head over the past year. However, AI tools, even Large Language Models (LLMs) such as ChatGPT and Bard, have been around for years (LLMs were first proposed in the 1960s!). The seismic shift of ChatGPT was not necessarily some ground-breaking new concept. Instead, what changed the world was its accessible, simple interface. 

LLMs have all but doomed traditional typed assignments to the realm of constant suspicion because of the prevalence and accessibility of LLM tools. Here’s a short list of other assessment formats used in K12 contexts, along with a summary of the state of the AI tools necessary for plagiarising on this type of assessment:

Visual Representations of Knowledge

What would it take for AI to create a pictorial model of the carbon cycle? Students already have LLMs at their disposal to generate the textual content and organize it appropriately. All that remains is the ability to organize that content graphically. This visual organization is quite tricky for Generative AI. I prompted the Bing Image Creator with, “Create a visual representation of the carbon cycle.” The following image is what I received:

Notice the distorted text, the arrows pointing in strange directions, and a collage of random objects strewn about (lots of plants, no animals, and half of a road bike?). Obviously, this is not something a student could turn in immediately. Even if a student were to replicate this on paper, they would produce a non-sensical model. That said, I don’t think it will take long for the technology to have more checks to prevent these oddities when creating this kind of product.

Video Presentations

President Biden’s staff used AI to illustrate that bad actors could use a snippet of an audio sample to create a convincing fake. The President’s response to seeing this video was, “When the hell did I say that?” as AI had made the entire video. We have already entered an age where videos AI can convincingly create convincing fake videos.

To my knowledge, the technology to create personalized videos has either yet to become easily accessible, or the easily accessible options are not sufficiently robust to be convincing. Yet, with the exponential pace of AI’s expansion, I don’t think it will be too long before this is more widely available.

Performance Assessments (Oral Exams, Practical Exams)

Perhaps the most challenging exam I’ve ever taken was a computer science practical exam. I had a time slot with 2 of the classes’ teaching assistants, a dry-erase board and marker, and 15 minutes to prove I knew the algorithm to balance an AVL tree.

Indeed, generative AI could have explained the process more clearly and correctly than I did, but if my robot twin had walked into the room, the teaching assistant would have suspected something fishy. The hardest assessments to fake are performance assessments (oral and practical exams), where students must demonstrate their knowledge in person.

Teacher’s AIed’s Recommendations

Where does this analysis lead us? We have a few recommendations for teachers.

First and foremost, avoid simple typed written assessment products. Teach students to use incredible GenAI tools but design assessments so that the technology cannot easily create the final product they turn in. Education has faced this similar challenge before when calculators were easily accessible during the third industrial revolution. At first, there was concern that calculators would lead to poor math outcomes, but now, math teachers teach students how to use calculators. Nevertheless, it would be ridiculous for math teachers to give their classes a multiplication fact quiz where students could use a calculator. As with GenAI, the product of the technology cannot directly match the product of the assessment; otherwise, students have nothing to contribute.

Secondly, replace easy to plagiarize assignments with more complicated to plagiarize assignments. Instead of writing a paragraph describing the carbon cycle, ask students to record a video, give a presentation, or create a graphic explaining the same concept. Students could even be allowed to use AI to prepare for this assessment, but as long as the AI tool cannot fully complete the assessment, some degree of cognitive lift remains on the student.

Lastly, consider pairing simple, typed, written assessment products with other components that will disincentivize cheating. The more intertwined these components are, the more authentic learning will occur. One common strategy that might be helpful here is conferencing. If students are doing a research project on the carbon cycle, have students periodically check in with you to discuss the content, brainstorm ways to improve, and set goals. If students know they will have to answer questions about their report with their teacher regularly, they will be less incentivized to cheat on writing the paper. They undoubtedly still could, but with the frequent check-ins, a teacher could most easily identify if a student completed the report using Generative AI.

In the world of education and AI, the battle against AI plagiarism in K12 classrooms has been nothing short of an adventure. As we've explored in this series, the rise of accessible AI tools has introduced both challenges and opportunities. We've witnessed how AI can easily create written content, but we've also seen its struggles in crafting visual representations of knowledge. The debate about personalized videos is ongoing, and oral and practical exams remain the toughest nuts to crack for AI.

So, where does this journey lead us?

Our advice to teachers is clear: create assessments that challenge students.

Remember, while AI transforms education, the human touch is invaluable, and by making classrooms AI-immune, we ensure that learning remains authentic and essential.