This post was contributed by Amy Hofer, Statewide Open Education Program Director, and Veronica Vold, Open Education Instructional Designer, Open Oregon Educational Resources. Thank you to Jeff Gallant, Program Director, Affordable Learning Georgia, and Stephen Krueger, Affordable Course Content Librarian, University of Kentucky, for their feedback and advice.
This post explains why the team at Open Oregon Educational Resources thinks that generative AI (GenAI) is not ready for the biggest open education projects that we’re working on right now. We’re not offering policy here, but rather guidance for authors along with an explanation of why we’re taking a wait-and-see approach instead of continuing to experiment with the new tools. By sharing this post while we’re still mid-stream with understanding this technology, we hope to provide a reference point for colleagues who would also like to slow down.
Context: our program received two grants to develop openly-licensed course materials with an equity lens in three disciplines. We’re developing 12 open textbooks that include self-check questions at the end of each chapter inviting students to test their understanding in a low-stakes format. Our project uses two accessible H5P question types: multiple choice and true/false. Questions always include automatic answer feedback to guide students in further study and review.
Two obvious uses of GenAI would be to draft chapters, or sections of chapters; and to draft H5P questions. We tried both, and it didn’t work for us. Our takeaway is that it’s not a good use of our program resources to put any more time and money into testing these tools right now. So the guidance below is what we are going to tell our authors until the tools improve.
Can Generative AI Write Your Chapter For You?
This is an emergent topic and our guidance for authors will likely change over time, but for now the answer is likely no – or, more precisely, to proceed with caution and prepare to spend extra time on your task.
If you already enjoy using GenAI tools as part of your writing process, we won’t stop you from continuing. Keep in mind, though, that we are unable to provide training or support for instructors who use GenAI on our projects. We also need you to take into account the three caveats below.
First, the outputs of GenAI tools can contain incorrect information due to a phenomenon known as “hallucination,” where a GenAI neural network’s response to a prompt includes what it deems to be the most correct-sounding response, without the response being factually correct. You must carefully fact-check anything that a GenAI tool produces.
Second, we are not given access to the materials used to train or validate the many GenAI tools you might use, and it is very likely that the corpus or validation methods were not chosen with an equity lens. This means that anything produced with GenAI tools will likely reflect historical biases and you will have a big revision job to do to keep your materials aligned with our project’s equity goals. Please consider the Seven Forms of Bias in Instructional Materials [Website] (invisibility, stereotyping, imbalance, unreality, fragmentation, linguistic, and cosmetic bias) and make sure that your tech tool is not introducing any bias into your work.
Third, you must understand the copyright and licensing issues that may affect your work. You’ll need to attribute revised GenAI generated content and keep track of the role of AI in the work so that you can clarify what tools were used, which sections were AI-generated, different copyright/licensing of those portions, etc. You will also need to review the licensing terms of the specific AI tool you used to ensure legal use of the output.
We suspect that after you’ve crafted the perfect prompt to get what you want, and gone through these steps to get it ready to share widely in an openly licensed textbook, you might have outsourced the creative parts to a fancy algorithm and left your human self with all the tedious tasks!
All this said, here are three resources that we recommend for anyone interested in experimenting with GenAI:
- Affordable Learning Georgia’s Guidelines for Using Generative AI Tools in Open Educational Resources [Website] explains copyright and ethical considerations for OER authors.
- University of Kentucky’s Generative AI Policy for OER [Website] provides concise guidelines for authors.
- Open Oregon Educational Resources’ Attribution Style Guide and Examples [Google Doc] provides examples of how to attribute revised GenAI generated content.
Can Generative AI Write Your H5P Interactives For You?
In a word, no.
Our team experimented with two GenAI tools, Nolej and ChatGPT, to generate H5P questions. We determined that it cost our program far more time and money than we would have invested in simply writing questions from scratch. We estimate that generating H5P interactives with GenAI approximately doubled our time investment when we tried it out. We can’t commit program resources to support its use since it’s demonstrably inefficient.
The overarching issue we found was that GenAI didn’t align our H5P interactives with the chapter’s learning objectives, leaving us with extensive revision work even after trying multiple prompts. The revisions were tedious to do, and because we needed to generate so many questions for our projects, it was easy to get tired and overlook places where the errors or missteps were subtle.
Here are three kinds of specific issues that we needed to correct after using two popular GenAI authoring platforms to generate questions:
- Nolej and ChatGPT generated questions that focused on easily recognizable facts, like dates and definitions, rather than questions that require the application of new concepts or critical thinking. This might work for some course settings, but it’s not what we’re looking for in the self-check questions for our projects.
- Nolej and ChatGPT marked answer options incorrectly – some answers that are actually true were coded as false. This would be confusing and frustrating for students.
- Nolej didn’t generate answer feedback, and ChatGPT didn’t generate meaningful answer feedback from an equity perspective, even when explicitly prompted to do so (ex: “suffix the correct answer with ::: and an equity-minded explanation for why this was the correct answer”). Instead, it repeated itself or introduced content unrelated to the original question. Sometimes the feedback lacked important nuance, and sometimes it missed the point of the question entirely.
ChatGPT seemed to perform less reliably the more the content related to equity concepts. In one example, generated answer feedback overlooked the concept of intersectionality as it relates to feminism, even though that is a stated learning outcome of the chapter. This is consistent with the findings shared by open educator Maha Bali in the webinar Actually Scary Things About Artificial Intelligence in Open Education [Streaming Video]. Bali points out that even GenAI tools like NotebookLM that are designed to extract and paraphrase material from provided content tend to reinforce systemic bias. Because GenAI is trained on biased material, it will miss equity concepts in ways that may be subtle or overt.
Further, as we saw in Florida Scours College Textbooks, Looking for Antisemitism [Website], questions are often used as at-a-glance indicators of the textbook quality as a whole. If our H5P interactives fail to reflect the nuance and complexity of the books themselves, we not only risk reinforcing the bias that we are working against. We also may lose our audience.
Conclusion, For Now
We’ve concluded, for now, that GenAI isn’t the right tool for our open curriculum development projects. The tools are changing fast and we’re open to changing our minds in the future.
Speaking of the future, we want to end on a downer: the environmental costs of AI. Of course many of the tech tools we use regularly – and recommend – have an environmental impact. However, GenAI’s impact when compared to the crummy returns we got just didn’t balance out. For more on this important consideration, we recommend Marco Seiferle-Valencia’s talk, AI Spring: Generative AI, Open Education, and Climate Catastrophe [Streaming Video].
Funding
Our grants drew from Governor’s Emergency Education Relief funding and the Fund for the Improvement of Postsecondary Education (FIPSE) in the U.S. Department of Education (eighty percent of the total cost of the program is funded by FIPSE, with the remaining twenty percent representing in-kind personnel costs funded by Open Oregon Educational Resources).
The contents of this post were developed under a grant from the Fund for the Improvement of Postsecondary Education, (FIPSE), U.S. Department of Education. However, those contents do not necessarily represent the policy of the Department of Education, and you should not assume endorsement by the Federal Government.
