The Impact of AI's Expanded Horizons in Education
Sora, Gemini and Education
Summary of main points:
Sora's Potential in Education:
Offers the ability to create high-quality visual content easily but struggles with producing precise imagery required for certain educational content.
Limitations and Concerns:
High-quality visuals could mislead about the overall quality of educational content as these become easier to produce.
Critical evaluation of content by educators becomes more important as the ability to produce high-quality visuals becomes widespread.
Google Gemini's Impact:
Its large context window allows for richer summarization of complex academic material, potentially revolutionizing research and literature review processes.
Enhances the creation of educational materials (e.g., assessments, study guides) by incorporating extensive course content and reading materials.
Promises more personalized learning experiences by tailoring responses based on a comprehensive understanding of a student's background and work.
General Considerations:
It is necessary to plan for the integration of AI technologies like Sora and Gemini into educational practices, given the rapid pace of development in the AI landscape.
The AI landscape changes every week, with new products, ideas and applications being developed. Two developments with significant potential for education recently occurred, the promotion of Sora from OpenAI, which generates realistic video from simple text prompts, and the release of the updated Google Gemini, which amongst other things has a very large context window.
What do these mean for education? As with other developments in AI we do not truly know yet, and there is much experimentation to do to find out, however let me outline what I think could be the effect. I’ll start with AI video, where I think the effect is smaller.
Sora
Firstly, if you haven’t had a look at some of the material yet, visit the OpenAI site and take a look. As with any promo, this material is picked to be some of the best, and will disguise some mistakes and frustrations that will likely be a feature of using this system, as it is with other AI products. Nonetheless, the results are impressive.
How do we currently use video as educators? Outside of courses specifically teaching media technology, video production, art and other directly related fields, the main use of video is as an accompaniment to a lecture.
Something like this for example, has some nice simple visuals with a voiceover explaining the economic topic of monopoly.
It will now be much easier to produce this sort of imagery to accompany an instructional video.
Careful analysis of videos produced by Sora, as with many static images produced by AI, shows errors. This may be a temporary state of affairs, but for now AI generated videos and images are unlikely to be helpful when you need very precise imagery. Contrast the video above to the one below.
The latter video uses more precise imagery of demand and supply diagrams, these have to be exactly correct, and for now AI will struggle with this detail1.
AI may also be helpful in improving accessibility of these resources. Making it easier to produce translations, subtitles or to change aspects of videos to fit to different cultural contexts or make them more appropriate to different age groups.
As an aside there is now the possibility to use AI voice-overs, so if you do want to produce instructional videos but think your own voice is insufficient to the task, this is another option that can enhance your educational content.
What might this easily available video allow that we cannot do now? If the technology develops further, then connecting it to virtual or augmented reality may offer easily adaptable immersive experiences. For now creating a virtual environment to simulate a business, or trading floor, or workshop floor will be very expensive and time-consuming, soon it may be easy.
In summary I don’t think for now AI video is likely to have much of an effect on education. It will make it easier to produce high quality visuals to give a general background effect, but will be of limited use in the production of very specific and accurate video. This may mean that some educators who do not have access to the resources to produce visuals of the type used in these examples can now access this market, meaning more high quality content for students to use, though also some more difficulties in ascertaining what is good and what isn’t. At the moment the presence of high quality visuals may act as a signal that the rest of the content is also high quality. This may not always be true, but soon this signal will be totally meaningless as everyone can produce these high quality videos relatively easily.
Our job as curators of this sort of content may therefore be strengthened, needing to stay on top of the what is available through YouTube and similar platforms and ensuring our students critically evaluate the content that is available.
Gemini
There has been a lot of discussion of Gemini, in terms of its enhanced performance and controversy about ideological bias. Here I want to focus on the context window.
First an explanation, from ChatGPT4:
As nicely explained (ignoring the slightly creepy use of “me”) the output of an LLM is limited by the size of a context window. For example I might want to summarise some academic ideas by a specific author. In a smaller context window I might be able to include one or two papers or a small book, in a larger context window I could summarise many books and articles, picking out themes across them to get a richer summary. Here is Ethan Mollick doing just this with his own work. Another interesting example is the use for translation, in this example using one book to translate a language.
Summarising material is already one of the most useful features of current LLMs, as witnessed by the explosion of research apps that summarise papers, find synergies between them etc. The larger context window can scale up this activity and improve the quality, making research and reviewing of literature an easier task.
Here is one of Google’s demos showing off the capabilities:
From the perspective of the teacher this could also have big implications. I have been using LLMs to help me write tests, assessment questions and model answers. For now I have mainly used the raw output of an LLM and its own training data, as well as a small amount of course material. Now I could potentially input all the material from my course to generate more complete assessment details (or at least the draft of them).
Similarly I could input a range of materials from the reading list of the course and output material to use in my handbook, on my slides etc.
Essentially the whole process of generating documents for a taught course is now going to change dramatically. You will not likely generate finished products that you can use straight away, but drafts of reading lists, study guides, assessments, presentation slides and other documents you will use will be easy to generate based on a repository of the main information you want to use in the course.
In a future post I will examine these capabilities via my own material to demonstrate what can be done.
Returning to the student perspective and considering what new possibilities emerge, more personalised learning now seems more attainable. The larger context window could include more information on past assessment attempts of the student for example, including feedback they received, drafts of other work etc. With that information an LLM could tailor its output more specifically to the needs of that student, their ability, the things they find more difficult, the things they find more attainable and so forth.
These personalised responses are available now with some suitable prompting but with the potential to include a student’s entire background and corpus of work, the level of personalised detail will be enhanced considerably.
It is worth making the point that whilst I am positive about the impact that AI can have on education, we are still experimenting, and every new development brings risks as well as the potential for rewards. As highlighted in the first section, could a flood of cheaply made video make it hard to discern what is truly useful content? Could reliance on AI generated university course material lead to a bland homogenisation that strips away some of the uniqueness of an individual academics courses? These are real dangers and something we can adapt to in developing both our own and our student’s skills in evaluating material, and ensuring that when we do use AI for course creation we do so in a way that retains the features we want as part of our programmes and our own knowledge.
Neither of these products are widely available to everyone currently, but given the pace of developments it seems unlikely that it will be too long before these aspects of the AI landscape become normal. We should plan accordingly.
The image at the end of this article is a good example, I am happy to use it to give a general ‘vibe’ but if I worried about the details there would clearly be a problem (just look at the words).