Despite their increasing use, there's limited evidence that AI-generated avatars significantly improve learning outcomes.

A study by Sung and Mayer (2012) suggests that any graphic in a learning experience will fall into one of these three categories: Instructive images: These visuals directly support learning and facilitate essential cognitive processing of core concepts. For example, a diagram illustrating Porter’s Five Forces can help students better understand this business strategy framework. Decorative images: These graphics enhance aesthetics but don’t influence learning. For example, an image of a business handshake can be visually appealing but won’t support or obstruct students’ understanding of negotiation strategies. Distracting images: Sung and Mayer call this category “seductive” images. While these visuals may relate to the topic, they impede learning because they require extraneous cognitive processing. As an example, consider a complex organizational chart of a full corporation in a lesson on team leadership. The image connects broadly to the lesson but also highlights a lot of irrelevant details, distracting students from the key concepts.

While LLMs generate text one word at a time, always moving forward, diffusion models start with random static and transform the entire image simultaneously through dozens of steps. It is like the difference between writing a story sentence by sentence versus starting with a marble block and gradually sculpting it into a statue, every part of the image is being refined at once, not built up sequentially.

But what makes diffusion models interesting is not their increasing ability to make photorealistic images, but rather the fact that they can create images in various styles.

Unlike diffusion models that transform noise into images, multimodal generation lets Large Language Models directly create images by adding tiny patches of color one after another, just as they add words one after another. This gives AIs deep control over the images it creates.

Indeed, there is evidence indicating that feedback delivered immediately after an item may optimise error correction (at least in some settings),43 whereas delayed feedback may enhance retention and transfer for correctly answered questions by providing an appropriately spaced second learning opportunity.44 Again, findings are mixed, and others have found delayed feedback to augment learning equally for both correctly and incorrectly answered questions.41 Interestingly, there may also be a greater advantage to delayed feedback when an initial question is more difficult

Contrary to our hypotheses, the feedback timing effect was non-significant—there was no discernible difference between feedback delivered immediately versus delayed feedback.

Logically if we think about feedback as correcting errors, then it makes sense that we would want pretty immediate feedback. But if we think about feedback as another presentation of the information, then a space ought to improve learning.