AI-Generated Virtual Instructors Based on Liked or Admired People Can Improve Motivation and Foster Positive Emotions for Learning

In recent years, machine learning (ML) algorithms have become increasingly adept at generating realistic-looking images and videos of people. This technology is being used for generating "deepfakes'' or "AI-generated characters'', which are synthetic images or videos where faces or bodies are digitally altered in ways that make them difficult to distinguish from real images or video content. While deepfakes have recently been used mostly for nefarious purposes, such as creating fake news stories or spreading misinformation, we believe they have the potential to be used for good.

One potential use case for AI-generated characters is in the field of education.  A shift towards remote learning during the COVID-19 pandemic has burdened teachers to transform their content, and has challenged students to keep focused and motivated. AI-generated characters present an opportunity for educational content to be personalized in order to foster interest and engagement. They also hold the potential to assist real-life instructors and perhaps improve access to education. 

Prior research suggests that instructors' identities and  student-teacher relationships can impact students’ attitudes, motivation  and even their academic outcomes . For example, one study found that learning from someone from the same race or gender can increase engagement and learning motivation. Another study found that fictional characters can be used to foster stronger motivation and growth mindsets in learning. These findings suggest that the way a student relates to the instructor can have a significant impact on a student's attitude and motivation levels, even if all other variables are constant. Motivation has in several studies been associated with better overall learning outcomes.

Given this, it is intriguing to consider how AI-generated instructors could be used to enhance motivation in online learning.  In this paper, we investigate the effects of learning from videos of AI-generated instructors that resemble characters that people like and admire. Using an open-source platform for generating synthetic characters, we conducted an extensive study with 134 participants to explore the effects of a personalized virtual instructor on students’ learning outcomes, as outlined in the following research questions:

We believe AI-generated characters can be used to create compelling learning experiences, from delivery of content online, to novel classroom experiences and engagement with content at museums, historical monuments, or even in nature. 

One potential benefit is that they could be used to personalize and motivate learning experiences by using characters that a student likes or admires. Using state-of the-art generative machine learning models, prominent historical, modern-day, or fictional figures can be brought to life to engage learners with "lived'' experiences of scientists making their discoveries, historical figures narrating battles, or painters discussing their inspiration and process. Young children could have their teacher or a classroom guest take the form of a favorite cartoon or movie character. For instance, these students could get excitement from having an AI-generated version of Elsa from the movie Frozen teach them about the formation of snow and ice. Similarly, high school students could experience a realistic recreation of a lecture or a historical event by using AI-generated characters that are part of that narrative. Using AI-generated characters, information that may only have been available in writing or told by a third party can now be delivered by a virtual representation of the real person discussing their work or story.

To evaluate the effects of virtual characters in online learning, we (1) designed educational videos based on information from the World Health Organization, (2) used an open-source AI-generated characters pipeline to generate videos of virtual instructors narrating these videos, and (3) performed a human subject study to evaluate the impact of using AI-generated instructors that resemble people that students may like or admire, on students' learning performance, emotions, motivation, and perceptions of the instructor.    

Study participants watched the educational video comprising a slideshow and a talking-head video of a virtual instructors. Two versions of the lecture were created. In one version, the lecture was delivered by an AI-generated instructor resembling Elon Musk (a well known American innovator and tech entrepreneur). In the other version, the exact same content was taught by an unknown person of the same age, race, and gender, whose appearance was generated using a free online service that generates faces of people that do not exist. In these videos, only the face was changed, whereas the voice and gestures were identical. Elon Musk was selected since he is a somewhat controversial figure that participants would like, dislike, or admire to various degrees. This variation would help us to study how liking or admiring the portrayed character can impact students' learning outcomes, motivation, emotions, and their perception of the virtual instructor.                                   

To generate realistic looking virtual instructors that resemble inspirational, historical, or fictional figures, we used an open-source unified pipeline. It takes audio or text input along with a target image in order to output a video of a talking AI-generated character based on the image. The pipeline was selected as it is 1) easy to set up, 2) provides realistic outputs, and 3) requires modest resources. Hence, it can easily be used by educators to generate learning materials.                                      

While the degree to which participants liked or admired the instructor did not result in a significant difference in test scores between the Experimental and the Control group, students who were neutral towards the AI-generated instructor scored slightly higher than the others. It is possible that this is due to being distracted, but more investigation would be needed. 

However, their degree of liking or admiration notably impacted their feelings and attitudes towards learning and their perception of the character as an instructor. This suggests that enabling students to learn from virtual characters modeled after people they like or admire may have the potential to boost students' positive emotions, increase their motivation to learn, and enhance their perception of their instructor.

Overall, the results suggest that AI-generated instructors could be useful in multiple ways in education. They could be used to completely stand in for a teacher in an online lecture, but they could also be leveraged as guest lecturers to complement existing lesson plans. For example, a teacher could either invite a virtual Einstein to teach about the theory of relativity or deliver the lecture while puppeteering the likeness of Einstein. The approach could also be used to create personalized learning experiences.

Furthermore, specific characters could be generated to suit the lecture content and add a special touch based on their unique backstories (e.g., Einstein for physics, Picasso for painting). These virtual teachers could also be used by students in active learning scenarios, where the students can drive the virtual characters through acting or puppeteering and could craft reenactments of important events. AI-generated characters can spark imagination and creativity by blending fiction with reality. There's also the possibility for such technology to increase the representation of minorities in teaching videos by modeling virtual teachers based on generic characters or popular role models. Recent studies have shown that students more positively appraise teachers , score higher on tests, and enter more gifted programs when their teachers are of a similar ethnicity to their own. This factor could likely be replicated in the use of virtual instructors. 

The ethics issues of AI-generated media go beyond the educational setting and are the subject of an ongoing, expansive conversation happening across different scales from personal usage to national policies. Here, we focus on ethics around AI-generated characters in education.

In this paper, we focus on the ethical considerations of using AI-generated characters in the context of education. AI-generated characters can be used to create educational content that is inaccurate as well as non-representative of the person being portrayed. For example, if a deepfake of a scientist is created, they could be made to say things that are not supported by scientific evidence. This could lead to students more readily believing false information or being confused when the supposed authority in a topic provides conflicting information. AI-generated characters can be used to deliberately spread inaccurate information.  For example, a deepfaked scientist could be made to say things that are not supported by scientific evidence. Students could be lead to believe false information or could be confused when the supposed authority on a topic provides conflicting information. It is important to respect a person's privacy and seek the consent of the person to be portrayed. Deepfakes can easily be used to publicly misportray people and their beliefs, which can inflict  profound harm (e.g., defamation, emotional distress). The potential for wide distribution can compound these negative effects. One open question is how to handle consent when the person is deceased. While AI-generated characters for teaching may have economic benefits and increase access to education in low-resource areas, they should primarily be used to augment or supplement  human teachers rather than replace them. Research has shown the student-teacher relationship to be a key factor for fostering positive student attitudes, behaviors and development. Moreover, research indicates that a lack of emotional attachment, as experienced in video-conferencing-based classes, decreases  learning effectiveness, and that a reduction in social relationships adversely affects mental health, physical health, and mortality risk. Hence, substituting real teachers with virtual instructors could pose a threat to student learning and well-being. 

We acknowledge the support of the Natural Sciences and Engineering Research Council of Canada (NSERC), [PGS D3-545858-2020]. Cette recherche a été financée par le Conseil de recherches en sciences naturelles et en génie du Canada (CRSNG), [PGS D3-545858-2020]. We acknowledge the support from NTT DATA.