Over the past couple of years, the learning and development industry has been rocked by the advent of LLMs. Of the jobs that AI is going to make obsolete, instructional designers have been warned that theirs is among the first on the chopping block. If you talk to an instructional designer, they’ll warn you against using an LLM to spit out content without one of them in the loop; that there’s a je ne sais quoi to instructional design that can’t be replicated by a machine. But what does research say? Do people learn as well from AI generated content as they do content created by a person?
First, let’s talk about what we would expect to find. It should not surprise you that these conversations have already been had in a slightly different context. In the 1990s, with the advent of multimedia learning content, researchers were asking questions about whether learners were getting as much out of content that included videos, pictures, and other types of media. Richard Clark (1994) argued that media is a vehicle for learning, but doesn’t cause it. He points out that there’s a difference between instructional design and content delivery, and it’s the instructional design that drives learning. So, if the same design principles were used, people should learn just as well from AI-generated content as from content created by a person.
Why would they be different? Richard Mayer’s (2021) work on the voice principle holds that people learn better from content delivered by a person versus a synthetic voice. However, a lot of the research comparing learning outcomes and preferences was done using older text-to-speech technologies, and this technology has come a long way since then. It may be the case that people don’t learn as well from AI voices when they suck.
One exciting upside of recent AI advances: researchers can now isolate specific aspects of video content in experiments. Deng and colleagues (2026) did just that, testing whether AI-generated instructors and synthetic voices affect learning. They assigned 130 participants to one of four video conditions, crossing human vs. AI-generated instructors with human vs. synthetic narration. Neither factor had a significant direct effect on motivation or retention. However, participants saw AI instructors as less human-like, and that perception was linked to lower motivation and, indirectly, poorer retention. The findings suggest that AI-generated instructional videos can support learning outcomes comparable to human-delivered instruction, but perceptions of the instructor may still influence the learning experience.
As always, this is one study and should be taken in the context of the literature as a whole. But I’m excited for the next few years of research in this area. As AI-generated content becomes indistinguishable from what people create, we have the opportunity to separate the design from the delivery and investigate these principles more thoroughly.