No People Required? How AI Designed, Curated and Delivered a TEDxMelbourne Event

Imagine attending a TED-style event where the keynote speakers aren’t people but artificial intelligence. You might be wondering if AI is capable of ideating, developing and delivering a TED-inspired talk, and if so, could it match the delivery, nuance and provocation of a human speaker? 

To test this idea and explore the boundaries of AI, the TEDxMelbourne team, drawing on the power of a range of AI technologies, curated a series of TED-themed talks delivered entirely by AI avatars. Considering the spectrum of AI tools available to help us write more efficiently, research and analyse information at record speed, you may assume AI has the capacity to develop ideas worth spreading. However delivering an exclusively AI-led event wasn’t as straightforward as you might expect. 

The art and science of AI

Before considering what topics to present, the team had to select the right tools and understand how best to use them to generate interesting content. The key to producing the right type of information was learning to ask the AI the right questions. This involved designing prompts to help guide and coach the AI to generate a TED-like talk based on a given topic. 

Prompting questions not only helped develop the content, but also to achieve the right language and tone to reflect TEDxMelbourne’s unique style. A great TED talk is engaging and dynamic and asks provocative questions while taking the audience on a narrative journey. It was important that content generated by AI had the same effect, but at the time of development, AI texts often lacked the natural rhythm, flow, and coherence of human language. To counteract this, the team used AI to recognise patterns, themes and speech styles from over 400 previous TEDxMelbourne talks and drew on these to further refine the AI prompting questions. 

As a final measure of quality, the team fact-checked the information and made tweaks for flow and structure. The written content was then fed into a text-to-speech program to bring the words to life through a human-like avatar supported by D-id Labs video generation software. Out of this process the curation team developed eight distinct TED-style talks. 

AI as the storyteller 

As the audience gathered around a central projector screen, chatter focused on what to expect from an event where AI was not only the topic of discussion but the medium. Would AI avatars command and engage the audience and would the content of the talks inspire deeper reflections on AI’s place in the world? The talks were preceded by a short introduction by Head of Curation, Jon Yeo, and audience members were invited to share short reflections in the breaks. The lights dimmed and a two-dimensional figure appeared on the projector screen. 

The topics of discussion were broad ranging and included AI ethics, the role of AI in the medical industry and AI as a force for good. Each avatar had a unique voice and aesthetic look, specifically designed by the team to mimic different speaker styles, backgrounds and personalities. Limitations of the text to speech program Eleven Labs, used to create the avatar voices, meant the accents leaned towards either American or British English. This was the case despite using AI-voice cloning technology to train the AI using voices of previous TEDxMelbourne speakers. Facial expressions somewhat mimicked those of a human with a slight robotic quality and now and again the avatars gesticulated awkwardly. While mechanical and lacking emotional depth, they undoubtedly had a human likeness that was recognisable, if peculiar and unnatural. 

‍

What we can learn from an AI-led event

Interestingly, audience reflections focused less on the topics presented by the AI and more on the curation process which brought the event to life. A recurring theme was the human effort it took to remain engaged with the AI as a presenter, when the normal visual and aural cues of a speaker were distorted or strange. One audience member admitted closing their eyes and imagining the event as a podcast which helped to draw attention away from the gawky and emotionless style of the avatar. There was agreement that without a person to deliver the information in a compelling and clear way, the content of the talks was challenging to absorb. Audience members reflected on some of the strengths and weaknesses of AI technology. As a tool to collate and analyse information quickly it is invaluable, but to present and persuade the same information to other people? There is a long way to go. 

Thinking about the effort and time to curate an AI-led event, Jon shared that in comparison with working with real speakers to develop and refine a TED talk, AI has clear challenges. Yes, AI can draft a first, second or third draft in minutes but it lacks the subtlety and human intelligence to tell a compelling narrative. For now, we remain the most important storytellers.