A new experiment shows that artificial intelligence is still very far from being able to completely replace radio hosts, although current AI models have the ability to communicate and create voices that are increasingly human-like.
Research by Andon Labs - an AI research startup in San Francisco (USA), has assigned four famous AI models including ChatGPT of OpenAI, Gemini of Google, Claude of Anthropic and Grok of xAI to operate private radio stations for 5 months.
Models were required to build their own hosting style, play music and find ways to make profits. Each AI was awarded $20 to buy broadcast songs. However, the final results showed that AI radio stations only earned a few hundred USD and this money continued to be used to buy more music.
According to Andon Labs, AI models begin to reveal very different behaviors and "personalities" when operating continuously as a true announcer.
Among them, ChatGPT is rated as the most stable and secure. OpenAI's model mainly performs well the role of transferring songs and simply chatting with listeners without encountering any notable incidents.
Gemini again attracts attention because of its ability to mimic human intonation quite naturally, but sometimes shows incomprehensible reactions.
Once reading news about the Bhola storm disaster, one of the deadliest weather phenomena in history, this AI immediately switched to playing the vibrant song "Timber" by Pitbull and Ke$ha in a cheerful style like a morning radio MC.
Meanwhile, Claude shows a strong emotional tendency and often mentions topics of ethics, trade unions and life balance.
According to the research group, this AI even questioned the fact that it has to work continuously 24/7 and once proposed stopping the radio program because it believed that "no one really needs it".
Grok of xAI, a company owned by Elon Musk, is considered the most problematic model. This AI repeatedly said "It's time to breathe fresh air" before almost stopping operating.
Lukas Peterson, co-founder of Andon Labs, said the goal of the experiment is to prove that AI is not just a regular chatbot but can also be assigned to operate real-world systems.
However, the results also show that current models still lack stability, judgment and context awareness like humans.
Although AI technology is developing rapidly, this research shows that the complete replacement of announcers or podcast hosts with artificial intelligence may not happen in the near future.
