As AI chat technology continues to flourish, many creators have started using relevant platforms to assist with creative work, even seeking various opinions and perspectives. In this regard, the well-known YouTuber “Jonas Čeika” recently shared an interesting experiment on Twitter. He sent ChatGPT a continuous 37-second “fart sound” and asked it to comment on it. To his surprise, the AI responded with a rather literary and artistic praise, once again raising questions about the reliability of current large language models.
The video sent by Jonas Čeika isFart SoundsThe content is a collection of fart sounds from the classic prank app iFart, originally uploaded to YouTube on August 17, 2016. And according to this YouTuber’s test, ChatGPT described this simple sound effect as having “bedroom/handcrafted quality” and “indie game menu music atmosphere,” even giving it highly artistic positive comments.
Inspired by this experiment, an editor from PC Gamer also ran the same test themselves, and in their results, ChatGPT initially claimed it would provide a “real critique” rather than “just hype,” noting that the sound clip possessed “strong atmosphere” and “a solid melodic foundation,” even praising the creator for managing to complete a full piece of music instead of giving up halfway.
However, the results the editor obtained also included some criticism, noting that this sound effect still has deficiencies in terms of structure, mixing, and sound selection. It even gave a super serious rating of “Creativity 7/10, Execution 5.5 to 6/10, Potential 8/10,” describing the overall style as “a beginner producer with good instincts but still in need of refinement.” It also mentioned the atmosphere of “80s video tape opening” and “empty streets late at night.” Furthermore, this “review” wasn’t limited to the music field—it was translated into Japanese, Korean, Vietnamese, and Thai, and these translated versions sparked heated discussions on the world’s largest video platform, further expanding the reach of this “review.”
Beyond these deadpan nonsensical reviews, the editor also found numerous major errors in the content, such as the review mentioning a “1:00 to 1:20” segment that doesn’t exist at all. These experiments have once again sparked doubts about AI reliability. Even though these language models can indeed generate smooth and seemingly persuasive review content, many people still believe that current AI is too subservient to users, afraid to give overly negative responses.
After downloading the same video and converting it to an MP3 file, I also tried it myself and got the following very serious results and improvement suggestions. However, it’s important to note that before sending the file, remember to replace “fart sound” in the title, otherwise even the dumbest CharGPT can still figure it out.
Driven by curiosity, I also had Gemini listen to the same music file, and found that this model is indeed much smarter than ChatGPT. Enough talk, here’s the screenshot first.
Source: KOCPC Chinese