“At the same time, 35% of the study participants stated that both songs were generated.”, — write: www.unian.ua
At the same time, 35% of the study participants stated that both songs were generated.
Only 60% of people can distinguish a live song from an AI-generated one / Photo – Velvet SundownScientists from the Federal University of Minas Gerais in Brazil conducted an experiment to find out how well people distinguish music created by artificial intelligence.
Subjects included pairs of tracks – one written by a person, the other generated. The task was simple: to determine where the real one is.
The researchers collected AI compositions from open sources: YouTube, Reddit, and the library of the Suno service. The test included songs created between July 2023 and February 2025 on Suno v3.5 and v4 models, without the latest version. In the control pair, the introduction to Beethoven’s Fifth Symphony was played – those who did not recognize it were excluded from the sample.
308 out of 653 people reached the end of the experiment, the average age of the participants was 31 years. In pairs where tracks were selected by genre, AI was correctly identified only in 60% of cases. And if the songs were from different genres – only 53%. Moreover, in 35% of cases, people decided that both compositions were artificial, even if one was written by a person.
Participants with a musical education or experience playing instruments coped a little better with the task. Those who themselves once used music generators also had higher accuracy. But with age, the ability to recognize AI tracks decreased.
Earlier, the expert told how to distinguish a neural network video from a real one. Users have already learned how to remove Sora’s watermark, so it becomes increasingly difficult to recognize the AI.
You may also be interested in news:
- A unique 3D printer designed by artificial intelligence was shown in London
- ChatGPT creator warns AI will replace ‘not real jobs’
- Elon Musk launched his own neural network analogue of “Wikipedia” – Grokipedia