Detection of AI-Generated Music Using Convolutional Neural Networks

Overview

The rapid growth of generative AI music is causing unprecedented content overload and a flood of spam on streaming platforms, posing significant reputational and income risks to human artists. For instance, platforms like Deezer report that 34% of daily uploads are now AI-generated, yet these tracks account for only 0.5% of total listens, illustrating a massive surge in AI-generated spam. Traditional methods for identifying AI-generated music, such as user reporting, human reviews, and rule-based algorithms, face significant challenges in speed, reliability and scalability of detection. The goal of this project is to develop a novel approach, using a Deep Learning-model based on Convolutional Neural Networks (CNNs), to distinguish AI-generated music from human-generated music, enabling online music platforms to quickly detect AI-generated content and take appropriate remedial action at the source. Unlike prior work, the current approach analyzes the underlying structural patterns, such as Mel Spectrograms, rather than sequentially analyzing audio samples. The hypothesis is that CNNs, which treat spectral patterns in music as visual data, will pick up the differences between AI-generated and human-generated music more effectively and outperform traditional models in this classification task. The proposed method involved training a CNN-based Deep Learning model on Google Colab (using Google TensorFlow) with Mel Spectrograms generated from 1140 labeled AI-generated and human-generated songs, in a variety of genres. The model was then tested on a random sample of 286 AI-generated and human-generated songs and the accuracy was used to evaluate its predictive performance. The results were promising with the model predicting the outcome with 88% accuracy and taking an average time of 0.138 seconds per song, to complete the task, making it an effective and practical approach. The current approach has some limitations (e.g., access to publicly available AI-generated music samples, using alternate image recognition models like Alexnet) and had to make some trade offs (e.g., reduce the resolution of Mel Spectrogram images) to fit the project constraints. These present opportunities for future work and further improve the accuracy of the current approach.

Competition history

• CSEF 2026 Computational Science (Senior Division) · Entry S-07-20

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Browse more like this

• All Computational Sciences projects
• CSEF Computational Sciences projects
• Every CSEF 2026 project