Title: HAIM: A Benchmark for Tracking Human-AI Integration in Music Production Datasets

Abstract:

With generative platforms like Suno and Udio achieving audio fidelity comparable to human standards, the application of artificial intelligence has permeated every stage of the music production pipeline. These technologies have moved far beyond basic track creation, driving the adoption of AI-driven techniques for vocal synthesis, arrangement, and professional mastering. However, existing detection research is predominantly limited to a binary "AI versus human" framework, which does not accurately mirror the complexities of modern production environments. In practice, AI tools are frequently utilized to polish or master tracks originally created by humans, while human engineers often refine AI-generated content to meet professional standards. Additionally, users may employ adversarial strategies, such as applying human-style mastering to AI outputs, to evade detection systems. This ambiguity creates a nuanced landscape that simple binary classification cannot adequately address.

To tackle this issue, this paper introduces and explores "AI Music Tracking," a framework aimed at identifying specific points of AI integration throughout the multifaceted spectrum of music production. We present HAIM, a novel dataset featuring granular labels for various production stages. HAIM is engineered to isolate distinct phases of AI intervention, enabling the tracking of hybrid production methods and agent-level involvement. Our assessment of current state-of-the-art detectors exposes significant systemic weaknesses. By publishing HAIM, we aim to establish a new benchmark that transitions the field away from binary classification toward a more detailed and structured evaluation of AI's role in music.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC