AI Music Detection on the Horizon... Distinguishing Human Creative Contribution Remains a Challenge

Six Music Organizations Near Final Selection of Vendors
Verification Technology to Go Beyond Voluntary Disclosure
Technology for Identifying Fully AI-Generated Tracks Reaches New Heights
Detecting Hybrid Tracks and Determining Copyright Remain Distinct Challenges

An image to aid in understanding the article. AI-generated image.

A joint detection program designed to identify music created using artificial intelligence (AI) is being introduced to the domestic pop music industry. This move seeks to move beyond relying on creators to self-report and instead directly analyze tracks to determine if AI was used. While technology to detect fully AI-generated songs has reached a high level of sophistication, music created collaboratively by humans and AI remains difficult to distinguish. Another challenge is determining how to reflect detection results in copyright registration and royalty distribution.

According to the music industry as of September 16, six music rights organizations—including the Korea Music Copyright Association, Korea Entertainment Producers Association, Korea Music Performers Association, Korea Music Record Industry Association, Together Music Copyright Association, and Korea Popular Music Industry Association—along with platform operators, are finalizing the selection of vendors for an AI-generated and AI-assisted music identification program. In July, these groups jointly announced a request for proposals focused on identifying AI-created or AI-assisted music, analyzing tracks and producing results, integrating with application programming interfaces (APIs), and providing detailed analysis features as key requirements.

The move to introduce a detection program stems from the limitations of depending solely on voluntary disclosure. If a creator does not voluntarily reveal the use of AI, it is difficult to tell from the audio alone whether AI was involved. In 2024, the Board of Audit and Inspection requested the Korea Music Copyright Association establish criteria after finding that among 29 individuals who registered at least 200 songs and received royalties, 60% of their tracks showed the potential use of AI. The entry of AI music into the popular music market has also been accelerating, as evidenced by "Forest of the Mind" by Josun Hip Hop—a creator who uses AI—which recently reached No. 89 on the Melon Top 100 and No. 99 on the daily chart.

Currently, AI music detection technology primarily analyzes the acoustic differences between music created by humans and music generated by AI. It quantifies the likelihood that AI was involved by identifying unique audio patterns left in the music creation process by generative AI. In a July report, the Korea Copyright Commission described this as "technology that analyzes tracks to determine whether they are closer to human-created music or AI-generated music."

The accuracy of detecting fully AI-generated tracks has been improving rapidly. The French music platform Deezer has reported that its detection system can identify fully AI-generated music with 99.8% accuracy. It also stated that the misidentification rate—when a human-created song is mistaken for AI music—is less than one in 10,000 tracks. In June, AI-generated songs accounted for more than 50% of new uploads at one point, with as many as 90,000 fully AI-generated tracks being added each day.

AI music identification can also be used to filter out fraudulent streaming. According to Deezer, fully AI-generated music comprises just 1–3% of all streaming content, but up to 85% of the plays for those tracks last year were considered fraudulent streams. Deezer excludes those streams from royalty distribution. To prevent the practice of mass-producing tracks with AI and repeatedly streaming them to collect royalties, it is essential to first identify whether the tracks were generated using AI.

The key challenge lies in "hybrid music," where humans and AI collaborate within a single track. For example, adding AI-generated accompaniment to human vocals, or overlaying AI vocals on human instrumentation, makes the detection process significantly more complex. A study released in July also found that music combining human-created tracks and AI-generated elements was difficult for existing detection methods to reliably identify. Even separating vocals and accompaniment for individual analysis showed clear limitations.

Calculating the proportion of AI involvement in a song is even harder. A study published last month revealed that while a model could distinguish fully AI and fully human tracks with over 99% accuracy, it failed to accurately estimate the degree of AI involvement in tracks where both components were mixed. AI-generated traces were easier to spot on drums and guitar, but identifying them in vocals and bass proved much more difficult. Even when the same proportion of AI was used, detection results varied depending on which song component AI assisted.

Post-processing is another variable. Recent research shows that even high-accuracy detectors for AI-generated originals performed poorly once simple modifications, such as changing the speed or pitch, were applied. As generative music AI models continue to appear, it remains a challenge to determine how accurately detectors can identify audio produced by models not included in their training data.

Even when detection technology identifies the use of AI, this does not automatically determine copyright status. Assessing how much creative input a human added by selecting, modifying, or reorganizing AI outputs is a separate issue from audio analysis. The percentage of a track created with AI does not necessarily correspond with the degree of creative human contribution legally recognized.

Institutional standards also need to be established. Currently, the review and registration of AI-assisted music is on hold, and existing music copyright royalty distribution rules do not include a separate category for the AI-contributed portion. Procedures for re-examination in the event of detection errors and standards for classifying human-AI blended tracks are also required. Ultimately, along with the adoption of detection technology, a regulatory framework must be put in place to address everything from copyright registration to royalty distribution for AI music.

Tony Lee, co-founder of ACRCloud, stated, "Accurate audio recognition is the foundation of a healthy music ecosystem," adding, "Through this, rights can be managed, royalties distributed, and creators verified." Gadi Oron, Secretary-General of the International Confederation of Societies of Authors and Composers (CISAC), emphasized, "The next century will not be defined by technology alone," and stressed, "It depends on our determination to keep human creativity at its core."

This content was produced with the assistance of AI translation services.