Does song presence in datasets guarantee AI training use?

Question

Accepted Answer

Song presence in datasets does not guarantee AI training use
The presence of a song in AI development datasets is not definitive proof that the track was used to train a generative AI model. The Atlantic, which launched The AI Watchdog search engine, stressed that a listing in the datasets cannot be treated as conclusive evidence of training. It also said that absence from the search engine does not prove a song was never used by an AI company.
What the AI Watchdog search engine can and cannot confirm
The AI Watchdog lets users check whether works appear in massive datasets used within the AI development community. An investigation found more than 21 million songs across four datasets, and a check identified at least 1,000 songs by Malaysian artists appearing in those listings. However, the dataset listing is only an indicator of whether a work is present in the database, not proof of training.
Why the listing is treated as non-definitive evidence
At least one dataset includes links to music hosted on widely used platforms such as YouTube and Spotify, which can help explain why many titles appear across different Malaysian discographies. Even so, The Atlantic explicitly stated that appearing in a dataset does not provide definitive proof that an AI system trained on it. The same point applies across examples such as Datuk Seri Siti Nurhaliza, the late Tan Sri P. Ramlee, Yuna, and Butterfingers, all of whom have multiple songs listed in the datasets.
What Malaysian publishers say they are monitoring
Malaysian music publishers are monitoring whether locally produced songs have been used without permission to train generative AI models. Taja Archive Sdn Bhd director Kimmy Khir said the publisher is aware of the AI Watchdog database and is closely monitoring developments, with the priority of protecting songwriters and copyright owners. Taja Archive represents about 900 songwriters and administers almost 20,000 registered songs, including works by Yuna, Noh Salleh, Altimet and MimiFly.
What is being sought beyond dataset appearance
The monitoring is paired with calls for greater transparency and clearer licensing practices between AI developers and rights holders. Kimmy Khir said the industry needs meaningful engagement so innovation can coexist with respect for copyright, and she is involved in the Malaysia Music Rights Alliance formed by the Music Publishers Association of Malaysia and Recording Industry Malaysia, together with collective management organisations such as MACP, PPM and RPM. The alliance has distributed an open letter to MPs and key ministries including the Prime Minister’s Office, the Digital Ministry and the National AI Office.

Reading history

Hide →

This history is cleared when the page is closed.

Does song presence in datasets guarantee AI training use?

Loading

Was this article helpful?

Helpful

Not helpful

Report an issue

1

0

0

Share with friends