Three music AI trends to watch in 2025: Recording & takeaways
On October 30, 2024, Water & Music hosted a webinar diving into three emerging trends reshaping AI's role in the music industry: Synthetic data, model controllability, and content/copyright detection.
Our goal was to move beyond surface-level AI discussions to examine how these developments are concretely impacting industry power dynamics and value creation.
Below are the full webinar recording and slides, followed by a brief recap of key themes discussed. Links to all the companies, research papers, and news articles cited in the webinar are also provided.
Follow along
Webinar recording:
Slides:
Current market context: Beyond early adoption
Our session began by acknowledging an important shift in the market: We're firmly past the early adoption phase of music AI.
- Recent surveys show approximately 25% of music creators are already using AI tools — primarily for specific parts of their creative process, like stem separation or mastering, rather than full-stack song generation.
- On the business side, adoption is even higher: In our State of Data in the Music Industry survey, nearly 40% of industry professionals reported that they were incorporating AI into their data operations.
The mainstreaming of AI tools comes amid a complex funding landscape. While early 2024 saw major rounds for companies like Udio ($10M), ElevenLabs ($80M), and Suno ($125M), investment has cooled significantly following high-profile lawsuits from major labels and publishers. This behavior closely mirrors previous hype cycles we've seen with technologies like Web3 and livestreaming.
Key trends and implications
1. Synthetic data
Synthetic data refers to artificially generated information used to train AI models when real-world data is scarce, expensive, or raises privacy concerns.
Due to a mix of rising licensing costs and a desire for more granular control and flexibility over the model training process, AI developers are increasingly exploring toward synthetic alternatives to acquiring data, in key areas like:
- Plugin generation (e.g. new research from Positive Grid)
- Stem mixing and shuffling (e.g. Sony's Diff-a-Riff)
- Data labeling (e.g. Tencent using LLMs to generate labels for 22,000 recordings compared to traditional institutional methods yielding only 700)
This trend could disrupt traditional licensing models between developers and rights holders, by challenging the assumption that there will always be an appetite from developers to acquire “traditional” data from the music industry.
However, current AI-generated audio isn't yet high-quality enough to serve as training data for music models; otherwise, degradation would compound with each successive generation. The likely path forward involves hybrid approaches, combining synthetic data generation techniques with properly licensed content.
Regardless, short-term market forces are significantly increasing research activity in the realm of synthetic data. As our analyst Yung noted: "Your data right now is about as valuable as it's ever going to be."
2. Model controllability
Model controllability measures how precisely users can guide and edit AI outputs. Modern tools offer controllability at two levels: Detailed prompt specifications (e.g. key, BPM, structure), and granular editing of outputs (versus having to re-prompt an entirely new output from scratch).
We demonstrated several breakthrough examples, including:
- Google's MusicFX DJ, for real-time music generation with several adjustable parameters
- Udio’s audio inpainting
- ACE Studio’s granular voice modeling and editing tools, with precise control over vibrato, tension, and even voice blending
The market implications here are significant. As music AI creation tools become more sophisticated, the traditional advantage of professional composers and producers — delivering highly customized, bespoke work — may erode.
The grounded example we offered in our webinar was that of a music supervisor on a tight budget and timeline. Given the option of being able to generate a piece of music to fit a need in that scenario, versus spending weeks hunting for an existing song and negotiating with rights holders, it's easy to see how market forces might lead the supervisor to choose the first option.
Success in the music industry will increasingly depend on taste and creative direction rather than technical expertise — similar to Rick Rubin’s approach of being valuable through decisive creative judgment, rather than technical knowledge.
3. AI content & copyright detection
AI content & copyright detection encompasses the technical infrastructure that rights holders and platforms use to protect copyrighted music at every step of the AI pipeline — from auditing training data and scanning user uploads, to detecting AI-generated tracks and voice clones. (We covered this trend in-depth in a previous Water & Music article.)
The industry is developing multi-layered approaches to protect copyrighted music in the AI age:
- Auditing training data (e.g. manual certifications like Fairly Trained and AI:OK, or direct technical integrations like BMAT x Voice-Swap)
- Attributing outputs to the influence of specific pieces of training input (e.g. new technologies from ProRata.AI, Musical AI, Human Native AI, and Sureel)
- Detecting music and voice deepfakes (e.g. Pex's voice clone detection, IRCAM's AI music detector)
- Scanning user reference uploads (e.g. Suno x Audible Magic)
- Watermarking social media content (e.g. “Made with AI labels” across Meta, TikTok, and YouTube)
Each approach has its own technical and political limitations. For instance:
- Deepfake detection tools need constant updating as AI models improve, and will never have 100% coverage as the open-source developer community continues to experiment with their own homegrown models.
- Watermarking systems rely heavily on voluntary participation from users, and can also be easily circumvented.
- There’s no reliable way to identify and label partially AI-generated content (e.g. if you used AI to make the vocals, but hand-produced every other stem in the song), leading to potential misrepresentation of hybrid creative processes.
Looking ahead: Diverging incentives
The central question emerging from these trends is: Who will have leverage in an AI-first music industry?
Our analysis suggests that different stakeholders have conflicting interests.
- Artists need strong brands and communities, and could potentially benefit from loosening up their IP to maximize reach.
- Rights holders need to protect their market share and maximize catalog value as it peaks in the context of synthetic data.
- Developers and founders need the most cost-effective ways to build products people actually want to use.
As Alexander observed during our webinar: "It's not a new trend... artists are increasingly having to leverage their brand and community. If you really take that to its logical conclusion, to do that effectively requires a looser approach to IP. And I think that kind of contradicts the approach that rights holders are going to want to take."
To dive even deeper, our full music AI market tracker includes detailed profiles of 200+ music AI startups and tracking of 300+ AI news developments.
What's next? Stay tuned for upcoming webinars and articles where we'll continue exploring the intersection of music, technology, and business. For any further questions or to share your thoughts, please reach out to our inbox at members@waterandmusic.com!