AI music generators like Udio and Suno can produce remarkably polished songs from a text prompt, yet they lack something essential that human composers have always relied on: a deliberate design phase before execution. According to François Pachet, a legendary AI researcher who has spent decades studying how machines can create music, this missing step explains why even the most impressive AI-generated tracks often feel directionless and unsatisfying.

What's Actually Missing From Today's AI Music Tools?

Pachet, who previously led research labs at Spotify and Sony Computer Science Lab and created influential systems like the Continuator and Flow Machines, recently shared his analysis of why platforms like Udio and Suno, despite their technical sophistication, fall short of true musical creativity. The core issue, he explained, is architectural: these systems are "end-to-end," meaning they jump directly from a text prompt to a fully mixed, mastered song without any intermediate design phase.

"These systems, they are called end-to-end. You start by a prompt, which is a text, and you end up with a completely fully-fledged mixed multi-track, even master. So, you have the complete end product, from end to end. And obviously what's missing is the way humans create something," Pachet explained.

François Pachet, AI Researcher and Musician

To illustrate the problem, Pachet drew a parallel to how architects and composers actually work. When an architect designs a house, they first sketch plans, then refine them through trial and error, and only then hand them to contractors for execution. Similarly, when Paul McCartney composed "Yesterday," he woke with a melody in his head, played it on guitar while refining it, and only then went to the studio to record it.

This two-phase process, Pachet argued, has always been fundamental to human creativity across all domains: there is the composition phase, where ideas are shaped and refined, and then the production phase, where those refined ideas are executed. Udio and similar platforms collapse this into a single step, skipping the crucial iterative design work that gives music its character and intentionality.

How AI Music Generation Could Evolve Beyond Current Limitations

• Implement Search and Iteration: Rather than generating a complete song in one pass, AI systems could explore a space of musical possibilities, try different approaches, recognize when something doesn't work, backtrack, and pursue alternative paths, mimicking how human composers refine their work.
• Separate Composition From Production: Build systems that first generate a musical "plan" or skeleton, allowing users to review, critique, and refine that plan before the system produces the final polished recording.
• Combine Sampling With Search: Merge sampling techniques with search algorithms to create systems that can explore variations and combinations of musical ideas rather than committing to a single output immediately.

Pachet emphasized that the next frontier in AI music generation requires combining sampling with search, a shift that would fundamentally change how these systems approach creativity. Rather than treating music generation as a one-shot production task, researchers need to build systems that can explore, evaluate, and refine musical ideas iteratively.

Why the Mystery of Musical Creativity Remains Unsolved?

Despite decades of progress in AI and deep learning, the fundamental question of how humans compose music remains largely unanswered. Pachet noted that while researchers have made strides in understanding style, texture, and how to combine different musical influences, the core mystery of creating something genuinely new and compelling persists.

"The day I hear a song of the quality of the Beatles, I will say: 'Okay, we are done'. And I've never heard anything like that. Never," Pachet stated.

François Pachet, AI Researcher and Musician

This gap between current AI capabilities and true musical creativity is not a failure of engineering, Pachet suggested, but rather evidence that the problem itself remains fundamentally open. Unlike many domains where AI has surpassed human performance, music generation sits in a unique space: the technology can produce impressive results, yet no one fully understands what makes music genuinely compelling to listeners.

Pachet has personally experimented with Udio and Suno, and while he acknowledged their impressive technical achievements, he remained unconvinced that they solve the core creative problem. "I have used some of these tools, like Suno and Udio, and I'm still not satisfied. There is still a lot to do," he noted.

The broader implication of Pachet's analysis is that AI music generation is not a solved problem, despite the polished outputs these platforms produce. Instead, it represents an early stage in a much longer journey toward understanding and replicating human creativity. The systems that will truly advance the field, he suggested, will be those that embrace the messiness and iteration of human composition rather than trying to automate it away entirely.