What Professional Musicians Actually Do With AI

In the summer of 2025, producer Mark Ronson — known for work with Amy Winehouse, Lady Gaga, and Bruno Mars — gave an interview to The Guardian in which he described using AI tools for stem separation, variation generation, and scratch vocal production, and then declined to use them for the parts of production he found most enjoyable. “I use it to get rid of the parts that feel like data entry,” he said. “The actual music still has to come from somewhere.”

Ronson’s framing — AI for administration, human for creation — is the dominant model among working professional musicians in 2027. It is not universal, and it conceals a lot of variation in what “administration” means and where the line falls. But it is a more accurate picture of studio reality than either the replacement narrative or the “AI changes nothing” dismissal.

The Practical Integrations

The AI tools that professional musicians have actually adopted fall into a few clear categories, and the pattern of adoption is surprisingly consistent across genre and career level.

Stem separation is the simplest and most universally adopted. Systems like Spleeter (open source, 2019), and its successors in commercial DAW plugins, can separate a mixed audio track into instrument components — vocals, bass, drums, melody — with quality good enough for remix work, sample clearance analysis, and live performance backing tracks. This was technically possible before 2020 but required expensive proprietary tools. It is now a standard feature in Logic Pro and Ableton. Every mixing engineer uses it.

Melodyne-style pitch correction has been in studios since the late 1990s, so auto-tune is not an AI story, but the newer generation of AI pitch and timing tools (iZotope Nectar, Antares with its neural pitch correction) does something qualitatively different: it can correct pitch while preserving the expressive micro-inflections that make a vocal performance feel human, rather than just snapping to the nearest semitone. Several vocalists who had previously resisted pitch correction have adopted these tools specifically because the newer algorithms don’t destroy the expressiveness of the original take.

Variation generation is newer and more interesting. Tools like Suno (the commercial version, post-litigation), Udio, and the variation generators in several professional DAW plugins can take a melody, chord progression, or rhythmic pattern and generate variations — different arrangements, different harmonic treatments, different rhythmic approaches. Professional musicians use these primarily as inspiration tools and scratch-idea generators, not as final outputs. The workflow is: generate twenty variations, identify two or three that suggest a direction you wouldn’t have thought of, develop those manually. The AI as starting-point-generator rather than final-output-producer.

The Specific Things They Don’t Automate

The negative case — what professional musicians have tried with AI and abandoned, or explicitly chosen not to use — is as illuminating as the positive case.

Lyric generation is the most consistent refusal. Even musicians who use AI extensively for production describe lyric generation as something they actively avoid, with explanations that cluster around the same idea: that the writing of lyrics is the part of their work that is most directly tied to their identity, that audiences respond to lyrics precisely because they carry the credibility of having been written by a person with specific experiences, and that AI-generated lyrics are recognizably hollow in a way that AI-generated instrumentation is not. Whether this is a permanent feature of lyrics or a temporary quality gap is debated, but the current practitioner consensus is strong.

Melody generation as a primary compositional tool is also largely resisted by established professional composers, though with more nuance. The distinction is between using AI to generate variations on an existing melody (accepted) versus using AI to generate the initial melodic idea (rejected by most). The initial melodic idea is described as the “spark” — the thing that makes a song this song rather than some other song — and there’s a strong cultural consensus that the spark should be human.

The Economic Reality Below the Headline Artists

Below the level of Mark Ronson and his peers — the established artists with career equity, devoted audiences, and the contractual leverage to control how their work is produced — the picture is considerably less comfortable.

Session musicians whose primary income came from recording services have experienced genuine income decline. The specific work most affected: string arrangements for pop records, which can now be synthesized convincingly enough for commercial release without session players. Brass and woodwind sections for the same context. Orchestral mockups for film scoring, which have moved almost entirely to AI-generated sketches that are then developed by smaller human sections than previously.

The studios haven’t disappeared. The budgets have contracted. Where a mid-tier film score might previously have employed thirty session musicians for three days, it now employs eight for two days, with AI-generated material filling the rest. The work that remains is more technically demanding (the AI handles the straightforward passages) and less regular (fewer projects employ large live sections). The working musician community’s median income from session work has declined in real terms even as demand for music has not.

What Adaptation Looks Like

The adaptations that have actually worked — again, not at the headline artist level but at the working professional level — share a common structure: moving toward services that AI cannot easily provide.

Live performance is the most obvious and most discussed. AI cannot perform live. The concert economy, despite predictions to the contrary, has not been harmed by AI music generation — it has been robust because the value of live performance is social and experiential in ways that recorded music is not. Musicians who have reoriented toward live income have done better than those who depended on recording income.

Teaching is another. AI music tools are complex enough that a substantial teaching market has emerged for instruction in AI-assisted music production. The musicians best positioned to teach AI production workflows are professional musicians with studio experience who have also learned the tools — a combination that will remain relatively scarce for the near future.

Sound design for immersive experiences — video games, VR applications, interactive installations — has expanded rather than contracted, partly because these applications require bespoke sound design that cannot be adequately handled by stock AI generation, and partly because the overall market for interactive entertainment has grown.

The replacement narrative is wrong in specifics and right in aggregate. Specific professional musicians are not being replaced. The economic structures that supported a class of professional musicians — the session economy, the sync licensing middle market, the production music library — are contracting. People who held specific positions in those structures are facing real economic pressure. The adaptation is real but it is not sufficient for everyone who needs it to work.