Does AI help musicians practise independently?
A review of 16 heterogeneous studies maps how recording tools, error detectors and generative assistants may support practice. Short-term findings are encouraging, but no study tested whether gains persist after the support is removed.
The short version
- What happened
- The PRISMA review screened 127 unique records and included 16 empirical studies: seven traditional digital tools, four discriminative AI systems and five generative-AI studies.
- Who it affects
- Learners can use AI to surface errors or practice options while keeping the explanation and final judgment in human hands.
- What remains unproven
- Sixteen heterogeneous studies and no pooled effect estimate
Keep in mind
Sixteen heterogeneous studies and no pooled effect estimate
Uneven quality, including small, cross-sectional and quasi-experimental designs
English-language peer-reviewed literature only and no grey-literature search
Review was not prospectively registered
No test of retained gains after tool removal and no deliberate scaffold fading
Source analysis · 1 linked source. Primary analysed source date: 15 October 2026.
Read the full report · 6 min The evidence, methods, implications and limitations
The answer: tools may scaffold practice, but independence was not tested
Digital recording, automated error detection and conversational AI can help learners set goals, notice mistakes and reflect on a practice session. A systematic review of 16 empirical studies found examples across all three phases of self-regulated learning: planning before practice, monitoring during performance and evaluating afterwards. Yet the review cannot answer the most important autonomy question. None of the included studies tested whether improvements survived after the technology was removed, and none deliberately faded support as learners became more capable.
That means a tool can make a session feel more guided without proving that a musician has learned to guide themselves. The distinction matters for teachers, parents and conservatoires. Immediate correction may reduce repeated errors, but constant prompts can also move attention away from listening and make the system—not the learner—the judge of quality. The evidence supports careful use of AI as a scaffold. It does not support replacing teacher judgment, auditory attention or the slow development of metacognition.[1]
What the review included
The authors searched Web of Science, Scopus, ERIC and PsycINFO from the earliest records to March 2026, then traced references backwards. They found 145 records, removed 18 duplicates and screened 127 titles and abstracts. Eighteen full texts were assessed, three were excluded and one additional paper was found through citations, leaving 16 empirical studies. Two reviewers screened independently, with Cohen’s kappa of 0.82, and resolved disagreements by discussion.
Eligible work had to connect an AI or digital intervention with at least one part of Zimmerman’s self-regulated-learning cycle. The corpus included experiments, quasi-experiments, surveys, qualitative studies and two full conference proceedings. Samples ranged from a four-person case study to surveys of 941 and 1,125 learners. Ages, instruments, teaching contexts and outcome measures differed too much for a valid meta-analysis, so the authors mapped recurring mechanisms through thematic synthesis.[1]
Three kinds of support appeared
Seven studies used traditional digital tools such as video playback, e-portfolios, recording comparison or reflective journals. These systems chiefly helped learners plan, document work and look back on performance. Four studies used discriminative AI to detect pitch, rhythm, gesture or other errors in real time. Five used generative AI for dialogue, suggestions, composition or explanations. The categories describe a primary function, not watertight technologies; a modern product may combine detection and language generation.
Across the review, traditional tools were strongest where a learner already had a goal and needed a record. Error detectors made otherwise invisible mistakes more noticeable during performance. Generative systems could translate feedback into next steps or sustain motivation through dialogue. Those patterns are plausible, but they are not common effect sizes. A positive survey association, a small classroom comparison and a usability study cannot be added together as if they measured the same outcome.[1]
Evidence quality was uneven
Two reviewers appraised every study with the Mixed Methods Appraisal Tool and reached 87.5% initial agreement, again with kappa of 0.82. Four studies met all or nearly all applicable criteria, while others had weak control of confounding, unclear sampling or incomplete reporting. Several recent generative-AI findings came from small quasi-experiments, cross-sectional surveys or qualitative evaluations. Large samples do not solve causality when learners choose whether and how to use a tool.
The review itself followed PRISMA 2020 but was not prospectively registered and had no pre-published protocol. It included only English-language peer-reviewed work and did not search grey literature, which can produce publication and language bias. Its 16 studies are a snapshot of a fast-changing field and skew toward work published in the last three years. The authors report no external funding or commercial conflict and state that generative AI was not used to create the manuscript.[1]
What teachers and learners should do differently
Teachers can decide in advance which decisions should remain with the learner. An error detector might flag a note while the student still explains why it happened and chooses the next exercise. A generative assistant might propose practice strategies while the student predicts which will work, listens back and records their own assessment. Periodic sessions without the tool can reveal whether monitoring and planning have actually transferred. The goal is to make support unnecessary, not merely pleasant.
Visual feedback deserves restraint. A dense display can make a learner watch the screen rather than listen to tone, balance and phrasing. Feedback should therefore be timed, selective and appropriate to proficiency. Beginners may benefit from a small number of clear cues; advanced musicians may need space for interpretive judgment that an error label cannot capture. Teachers should also inspect privacy terms before students upload recordings and avoid systems that train on work without meaningful consent.[1]
What would change the assessment
The next useful studies would randomly compare ordinary instruction, stable AI support and support that deliberately fades. They should measure performance with blinded human assessment, track independent practice behaviour and retest after the system is removed. Longer follow-up could show whether learners keep planning, detecting and correcting problems on their own. Results should be reported by age, expertise, instrument and disability rather than assuming one practice model fits everyone.
Confidence would also rise if researchers preregistered protocols, shared measures and replicated promising tools outside the developer’s institution. Student interviews should examine frustration, trust and attention as well as convenience. For now, AI appears capable of supporting parts of the practice cycle, but the review does not prove that it produces autonomous musicians. The best near-term test is simple: after the prompt disappears, can the learner still hear the problem, choose a strategy and judge the result?[1]