The Occasion

“Next level?” The question mark is doing real work. The framing HfMT chose for the day was appropriately provisional: not a declaration that AI has already transformed music education, but an invitation to ask whether, in what direction, and at what cost.

The invitations that reach me for events like this tend to come with one of two framings. The first is enthusiasm: AI is coming, we need to get ahead of it, here are tools your students are already using. The second is anxiety: AI is coming, it threatens everything we do, we need to protect students from it. Both framings are understandable. Neither is adequate to the curriculum question, which is slower-moving and more structural than either suggests.

I prepared three sets of handouts. The first covered data protection — the least glamorous topic in AI education, and the one that most directly determines what can legally be deployed in a university setting. The second covered AI tools for students: what exists, what it does, and what critical thinking skills you need to use it without being used by it. The third covered AI for instructors: where it helps, where it flatters, and where it makes things worse.

This post does not recapitulate the handouts. It addresses the question I kept returning to across all three workshops: what does this change about what a music student needs to learn?

What the Technology Actually Is

My physics training left me professionally uncomfortable with hand-waving — including my own. Before discussing curriculum implications, it is worth being specific about what these tools are.

The dominant paradigm in current AI — responsible for ChatGPT, for Whisper, for Suno.AI, for Google Magenta, for the large language models whose outputs are now visible everywhere — is the transformer architecture (Vaswani et al., 2017). A transformer is a neural network that processes sequences by computing, for each element, a weighted attention over all other elements. The attention weights are learned from data. The result is a model that can capture long-range dependencies in sequences — text, audio, musical notes — without the recurrence that made earlier architectures difficult to train at scale.

What this means practically: these models are trained on very large corpora, they learn statistical regularities, and they generate outputs that are statistically consistent with their training distribution. They are not reasoning from first principles. They do not “know” music theory the way a student who has internalised harmonic function knows it. They have learned, from enormous quantities of text and audio, what tends to follow what. For many tasks this is sufficient. For tasks that require understanding of underlying structure, it is not — and the failure modes are characteristic rather than random.

BERT (Devlin et al., 2018) showed that pre-training on large corpora and fine-tuning on specific tasks produces models that outperform task-specific architectures on a wide range of benchmarks. The same transfer-learning paradigm has spread to audio (Whisper pre-trains on 680,000 hours of labelled audio), to music generation (Magenta’s transformer-based models produce melodically coherent sequences), and to multimodal domains. The technology is mature, improving, and available to students now. Knowing what it is — not just what it produces — is the starting point for any sensible curriculum discussion about it.

The Data Protection Constraint

Before any discussion of pedagogical benefit, there is a legal boundary that most AI-in-education discussions skip over. In Germany, and in the EU more broadly, the deployment of AI tools in a university setting is governed by the GDPR (DSGVO, Regulation 2016/679) and, at state level in NRW, by the DSG NRW. The constraints are not abstract: they determine which tools can be used for which purposes with which students.

The core principle is data minimisation: only data necessary for a specific, documented purpose may be collected or processed. When a student uses a commercial AI tool to get feedback on a composition exercise and enters text that could identify them or their institution, that data may be stored, processed, and used for model improvement by an operator whose servers are outside the EU. Whether such transfers remain legally valid under GDPR after the Schrems II ruling (Court of Justice of the EU, 2020) is contested — and “contested” is not a position in which an institution can comfortably require students to use a tool.

The practical upshot for curriculum design is this: AI tools running on EU servers with documented processing agreements can be integrated into formal coursework. Commercial tools whose terms specify US-based processing and model training on user data cannot be required of students. They can be discussed and demonstrated, but making them mandatory puts students in a position where they must choose between their privacy and their grade.

This is not a reason to avoid AI in teaching. It is a reason to be honest about the regulatory landscape, to distinguish clearly between tools you can require and tools you can recommend, and to make data protection literacy part of what students learn. The skill of reading a terms-of-service document and identifying the data flows it describes is not a legal skill — it is a general literacy skill that matters for every digital tool a music professional will use.

What Changes for Students

The question I was asked most often across the three workshops was some version of: “If AI can already do X, should students still learn X?”

The question is less simple than it appears, and the answer is not uniform across skills.

Skills where automation reduces the required production threshold do exist. A student who spends weeks mastering advanced music engraving tools for score production, when AI can generate a usable first draft from a much simpler description, has arguably spent time that could have been better allocated elsewhere. Not because the underlying skill is worthless — it is not — but because the threshold of competence required to produce a working output has dropped. The student’s time might be more valuable spent on something that has not been automated.

Skills where automation creates new requirements are more interesting. Transcription is a useful example. Automatic speech recognition — using models like Whisper for spoken-word transcription, or specialised models for audio-to-score music transcription — is now accurate enough to produce usable first drafts from audio. This does not eliminate the need for transcription skill in a music student. It changes it. A student who cannot evaluate the output of an automatic transcription — who cannot hear where the model has made characteristic errors, who does not have an internalised sense of what a correct transcription looks like — is unable to use the tool productively. The required skill has shifted from production to evaluation. This is not a lesser skill; it is a different one, and it is not automatically acquired alongside the ability to run the tool.

Skills that automation cannot replace are those that depend on embodied, situated, relational knowledge: stage presence, real-time improvisation, the subtle negotiation of musical meaning in ensemble, the pedagogical relationship between teacher and student. These are not beyond AI in principle. They are far beyond it in practice, and the gap is not closing as quickly as the generative AI discourse sometimes suggests.

The curriculum implication is not “teach less” or simply “teach differently.” It is: be explicit about which category each skill falls into, and design assessment accordingly. An assignment that asks students to produce something AI can produce is now testing something different from what it was testing two years ago — not necessarily nothing, but something different. The rubric should reflect that.

What Changes for Instructors

The same three-category analysis applies symmetrically to teaching.

Routine task automation is genuinely useful. Generating first drafts of worksheets, producing exercises at different difficulty levels, transcribing a recorded lesson for later analysis — these are tasks where AI can save meaningful time without compromising the pedagogical judgment required to make use of the output. Holmes et al. (2019) identify feedback generation as one of the clearer wins for AI in education: systems that provide immediate, targeted feedback at a scale that human instructors cannot match. A transcription model listening to a student practice and flagging rhythmic inconsistencies does not replace a teacher. It extends the feedback loop beyond the lesson hour.

Content generation with limits is where AI is most seductive and most dangerous. A model like ChatGPT can produce a reading list on any topic, a summary of any debate in the literature, a set of discussion questions for any text. The outputs are fluent, plausible, and frequently wrong in ways that are difficult to detect without domain expertise. Jobin et al. (2019) and Mittelstadt et al. (2016) both document the broader concern with AI opacity and accountability: when a model produces a confident-sounding claim, the burden of verification falls on the user. An instructor who outsources the construction of course materials to a model, and who lacks enough domain knowledge to catch the errors, is not saving time — they are transferring risk to their students.

Hallucinations — outputs that are plausible in form but false in content — are not bugs in the usual sense. They are a structural consequence of how generative models work. A model trained to predict likely next tokens will produce the most statistically plausible continuation, not the most accurate one. For music education, where historical facts, composer attributions, and music-theoretic claims need to be correct, this matters. The model’s fluency is not evidence of its accuracy.

Personalisation is the most-cited promise of AI in education (Luckin et al., 2016; Roll & Wylie, 2016) and the hardest to evaluate in practice. The argument is that AI can adapt instructional content to individual learners' needs in real time, producing one-to-one tutoring at scale. The evidence in formal educational settings is more mixed than the boosters suggest. What is clear is that personalisation at scale requires data — and extensive data about individual students’ learning trajectories raises the same data protection concerns already discussed, in more acute form.

The Music-Specific Question

I want to be direct about something that came up repeatedly across the day and that the general AI-in-education literature handles badly: music education is not generic.

The skills involved — listening, performing, interpreting, composing, improvising — have a phenomenological and embodied dimension that does not map cleanly onto the text-prediction paradigm that most current AI systems instantiate. Suno.AI can generate a stylistically convincing chord progression in the manner of a named composer. It cannot explain why the progression is convincing in the way a student who has internalised tonal function can explain it. Google Magenta can generate a continuation of a melodic fragment that is locally coherent. It cannot navigate the structural expectations of a sonata form with the intentionality that a performer brings to interpreting one.

This is not a criticism of these tools. It is a description of what they are. The curriculum implication is that music education must be clear about what it is teaching: the product — a score, a performance, a composition — or the process and understanding of which the product is evidence. Where assessment focuses on the product, AI creates an obvious challenge. Where it focuses on demonstrable process and understanding — including the ability to critically evaluate AI-generated outputs — it creates new opportunities.

The more interesting question is whether AI tools can make musical process more visible and discussable. A composition student who uses a generative model, notices that the output is harmonically correct but rhythmically inert, and can articulate why it is inert — and then revise it accordingly — has demonstrated more sophisticated musical understanding than a student who produces the same output without any generative assistance. The tool does not lower the standard; it shifts where the standard is applied.

There is an analogy in music theory pedagogy. The availability of notation software that can play back a student’s harmony exercise and flag parallel fifths changed what ear training and harmony teaching emphasise — but it did not make harmony teaching obsolete. It changed the floor (students can check mechanical correctness automatically) and raised the ceiling (more class time can be spent on voice-leading logic and expressive intention). AI tools are a larger version of the same displacement: the floor rises, the ceiling rises with it, and the pedagogical question is always what you are doing between the two.

Copyright and Academic Integrity

Two issues that crossed all three workshops and deserve direct treatment.

On copyright: the training data of generative music models includes copyrighted recordings and scores, the legal status of which is actively litigated in multiple jurisdictions. When Suno.AI generates a piece “in the style of” a named composer, it is drawing on patterns extracted from that composer’s work — work that is under copyright in the case of living or recently deceased composers. The output is not a direct copy, but neither is the relationship to the training data legally settled. Music students who use these tools in professional contexts should know that they are working in a legally uncertain space, and institutions should not pretend otherwise.

On academic integrity: the issue is not that students might use AI to cheat — they will, some of them, and they have always found ways to cheat with whatever tools were available. The issue is that current AI policies at many institutions are incoherent: prohibiting AI use in assessment while providing no clear guidance on what counts as AI use, and assigning tasks where AI assistance is undetectable and arguably appropriate. The more useful approach is to design tasks where AI assistance is either irrelevant (because the task requires live performance or real-time demonstration) or visible and assessed (because the task explicitly includes reflection on how AI was used and to what effect).

Three Things I Came Away With

After a full day of workshops, discussions, and the conversations that happen in the corridors between sessions, I left with three positions that feel more settled than they did in the morning.

First: the data protection question is not separable from the pedagogical question. Any serious curriculum discussion of AI in music education has to start with what can legally be deployed, not with what would be useful if constraints were not a factor. The constraints are a factor.

Second: the skill most urgently needed — in students and in instructors — is not AI literacy in the sense of knowing which tool to use for which task. It is the critical capacity to evaluate AI-generated outputs: to notice what is wrong, to understand why it is wrong, and to correct it. This requires domain expertise first. You cannot critically evaluate an AI-generated harmonic analysis if you do not understand harmonic analysis. The tools do not lower the bar for domain knowledge. They raise the bar for its critical application.

Third: the curriculum question is not “how do we accommodate AI?” It is “what are we actually trying to teach, and does the answer change when AI can produce the visible output of that process?” Answering that honestly, skill by skill, for a full music programme, is slow work. It cannot be done at a one-day event. But a one-day event, if it is well-designed, can start the conversation in the right place.

HfMT’s Thementag started it in the right place.

References

• Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018). BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805. https://arxiv.org/abs/1810.04805

• Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press. https://www.deeplearningbook.org

• Holmes, W., Bialik, M., & Fadel, C. (2019). Artificial Intelligence in Education: Promises and Implications for Teaching and Learning. Center for Curriculum Redesign.

• Jobin, A., Ienca, M., & Vayena, E. (2019). The global landscape of AI ethics guidelines. Nature Machine Intelligence, 1, 389–399. https://doi.org/10.1038/s42256-019-0088-2

• LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521(7553), 436–444. https://doi.org/10.1038/nature14539

• Luckin, R., Holmes, W., Griffiths, M., & Forcier, L. B. (2016). Intelligence Unleashed: An Argument for AI in Education. Pearson.

• Mittelstadt, B. D., Allo, P., Taddeo, M., Wachter, S., & Floridi, L. (2016). The ethics of algorithms: Mapping the debate. Big Data & Society, 3(2). https://doi.org/10.1177/2053951716679679

• Roll, I., & Wylie, R. (2016). Evolution and revolution in artificial intelligence in education. International Journal of Artificial Intelligence in Education, 26(2), 582–599. https://doi.org/10.1007/s40593-016-0110-3

• Russell, S., & Norvig, P. (2020). Artificial Intelligence: A Modern Approach (4th ed.). Pearson.

• Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30. https://arxiv.org/abs/1706.03762