Artificial Intelligence in Music Pedagogy: Curriculum Implications from a Thementag
Abstract
On 2 December 2024 I gave three workshops at HfMT Köln’s Thementag on AI and music education. The handouts covered data protection, AI tools for students, and AI in teaching. This post is the argument behind them — focused on the curriculum question that none of the tools answer on their own: what should change, and what should not?
Contents
On 2 December 2024, the Hochschule für Musik und Tanz Köln held a Thementag (a theme day): “Next level? Künstliche Intelligenz und Musikpädagogik im Dialog” (“artificial intelligence and music pedagogy in dialogue”). I gave three workshops — on data protection and AI, on AI tools for students, and on AI in teaching. The handouts from those sessions cover the practical and regulatory ground. This post is the argument behind them: what I think changes in music education when these tools become ambient, and what I think does not.
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.
Transformers [[1]] are a dominant architecture for current language and many audio models, including Whisper [[2]] and the models behind ChatGPT; they are not a complete description of every tool named here or of Google Magenta’s varied research projects. A transformer is a neural network that processes a sequence of positions (tokens of text, frames of audio, musical notes) by computing, for each position, a weighted sum over the positions it is allowed to see. An encoder’s self-attention sees every position of the input, itself included. A decoder that generates output one step at a time sees itself and every earlier position; later ones are masked out. In an encoder–decoder model the decoder also attends across to the whole of the encoder’s output [[1]]. A language model that writes text token by token uses the masked form over its context window. Whisper is an encoder–decoder: its encoder reads a 30-second stretch of audio, and its decoder, “an audio-conditional language model”, attends to the transcript written so far and, across, to the encoder’s representation of the audio [[2]]. I previously wrote “all other elements”; that describes none of these. The weights are computed for each input from learned query, key and value projections. The result is a model that can capture long-range dependencies in sequences without recurrence, in which earlier networks passed information along one step at a time. That design made long-range dependencies hard to learn [[3]] [[4]] and training hard to parallelise [[1]].
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 [[5]] showed that a model pre-trained on large unlabelled text corpora and then fine-tuned for each task could advance the state of the art on eleven language tasks, “outperforming many task-specific architectures”. The broader move, training once on a very large corpus and reusing the result for many tasks [[6]], has spread to audio, to music generation and to multimodal domains. Whisper was trained on 680,000 hours of audio paired with transcripts found on the internet, and it transfers to standard benchmarks without any dataset-specific fine-tuning [[2]]; Magenta’s Music Transformer carries the architecture over to music [[7]]. 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 (Regulation (EU) 2016/679; in German the Datenschutz-Grundverordnung, DSGVO) [[8]] and, at state level in NRW, by the Datenschutzgesetz Nordrhein-Westfalen (DSG NRW, the state data protection act). The constraints are not abstract: they determine which tools can be used for which purposes with which students.
The core principle is data minimisation: personal data must be “limited to what is necessary in relation to the purposes for which they are processed”, and those purposes must be specified and explicit (Art. 5(1)(b)–(c)) [[8]]. 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 outside the EU. Whether the GDPR’s transfer rules apply depends on who passes the data on. A student using an account of their own is not an “exporter” in the guidance of the European Data Protection Board (EDPB), so Chapter V does not apply. When the university requires or provides the tool and the vendor processes the data on its behalf, Chapter V does apply, and passing the data to a vendor in a third country needs a basis in Arts. 44–49 [[9]]. Schrems II struck down the EU–US Privacy Shield and held that transfers under Article 46, such as standard contractual clauses, must be checked against the destination country’s law, public authorities’ access to the data included, for a level of protection “essentially equivalent” to the EU’s [[10]]. Since 10 July 2023 a new adequacy decision under Article 45, the EU–US Data Privacy Framework, has covered transfers to US organisations on the Department of Commerce’s Data Privacy Framework List [[11]]. I previously wrote that every such transfer needs a supplementary assessment; that holds for transfers under Article 46, not for one under an adequacy decision. The General Court dismissed an action to annul the framework on 3 September 2025 [[12]], and an appeal followed [[13]]; the Publications Office’s database held no judgment on the appeal when I checked it on 28 October 2026. Either way, legality depends on the facts, safeguards and contract, not on server location alone.
The practical upshot for curriculum design is that a university needs a documented purpose, lawful basis, processor/transfer assessment and accessible alternative before requiring a tool (Fig. 1). For a public university the lawful basis is usually a task carried out in the public interest (Art. 6(1)(e)), which must itself rest on Union or Member State law (Art. 6(3)); consent fits badly with anything compulsory where the controller is a public authority (recital 43). When the university decides that a tool is used and for what, it is the controller and the vendor processes data on its behalf, under a contract that must say, among other things, that the vendor acts “only on documented instructions from the controller” (Art. 28(3)). Whoever determines the purposes and means of processing is a controller (Art. 4(7)), and a processor that does so in breach of the Regulation is considered one (Art. 28(10)); a vendor that trains its models on students’ prompts for its own purposes has left the processor’s role for that use [[8]]. EU hosting can help but is not sufficient; US processing is not automatically unlawful, but cannot be treated as administratively irrelevant. Students should not have to surrender unnecessary personal data to complete assessed work.
The GDPR is not the only instrument. The EU’s AI Act, Regulation (EU) 2024/1689, entered into force on 1 August 2024, four months before the Thementag, with its obligations phased in later [[14]]. Two provisions that matter here have applied since 2 February 2025. The prohibitions in Article 5 include the use of AI systems “to infer emotions of a natural person in the areas of workplace and education institutions”, except for medical or safety reasons. Article 4 required providers and deployers (a university using an AI system under its own authority is a deployer, Art. 3(4)) to take measures to ensure, “to their best extent, a sufficient level of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf”. An amending regulation in force since 27 July 2026 turned that into a duty to “support the development of AI literacy”, which “does not require providers or deployers to guarantee any specific level of AI literacy of any individual” [[15]]. AI systems for certain educational uses, including admission decisions, the evaluation of learning outcomes and the detection of prohibited behaviour during tests, are high-risk (Annex III, point 3) unless, under Article 6(3), they pose no significant risk of harm; the amendment moved the date from which those obligations apply from 2 August 2026 to 2 December 2027 [[14]] [[15]].
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. In Holmes et al.’s account [[16]], which I have not been able to check, feedback generation is 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) [[17]] and Mittelstadt et al. (2016) [[18]] both document the broader concern with AI opacity and accountability. For a teacher the consequence is practical: 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 are trained and evaluated. Kalai et al. [[19]], writing from OpenAI, argue that where incorrect statements cannot be told apart from facts, pre-training produces them “through natural statistical pressures”, and that benchmarks which reward a guess over an admission of uncertainty keep them alive. A model trained to predict likely next tokens produces a statistically plausible continuation, not a checked 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 a common promise of AI in education (Luckin et al., 2016 [[20]]; Roll & Wylie, 2016 [[21]], neither of which I could check in full) 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. In my reading, 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 can generate a stylistically convincing chord progression, as part of a complete song, in the manner of a named composer, although its help article on moderation, last edited 19 December 2025, says that songs “may not generate” if the prompt includes “the names of well known artists or people” [[22]]; what it accepted in December 2024 is not recorded here. 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, in my judgement, locally coherent; its developers describe minute-long pieces “with compelling structure” [[7]]. 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; in June 2024, for instance, Universal, Sony and Warner labels sued Suno in the US District Court for the District of Massachusetts [[23]]. Where Suno produces 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 for living composers and, in the EU, for 70 years after the composer’s death [[24]]. The output is usually not a direct copy, but neither is the relationship to the training data legally settled. Since I wrote this, a German court has decided a narrower question, memorisation. On 31 July 2026 the Landgericht München I (Munich Regional Court I) found largely (überwiegend) for the collecting society GEMA. It held that particular songs, whose original elements reappeared when Suno was prompted with their lyrics, title and a style, were contained in its models; that the training copies made in the US were not covered by fair use; and that Suno, not its users, was responsible for the outputs. The judgment is not final [[25]]. 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.
Literature, tool descriptions and EU law checked through 2026-10-28; the GDPR article numbers in Fig. 1 against the Official Journal text. Not opened: Holmes et al. (2019), Luckin et al. (2016) and the DSG NRW. Read in part: the abstracts of Mittelstadt et al. (2016) and Roll & Wylie (2016), and only the table of contents of Russell & Norvig (2020). The claims resting on these sources are the ones this post made before the check.
References
[1] 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. Sections 1–2 (parallelisation) and 3.2 (attention). arXiv:1706.03762. https://arxiv.org/abs/1706.03762
[2] Radford, A., Kim, J. W., Xu, T., Brockman, G., McLeavey, C., & Sutskever, I. (2022). Robust speech recognition via large-scale weak supervision. Preprint, arXiv:2212.04356. https://arxiv.org/abs/2212.04356
[3] LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521(7553), 436–444. DOI: 10.1038/nature14539
[4] Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press. Section 10.7, “The Challenge of Long-Term Dependencies”. https://www.deeplearningbook.org
[5] Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Vol. 1, pp. 4171–4186. DOI: 10.18653/v1/N19-1423. Preprint (2018): arXiv:1810.04805.
[6] Russell, S., & Norvig, P. (2020). Artificial Intelligence: A Modern Approach (4th ed.). Pearson. Section 24.5, “Pretraining and Transfer Learning” (checked against the table of contents only). https://aima.cs.berkeley.edu
[7] Huang, C.-Z. A., Vaswani, A., Uszkoreit, J., Shazeer, N., Simon, I., Hawthorne, C., Dai, A. M., Hoffman, M. D., Dinculescu, M., & Eck, D. (2018). Music Transformer: Generating music with long-term structure. Preprint, arXiv:1809.04281 (v3, 12 December 2018). https://arxiv.org/abs/1809.04281
[8] Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data (General Data Protection Regulation). Official Journal of the European Union, L 119, 4 May 2016, p. 1. Recital 43; Arts. 4(7)–(8), 5, 6, 28, 44–46. http://data.europa.eu/eli/reg/2016/679/oj
[9] European Data Protection Board. (2023, 14 February). Guidelines 05/2021 on the Interplay between the application of Article 3 and the provisions on international transfers as per Chapter V of the GDPR, version 2.0. Section 2.2 and Examples 2 and 5. Guidance, not binding law. https://www.edpb.europa.eu/system/files/2023-02/edpb_guidelines_05-2021_interplay_between_the_application_of_art3-chapter_v_of_the_gdpr_v2_en_0.pdf
[10] Court of Justice of the European Union (Grand Chamber). (2020, 16 July). Data Protection Commissioner v Facebook Ireland Ltd and Maximillian Schrems (Schrems II), Case C-311/18, ECLI:EU:C:2020:559. Operative part, points 2 and 5. http://publications.europa.eu/resource/celex/62018CJ0311
[11] Commission Implementing Decision (EU) 2023/1795 of 10 July 2023 pursuant to Regulation (EU) 2016/679 on the adequate level of protection of personal data under the EU-US Data Privacy Framework. Official Journal of the European Union, L 231, 20 September 2023, p. 118. Article 1. http://publications.europa.eu/resource/celex/32023D1795
[12] General Court (Tenth Chamber, Extended Composition). (2025, 3 September). Latombe v Commission, Case T-553/23, ECLI:EU:T:2025:831. http://publications.europa.eu/resource/celex/62023TJ0553
[13] Appeal brought on 31 October 2025 by Philippe Latombe against the judgment of the General Court in Case T-553/23 (Case C-703/25 P). Official Journal of the European Union, C/2025/6610, 22 December 2025. http://publications.europa.eu/resource/celex/62025CN0703
[14] Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union, L, 2024/1689, 12 July 2024. Arts. 3(4), 4, 5(1)(f), 6(3), 113; Annex III, point 3. http://data.europa.eu/eli/reg/2024/1689/oj
[15] Regulation (EU) 2026/1744 of the European Parliament and of the Council of 8 July 2026 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 as regards the simplification of the implementation of harmonised rules on artificial intelligence (Digital Omnibus on AI). Official Journal of the European Union, L, 2026/1744, 24 July 2026. Article 1, points (5) and (40); Article 4. http://data.europa.eu/eli/reg/2026/1744/oj
[16] Holmes, W., Bialik, M., & Fadel, C. (2019). Artificial Intelligence in Education: Promises and Implications for Teaching and Learning. Center for Curriculum Redesign. Not opened.
[17] Jobin, A., Ienca, M., & Vayena, E. (2019). The global landscape of AI ethics guidelines. Nature Machine Intelligence, 1(9), 389–399. DOI: 10.1038/s42256-019-0088-2. Preprint: arXiv:1906.11668.
[18] Mittelstadt, B. D., Allo, P., Taddeo, M., Wachter, S., & Floridi, L. (2016). The ethics of algorithms: Mapping the debate. Big Data & Society, 3(2), 2053951716679679. DOI: 10.1177/2053951716679679
[19] Kalai, A. T., Nachum, O., Vempala, S. S., & Zhang, E. (2025). Why language models hallucinate. Preprint, arXiv:2509.04664 (OpenAI and Georgia Tech). https://arxiv.org/abs/2509.04664
[20] Luckin, R., Holmes, W., Griffiths, M., & Forcier, L. B. (2016). Intelligence Unleashed: An Argument for AI in Education. Pearson. Not opened.
[21] Roll, I., & Wylie, R. (2016). Evolution and revolution in artificial intelligence in education. International Journal of Artificial Intelligence in Education, 26(2), 582–599. DOI: 10.1007/s40593-016-0110-3
[22] Suno. Does Suno moderate songs? Suno Knowledge Base, last edited 19 December 2025, accessed 28 October 2026. Vendor documentation. https://help.suno.com/en/articles/3198209
[23] UMG Recordings, Inc. et al. v. Suno, Inc., No. 1:24-cv-11611 (D. Mass.). Complaint filed 24 June 2024. https://www.riaa.com/wp-content/uploads/2024/06/Suno-complaint-file-stamped20.pdf
[24] Directive 2006/116/EC of the European Parliament and of the Council of 12 December 2006 on the term of protection of copyright and certain related rights (codified version). Official Journal of the European Union, L 372, 27 December 2006, p. 12. Article 1(1). http://publications.europa.eu/resource/celex/32006L0116
[25] Landgericht München I. (2026, 31 July). Urteil GEMA gegen SUNO. Press release 16/2026 on the judgment in case 42 O 763/25. https://www.justiz.bayern.de/gerichte-und-behoerden/landgericht/muenchen-1/presse/2026/16.php
Changelog
- 2026-07-11: Narrowed the technology claims and clarified that GDPR compliance depends on the full processing and transfer arrangement rather than server location alone; added the current EU AI Act implementation context to the source review.
- 2026-10-28: Corrected the description of attention from “a weighted attention over all other elements” to the three uses in Vaswani et al. (section 3.2.3): encoder self-attention over every input position, itself included; masked decoder self-attention over the position itself and the earlier ones; and decoder attention across to the encoder’s output. The weights are computed from learned projections, not learned themselves. Corrected the transfer passage, which said every transfer outside the EU needs a supplementary assessment: after Schrems II that applies to transfers under Article 46, such as standard contractual clauses, not to transfers under an adequacy decision, and since 10 July 2023 the EU–US Data Privacy Framework is one for listed US organisations (the General Court dismissed an annulment action on 3 September 2025; an appeal was lodged). Added, following the EDPB’s guidance, that the transfer rules apply when the university requires or provides the tool, not when a student uses an account of their own. Anchored the data-protection section in the GDPR (Arts. 4–6, 28, 44–46; recital 43) and added Fig. 1. The 2026-07-11 entry put the AI Act context into the research note, not the body; the body now says that the Act has been in force since 1 August 2024, that its Article 5 prohibitions (including emotion inference in education institutions) and Article 4 on AI literacy have applied since 2 February 2025, that Regulation (EU) 2026/1744 weakened Article 4 from 27 July 2026, and that it moved the high-risk date for education systems under Annex III from 2 August 2026 to 2 December 2027. Replaced “recently deceased” composers with the EU term of 70 years after death, and “The output is not a direct copy” with “usually not”. Added the 2024 US label suit against Suno and the Munich Regional Court’s judgment of 31 July 2026, largely for GEMA against Suno, on the memorisation of particular songs (not final). Added that Suno’s help article on moderation, last edited in December 2025, says songs naming well-known artists “may not generate”, and that what Suno accepted in December 2024 is not recorded. Described Whisper as used without dataset-specific fine-tuning and quoted BERT’s own claim (“outperforming many task-specific architectures”). Changed “the most statistically plausible continuation” to “a statistically plausible continuation, not a checked one”, citing Kalai et al. (2025). Marked the Holmes et al. claim as unchecked and the mixed evidence on personalisation as my reading, and changed “the most-cited promise” to “a common promise”. Cited LeCun et al., Goodfellow et al. and Russell & Norvig, previously listed but never cited, for the sentences they support. The verification line now names the sources that could not be opened or were read only in part. Converted the author–year list to numbered, linked citations.