It sounds like someone imagining us

“It sounds like someone imagining us.” That’s exactly what one of my students said after listening to AI-generated traditional music during my World Music lecture. The sentence stayed with me because it captured the problem so precisely. The music was polished. It contained familiar cultural signals. Yet something was missing.

So as a class, we wanted to understand what these systems recognise when we insert a prompt into AI music platforms about an instrument, ensemble or musical tradition. Do they understand the music, or do they simply assemble sounds that listeners associate with a region?

To explore this, my students used two platforms to generate the Chinese guqin and Indonesian gamelan. They tested each example twice. First with a name-only prompt and then with a more musically informed prompt. Across the two platforms, this produced eight audio outputs.

This was not a scientific benchmark. It was a listening exercise. We wanted to compare what the platforms produced with the musical features that give each tradition its identity.

It sounds Chinese but is that enough?

The guqin is a seven-stringed Chinese zither with a history of more than 3,000 years. Traditionally practised by scholars in intimate settings, it was one of four arts alongside calligraphy, painting and chess that an educated Chinese scholar was expected to master. According to tradition, proficiency could require approximately twenty years of training. UNESCO. The instrument’s identity lies not simply in its plucked sound, but in its characteristic register, harmonics, sliding tones, subtle changes in pressure and use of silence. 

The two prompts tested whether the platforms could recognise the guqin by name alone, and whether describing its defining musical features would produce a more convincing output.

Name-only prompt:

A solo performance of traditional Chinese guqin music.

Musically informed prompt:

Traditional Chinese guqin solo. A seven-stringed fretless plucked zither. Use a low-to-middle register, a quiet wooden plucked tone, open strings, clear harmonics and gently sliding stopped notes. Keep the rhythm free, the melody mostly monophonic and the phrasing slow. No accompaniment

Platform 1: Close enough to recognise the category

© Screenshot

Guqin Platform 1 - Name only prompt

Students felt that Platform 1 recognised that the guqin was a Chinese plucked zither with just the name of the instrument and its country of origin. However, they described the output as bright and metallic rather than quiet and wooden. Its register was too high, the bends were excessive, and an erhu-like instrument appeared midway despite the request for a solo performance. The pentatonic melody created a familiar “Chinese” sound, but this regional marker could not compensate for the inaccurate instrumental timbre. What emerged was a generic digital zither rather than a convincing guqin.

The platform’s expansion of the students’ deliberately simple prompt also revealed an important limitation. It added phrases such as “measured pauses that let each note bloom” and “mid-piece circles through spacious motifs.” These descriptions sounded expressive but offered little clear information about the instrument’s actual timbre, register or playing techniques. The added language created an impression of musical knowledge, but the inaccurate audio showed that it did not translate into a convincing understanding of how the guqin should sound.

© screenshot

Guqin Platform 1 - Musically informed prompt

Students identified several improvements in the second output. They noted that the audio began within the guqin’s register and contained clearer harmonics, sliding tones, pauses and ornamentation. However, they felt that some of the spacing sounded natural while other pauses appeared artificially inserted. They also observed that although the recording began as a solo plucked instrument, it eventually introduced an erhu-like sound, repeating the problem heard in the name-only output. 

The platform also expanded the musically informed prompt with descriptions such as “slow spacious phrases with natural decay and long silences,” “subtle breathing room between lines,” and an “intimate dry close-mic mix” that was “warm and organic” and focused on “wood and air.” Much of this restated the original instructions in more expressive but less precise language. It also added the word “bright,” which did not match the requested quiet wooden tone. As with the name-only prompt, the added description created an impression of musical detail without it being reflected in the audio. Overall, students found that the platform followed several individual instructions but could not sustain either the identity of the guqin or the requested solo texture throughout the piece.

Platform 2: The instrument disappeared

© screenshot

Guqin Platform 2 - Name only prompt

Unlike Platform 1, Platform 2 did not expand the name-only prompt with elaborate or unnecessary descriptions. Its description simply restated the request as “a solo performance of traditional Chinese guqin music.” However, the audio did not follow this description accurately. 

The name-only prompt in this platform produced a bowed string instrument that sounded closer to an erhu than a guqin. Without the words “plucked” or “zither”, the output contained neither the attack nor the timbre of a plucked stringed instrument. The additional accompaniment moved it even further from the requested solo performance. Students noted that the platform did, however, appear to respond to the word “Chinese” by generating a pentatonic melody and using broadly familiar markers of Chinese music. Unlike Platform 1, which at least attempted to imitate a Chinese plucked instrument, Platform 2 recognised the regional label but not the guqin itself.

© screenshot

Guqin Platform 2 - Musically informed prompt

For the musically informed prompt, Platform 2 added several descriptions of its own. It described the music as having a “sparse texture,” “flexible rhythm,” “subtle changes in timbre,” and an “intimate” sound, with “space and silence between phrases.” Some of these phrases interpreted or reworded the students’ instructions: “flexible rhythm” reflected the request for free rhythm, while “sparse texture” drew on the mostly monophonic melody and absence of accompaniment. Other additions, particularly “subtle changes in timbre” and “intimate,” were not stated in the original prompt. Unlike Platform 1’s more elaborate language, Platform 2’s additions remained relatively concise and closely connected to the requested musical qualities.

However, the resulting audio produced a plucked-zither sound for approximately five seconds, suggesting that the platform could respond briefly to words such as “plucked” and “zither,” but it could not sustain those features as a coherent instrumental identity. The platform then introduced several synthesised instruments. Some of my Chinese students associated its layered synthesisers, sustained string-like textures and sentimental melody with 1990s Cantopop and the background music of Hong Kong films from that period. It was striking that these Gen Z students recognised the style so readily, even though 1990s Cantopop belonged to an earlier generation. When I asked how they were so certain, they explained that they had heard it in films watched by their parents. This was an interesting observation, because their response showed that musical recognition is shaped not only by what we choose to hear, but also by the sounds that circulate around us over time.

What students heard across the four recordings

Across the four recordings, students identified different levels of musical recognition. Platform 1 appeared to recognise the guqin as a type of Chinese plucked zither, but not as an instrument with a distinct acoustic identity. The informed prompt improved individual features such as register, harmonics, sliding tones and pacing, yet these remained isolated details rather than a coherent guqin performance. 

Platform 2 performed more poorly. Students felt that Platform 2 responded more strongly to the word “Chinese” than to the word “guqin.” There were broad markers of Chinese music, particularly pentatonic melodies and cinematic textures, but little understanding of the timbre and acoustic characteristics of guqin itself. Even the musically informed prompt produced only a brief plucked sound before reverting to synthesised accompaniment. Students concluded that Platform 1 offered an inaccurate approximation of the instrument, while Platform 2 largely substituted a generalised idea of Chinese music for it. 

Several factors may have contributed to these failures. The platforms may have encountered too few clearly labelled guqin recordings, or the available recordings may not have been supported by accurate titles, tags and metadata. They may also prioritise producing a polished, complete composition over maintaining the identity of an unfamiliar instrument. The comparison shows that a detailed prompt can refine features a platform already partly understands, but it cannot supply musical knowledge that the model has not learned. The comparison therefore reveals an important difference between recognising a cultural category and understanding the specific instrument named in the prompt. 

A gong does not make it gamelan

Gamelan presented a different challenge. Unlike the solo guqin, gamelan is an ensemble built around metallophones, gong-chimes, hanging gongs and drums. Instruments like the saron, bonang, kenong, kempul and gong have different roles within a layered musical structure. In Javanese gamelan, the instruments are organised according to two principal tuning systems, or laras, known as sléndro and pélog. Their precise pitches vary between gamelan sets and do not conform to Western equal temperament. However, this does not mean that gamelan has no tuning rules or that any collection of pitches can sound correct. The intervals remain organised within a recognisable system. 

The test, therefore, was not simply whether the platforms could generate metallic or gong-like sounds. It was whether they could reproduce the tuning, instrumental roles and layered relationships that make those sounds recognisably gamelan. The name-only prompt tested what each platform understood from the word “gamelan” itself. The musically informed prompt then supplied the main instruments and structural features to determine whether clearer guidance would produce a more convincing ensemble. 

Name-only prompt:

An instrumental performance of traditional Indonesian gamelan music.

Musically informed prompt:

Traditional Javanese gamelan performed as an acoustic instrumental ensemble of metallophones and gongs. Feature saron, bonang, kempul , gambang and gendang as part of the instruments in the ensemble. Use interlocking and repeating melodic patterns, cyclical gong punctuation and the distinctive tuning and ensemble texture of Indonesian gamelan.

Platform 1: The sound of metal without the system

© screenshot

Gamelan Platform 1 - Name only prompt

Platform 1 recognised the most obvious surface feature, which was the metallic timbre of a gamelan ensemble. However, approximately eighteen seconds into the recording, a regular marching beat and a metallic-sounding plucked instrument entered, neither of which reflected gamelan’s timbre, interlocking rhythms or cyclical structure. The platform also added orchestral textures and other non-gamelan sounds, making the output resemble cinematic music rather than a true gamelan ensemble performance. 

 As with the guqin prompts, Platform 1 expanded the students’ simple name-only prompt with a much more detailed description. It referred to “steady interlocking gong and metallophone patterns,” “layered kendang pulses” and “shimmering bonang figures.” It also described “verse-like opening cycles,” “cascading patterns,” and “bronze overtones.” Interestingly though, some of these additions accurately described features of a gamelan ensemble. However the audio did not consistently reproduce the interlocking patterns, gong punctuation or ensemble relationships that the platform itself had described.

Overall, students felt the platform reproduced the sound of struck metal without reproducing the rhythmic and structural relationships that make it recognisably gamelan. Once again, the platform’s written description suggested more musical knowledge than could be heard in the generated audio.

© screenshot

Gamelan Platform 1 - Musically informed prompt

This audio included interlocking patterns, suggesting that Platform 1 responded to one of the structural features specified in the prompt. However, the pitches assigned to the metallophone-like sounds sometimes extended beyond the tonal register of the gamelan instruments they appeared to imitate. More importantly, the pitch relationships did not convincingly belong to either sléndro or pelog indicating that the platform was unfamiliar with this tuning system. Similar to the name-only prompt, this audio also introduced orchestral instruments and textures that were not characteristic of a gamelan ensemble. 

The platform again expanded the prompt with its own descriptions. It referred to a “driving kendang pulse,” “cyclical gong punctuation” and a “steady ceremonial tempo.” It also claimed that the middle section layered rebab (bowed string instrument) and suling (wind instrument) ornaments over the main pulse, while the ending introduced stronger gong accents and “brighter cascades.” Neither the rebab nor the suling appeared in the students’ original prompt, which specifically named the saron, bonang, kempul, gambang and gendang. The platform therefore did more than rephrase the supplied musical information: it introduced additional instruments and imagined how they would function within the arrangement. Other phrases, such as “soft room bloom,” “subtle bamboo texture” and “polished bronze shimmer,” sounded expressive but gave little precise information about the organisation of the music. More importantly, several features named in the description were not clearly sustained in the audio.

Although interlocking patterns were present, the cyclical gong punctuation and the different roles of the instruments did not become clearer. As one student observed, it was “a different arrangement of the same approximation.” The longer prompt changed the surface organisation of the music, but it did not produce a more convincing gamelan ensemble. As with the previous examples, the platform could describe several features associated with the tradition more convincingly than it could reproduce their relationships in sound.

Platform 2: What is a gamelan?

© screenshot

Gamelan Platform 2 - Name only prompt

Platform 2 showed no audible evidence that it understood the word “gamelan.” Instead of producing an ensemble of metallophones, gongs and drums, it generated a pop composition that broadly resembled Malay or Indonesian popular music. The recording relied on Western and electronic instrumentation, including saxophone-like sounds, a drum kit, strings and synthesisers. The platform therefore appeared to respond to the regional label “Indonesian” rather than to gamelan as a specific ensemble tradition. Unlike Platform 1, which at least reproduced gamelan’s metallic tone colour, Platform 2 produced no recognisable gamelan instruments, tuning or ensemble structure.

© screenshot

Gamelan Platform 2 - Musically informed prompt

After hearing the name-only output, students already expected the musically informed prompt to produce a similar result. Since Platform 2 had shown no audible evidence that it recognised the word “gamelan,” they doubted that additional descriptions of its instruments and musical structure would correct the problem. Unlike Platform 1, Platform 2 did not expand the prompt with new instruments or expressive descriptions. Its summary remained close to the information provided by the students, referring to the bonang, saron, gambang, gong and gendang, as well as interlocking patterns and cyclical gong punctuation. The contrast became clear in the resulting audio, because while the platform could repeat the relevant information in writing, it could not reproduce it in sound.

Their expectations were confirmed. Despite receiving specific information about the instruments, tuning and the ensemble’s structural features, the platform generated another synthesised pop composition.  None of the requested gamelan instruments or structural features could be clearly heard. Most strikingly, the platform titled the piece “Jalinan Bonang,” even though no recognisable bonang appeared in the music. The word entered the title but not the sound. The longer prompt influenced how the platform labelled and described the output, but it did not improve its musical understanding of gamelan.

What students heard across the four recordings

The students’ observations raised a larger question: why did both platforms struggle with gamelan when it is one of the most internationally recognised Asian ensemble traditions? Its global visibility does not necessarily mean that the platforms encountered enough accurately labelled recordings during training. Gamelan may be widely performed in concerts and universities outside Asia, but this is different from being well represented in the datasets used to train commercial music platforms. Research has found that non-Western music accounts for only a small proportion of existing music-generation datasets. MBZUAI

The recordings also suggest that the platforms may have learned gamelan as a general sound category rather than as an organised musical system. Metallic timbres are relatively easy to identify, but gamelan’s identity depends on relationships across an entire ensemble: instruments tuned together as a set, distinct musical roles, layered patterns and gong cycles unfolding over time. These relationships are more difficult to capture than a single instrument or timbre. The precise tuning also varies between gamelan sets, which may make it harder for a platform to learn a single acoustic template without detailed information about the tradition, tuning and ensemble represented in each recording. 

Labelling presents another possible problem. Recordings identified broadly as “gamelan” may include Javanese and Balinese traditions, contemporary compositions, fusion arrangements and music using only selected gamelan instruments. If these differences are not accurately described in the accompanying metadata, the platform may connect the word with scattered features rather than with a particular musical system. The students heard this fragmentation across the four recordings: metallic timbres and interlocking patterns appeared as isolated features, while Western pop, orchestral and cinematic textures filled the remaining musical space. This suggests that the platforms had learned partial associations with gamelan but not the relationships that organise those features into a coherent ensemble tradition.

Conclusion

Across both the guqin and gamelan recordings, students concluded that cultural visibility does not guarantee accurate machine representation. The informed prompts could name the defining features of each tradition, but they could not supply musical knowledge that the platforms had not already learned. For the guqin, the platforms reproduced broad markers of Chinese music without sustaining the timbre, register and playing gestures that distinguish the instrument. For gamelan, they recognised metallic sounds and cultural labels more readily than the tuning, instrumental roles and colotomic cyclical relationship within the ensemble. 

The gap was therefore not simply between an accurate and an inaccurate sound. It was between cultural recognition and musical understanding. The platforms could reproduce features that signalled “Chinese” or “Indonesian” to a general listener, but they did not consistently preserve the relationships that gave those features meaning within each tradition. Pentatonic melodies and cinematic textures stood in for the guqin, while metallic timbres and regional labels stood in for gamelan. In both cases, complex musical practices were reduced to familiar cultural signs.

This reduction also removes the community knowledge carried by the music. The guqin is shaped by an intimate tradition of listening, reflection and individual expression, while gamelan depends on collective coordination and instruments built, tuned and performed as an ensemble. When AI reproduces only their most recognisable sounds, it separates the music from the values, relationships and practices through which communities sustain it. 

The bias therefore affects more than musical accuracy. It determines which parts of a tradition become audible and which disappear. If these approximations circulate widely, they may gradually shape how unfamiliar listeners imagine the tradition, allowing an AI-generated stereotype to compete with the community’s own musical representation. The eight recordings suggest that a tradition can be visible enough for AI to imitate its surface, yet remain insufficiently represented for its cultural and musical meaning to survive the process.