The AI Composer in the Mind

by

Author: Denis Avetisyan

New research suggests our opinions of artificial intelligence, not the music itself, largely shape how we perceive and enjoy AI-generated compositions.

Human perception of authenticity varies significantly by genre, with stimulus characteristics playing a crucial role in determining perceived humanness.
Human perception of authenticity varies significantly by genre, with stimulus characteristics playing a crucial role in determining perceived humanness.

Listener attitudes toward AI are the primary driver of music evaluation, influencing both liking and emotional perception, rather than inherent qualities of the composition or composer identity.

Despite growing interest in artificial intelligence’s creative potential, understanding how audiences perceive AI-generated art remains surprisingly complex. This study, ‘Perception of AI-Generated Music — The Role of Composer Identity, Personality Traits, Music Preferences, and Perceived Humanness’, investigates the factors shaping evaluations of music created by AI, employing a mixed-methods approach to disentangle the influence of composer information and listener characteristics. Crucially, findings reveal that pre-existing attitudes toward AI, rather than the perceived origin of the music, are the strongest predictor of both liking and emotional response. As AI increasingly blurs the lines of artistic creation, how will these perceptions evolve, and what ethical considerations will become paramount in evaluating music composed by machines?

The Erosion of Symbolic Precision in Musical Creation

Early attempts at computational music generation were largely constrained by their reliance on pre-programmed rules and symbolic representations. These systems, often categorized as ‘Symbolic Music Generation’, operated by manipulating musical notes and structures according to explicitly defined parameters – essentially, a digital mimicking of compositional techniques. While capable of producing technically correct pieces, the resulting music often lacked the subtle emotional depth and expressive qualities inherent in human composition. The rigidity of these rule-based approaches struggled to capture the nuances of phrasing, timbre, and the organic variations that give music its life, ultimately limiting their capacity to create truly compelling or emotionally resonant experiences for listeners. The music felt constructed, rather than composed, and highlighted the gap between algorithmic precision and artistic expression.

The landscape of musical creation is undergoing a dramatic shift, fueled by recent breakthroughs in text-to-music models like Stable Audio and Suno. These innovative systems transcend traditional methods of algorithmic composition, moving beyond rigid rule-based structures to generate original music directly from textual descriptions. A user can now input a simple prompt – “a hopeful piano ballad” or “driving electronic dance music” – and the model will synthesize an entire musical piece, complete with instrumentation and arrangement. This represents a significant leap forward, democratizing music production by removing the need for formal musical training or instrumental proficiency. The ability to translate natural language into audible compositions not only streamlines the creative process but also opens up exciting possibilities for personalized music experiences and novel forms of artistic expression.

The emergence of AI-generated music compels a re-evaluation of how humans perceive and value musical creations, moving beyond purely aesthetic judgment. Studies indicate that a listener’s pre-existing attitudes towards artificial intelligence significantly influence their emotional response and overall liking of algorithmically composed pieces. This suggests that appreciation isn’t solely based on the sonic qualities of the music itself, but is heavily colored by expectations and beliefs about the composer – or, in this case, the absence of a traditional human composer. Consequently, even objectively similar musical passages may be received very differently depending on whether the listener knows they were generated by an AI, highlighting a complex interplay between technological innovation and deeply ingrained human biases in artistic evaluation.

The Quantifiable Subjectivity of Musical Preference

Individual preferences for music are not solely determined by objective qualities of the composition but are significantly modulated by personality traits. Research indicates a correlation between scores on the “Openness to Experience” dimension of the Five Factor Model and a greater enjoyment of a wider range of musical genres, including complex and unconventional styles. Conversely, higher levels of “Neuroticism” have been associated with a preference for music with melancholic or emotionally intense qualities. Furthermore, a pre-existing level of “Music Appreciation”-reflecting a general interest in and positive attitude towards music-serves as a foundational element influencing overall liking, independent of specific genre or stylistic preferences.

Musical competence, reflecting a listener’s accumulated knowledge of musical structures and techniques, demonstrably shapes evaluative processes. Individuals with higher levels of musical training or experience tend to prioritize compositional elements such as harmonic complexity, counterpoint, and formal structure when assessing a piece. Conversely, listeners with lower musical competence often focus on more superficial characteristics like melody, rhythm, and timbre. This difference in focus leads to divergent evaluation patterns; highly competent listeners may appreciate nuanced or unconventional compositions that less experienced listeners find dissonant or inaccessible. Consequently, ratings of musical quality or enjoyment can vary significantly based on the evaluator’s level of musical expertise.

Emotional response is a key determinant in musical evaluation, and is quantifiable using instruments such as the Geneva Emotional Music Scale-9 (GEMS-9). This scale assesses nine basic emotions elicited by music – wonder, tenderness, nostalgia, peacefulness, joy, excitement, sadness, anxiety, and tension – providing a metric for the emotional impact of a given piece. Studies indicate a correlation between high scores on GEMS-9 and increased overall liking of music, though the relationship is complex and modulated by individual differences. Aggregate data from multiple studies reveals an average liking rating of 3.15 on a standardized scale, underscoring the multifaceted nature of musical preference and the significant contribution of emotional intensity to subjective evaluation.

Average liking scores varied significantly across experimental groups.
Average liking scores varied significantly across experimental groups.

The Attribution Bias: A Flaw in Perceptual Logic

Evaluations of musical compositions are demonstrably affected by attribution, a phenomenon termed ‘Composer Bias’. Statistical analysis of participant ratings reveals a significant difference in ‘Liking’ scores based on perceived authorship; music labeled as created by a ‘Soundtrack’ composer received an average rating of 3.15, compared to 2.87 for music labeled as AI-generated and 2.89 for unlabeled music. This difference is statistically significant at the p < .05 level, indicating that listeners do not evaluate music neutrally and that attribution to a human composer positively influences perceived enjoyment. This suggests a predisposition to favor music believed to be created by a person, irrespective of objective musical qualities.

Evaluations of music generated by artificial intelligence are subject to cognitive bias, wherein listeners implicitly utilize differing assessment criteria compared to those applied to human-composed music. This can result in systematically lower ratings for AI-generated pieces, not necessarily due to objective musical quality, but rather a subconscious perception of diminished artistic value. This phenomenon suggests that listeners may prioritize attributes traditionally associated with human creativity – such as originality, emotional depth, or technical skill – and apply these expectations more stringently when the source is identified as AI. Consequently, music created by algorithms may be judged based on its perceived technical execution rather than its artistic merit, potentially leading to an undervaluation of its creative qualities.

Evaluations of AI-generated music are significantly influenced by the listener’s perception of its human qualities. Data indicates a substantial difference in perceived humanness between music generated by different AI models; Suno-generated music received a Cohen’s d of 1.39 compared to Stable Audio-generated music, demonstrating a large effect size. This suggests that AI music exhibiting characteristics more closely associated with human composition is evaluated more favorably, potentially reducing the negative bias often observed when listeners are aware the music is not human-created. The degree to which an AI can mimic perceived human musical traits appears to directly correlate with its acceptance by listeners.

The Algorithmic Soundtrack: A Shift in Auditory Ecology

Recent qualitative studies, employing rigorous thematic analysis, demonstrate the remarkably varied roles music plays in daily life. Investigations reveal a spectrum of ‘functional uses’ extending far beyond mere entertainment; music serves as subtle background ambience in commercial spaces, a focal point for active listening during leisure, and a powerful tool in therapeutic settings for managing stress and promoting wellbeing. This research highlights how individuals consciously and unconsciously leverage music to shape mood, enhance focus, and navigate emotional states, suggesting a deeply ingrained and multifaceted relationship between humans and sound. The identified patterns emphasize that the way music is used is as significant as the music itself, informing expectations and influencing perceptual experiences.

The proliferation of artificial intelligence has extended into the realm of auditory experience, with AI-generated music finding increasing application across diverse contexts. Beyond simple replication, these systems now craft personalized soundtracks tailored to individual preferences, activities, or even biometric data, enhancing experiences ranging from exercise and gaming to focused work. More remarkably, adaptive compositions are emerging – musical pieces that dynamically alter based on listener interaction, environmental factors, or emotional cues. This responsiveness offers potential in therapeutic settings, where music can be adjusted to promote relaxation or alleviate stress, and in commercial spaces, where ambiance can be optimized to influence mood and behavior. The ability of AI to generate and modify music in real-time represents a significant shift, promising a future where auditory environments are not static, but rather fluid and intelligently designed.

The pervasive integration of AI-generated music into daily life necessitates a careful examination of its impact on how individuals perceive and experience sound. As algorithms increasingly curate and compose sonic environments, listener expectations are subtly reshaped, potentially introducing perceptual biases. For instance, consistently receiving music tailored to specific activities might diminish an individual’s ability to appreciate sonic novelty or complexity in other contexts. Responsible innovation, therefore, demands a proactive approach to understanding these shifts; researchers must investigate how algorithmic personalization influences aesthetic preferences, emotional responses, and even fundamental auditory processing. Failing to address these potential biases risks creating a musically homogenized landscape, where genuine artistic exploration is overshadowed by algorithmically-driven predictability and the nuances of human musical appreciation are lost.

The study’s findings regarding the primacy of listener attitudes over musical origin resonate with a fundamental tenet of mathematical elegance. It suggests that perception isn’t solely dictated by objective qualities-the ‘correctness’ of the musical composition-but by pre-existing biases, a kind of subjective ‘algorithm’ running in the listener’s mind. As Henri Poincaré stated, “Mathematics is the art of giving reasons.” This mirrors the study’s implication that evaluation isn’t a purely logical process, but rather, a reasoned justification after an attitude has shaped the initial perception. The emphasis on attitude as a primary driver highlights that the listener constructs a reality based on internal models, much like a mathematical proof built upon axioms, rather than an unbiased assessment of the music itself.

What’s Next?

The demonstrated primacy of listener attitudes toward artificial intelligence, rather than demonstrable qualities of the musical output itself, presents a peculiar, yet predictable, challenge. The field has, for some time, pursued technical perfection – algorithms that flawlessly mimic, or even surpass, human compositional skill. This work suggests such efforts may be largely orthogonal to genuine acceptance. A perfectly rendered imitation, indistinguishable from a human creation, will still be evaluated through a lens of pre-existing bias. The pursuit of ‘believability’ may be a fundamentally flawed objective.

Future research must address the source of this bias. Is it a simple category error – a refusal to accept creative agency in a non-human entity? Or does it stem from deeper cognitive mechanisms related to expectation and emotional response? A rigorous, theoretically-grounded exploration of these factors is essential, moving beyond mere correlation of personality traits and preferences. The creation of demonstrably ‘correct’ music, in a mathematical sense, remains a worthwhile endeavor, but its evaluation requires a parallel investigation into the human cognitive architecture that receives it.

Perhaps the most pressing task lies in developing metrics for quantifying this ‘AI attitude’ itself. Subjective assessments, while valuable, lack the precision demanded by scientific inquiry. A robust, quantifiable measure of pre-existing bias would allow for controlled experimentation, moving the field beyond descriptive observation and toward predictive modeling. Only then can the truly elegant solution – one that addresses the root cause of listener perception – be formulated.

Original article: https://arxiv.org/pdf/2512.02785.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2025-12-03 16:15