Creative Homogenisation: When AI Makes Us All Sound the Same

Gist: Generative AI creates a creativity paradox: it makes individual work better and faster, but it also causes a homogenisation of collective output, pushing everyone toward the same average, statistical patterns. This loss of diversity and unique cultural input is a serious risk, as real innovation happens when different ideas clash, a process AI tends to smooth over. To avoid this algorithmic standardisation, we need to teach people to use AI critically and value the difficult, messy human effort that truly leads to novel ideas.

Share

Picture two seminars. In the first, students write short pieces without assistance. Some struggle with plot structure, others produce clumsy prose, but each piece feels distinctly different, messy. One leans into theory, another attempts practical guide, a third experiments with data storytelling. In the second workshop, students use GPTs to generate ideas and refine their work. The individual stories improve dramatically: better pacing, cleaner sentences, more coherent plots. But when you read all ten together, something uncanny emerges. They feel like variations on a theme rather than genuinely different creative visions.

This isn’t hypothetical. When researchers (Kosmyna et al., 2025) at Harvard and MIT studied creative writing with AI assistance which focused on finding out the cognitive cost of using an LLM in the educational context of writing an essay. They found that individual stories became more creative, better written, and more enjoyable to read. But all the AI-assisted stories became significantly more similar to each other than stories written by humans alone. We’re individually better off, collectively producing a narrower range of novel content.

This is the creativity paradox at the heart of generative AI: it amplifies individual creative capacity whilst homogenising collective creative output.

According Zhou & Less (2024) AI assistance ‘enhances human creative productivity’ by approximately 25% but decreases collective diversity, consistent with concerns about cultural homogenisation at scale. They analysed of over 53,000 artists and 5,800 known AI adopters for their study. At the same time creative productivity does not necessarily mean creativity. Zhou & Lee define this as ‘creative productivity as the log of the number of artifacts that a user posts in a month’ This is output linked to ‘automating the execution stage of the creative process’. For me this does not seem to be creativity but the generation or automation of digital creative products.

Why does this matter? Because creativity isn’t just an individual attribute, it’s a social phenomenon. Innovation develop from diverse perspectives colliding, from unexpected combinations, from people bringing different cultural frameworks and lived experiences. When algorithms mediate our imagination, trained on patterns extracted from existing work, they subtly push us toward convergence. We’re optimising for individual productivity whilst inadvertently narrowing the collective space of creative possibility.

The rest of this post examines what we’re maybe losing as generative AI becomes ubiquitous across creative domains. What insights from embodied creative practices, computational social science, and computing education offer for preserving genuine novelty.

To understand why AI assistance produces creative homogenisation (Doshi & Hauser, 2024; Kosmyna et al., 2025) we need to understand how generative models work. These systems are trained on vast corpus of existing creative work-millions of stories, videos, images, musical compositions. AI learn statistical patterns: which words typically follow which others, which visual elements commonly appear together, which chord progressions feel ‘natural.’ When you prompt AI, they’re essentially performing sophisticated pattern matching and interpolation within learned distributions.

A 2025 study challenged the tradeoff by using diverse GenAI ‘personas’ with different cultural backgrounds and thinking styles. While plot ideas from any individual persona showed high similarity (average cosine similarity of 0.92), ideas across different personas exhibited substantial variation (average similarity of 0.20) Boone et al., 2023

The above research findings align showing that the homogenisation can be counteracted through deliberately engineered diversity in AI systems. which supports the point that the convergence is real and happens when creators use the same AI tools trained on similar patterns.

This is simultaneously their strength and limitation. Generative AI excels at producing outputs that feel competent, polished, and ‘correct’ because they’re optimised to match patterns in training data. But genuine creative novelty often comes from breaking patterns, from combinations that feel unexpected or even ‘wrong’ initially.

The collective diversity problem: Research shows that when many creators use the same AI assistance, their outputs converge toward similar solutions because they’re all drawing from the same learned patterns. Individual stories might be more creative than what the writer would produce alone, but the distribution of creative approaches narrows. We lose the outliers, the weird experiments, the culturally specific approaches that don’t match dominant training data patterns.

Think of it as creative regression toward the mean. If you train a model on thousands of successful novels, it learns what ‘successful’ looks like statistically, certain narrative structures, character archetypes, pacing rhythms. When creators use AI for ideation or refinement, they’re subtly steered toward those statistically average patterns. The result is individually competent but collectively homogeneous.

Measuring what we’re losing: Computational social science methods (semantic distance measures, network analysis, cultural analytics…) allow researchers to quantify/qualify creative diversity at scales impossible through individual observation Hoffman, 2021’s Feminist Data Ethics and Drage & Fabretti (2024) AI that Matters provides interesting feminist lenses in which to view algorithmic mediate human imagination. When researchers analysed the semantic space occupied by AI-assisted versus human-only creative writing, they found AI-assisted work clustered more tightly. The stories covered a narrower conceptual territory despite individual quality improvements.

This is technological colonialism playing out in creative fields. Warning us that these systems risk homogenising global content, resulting in the loss of diverse cultural narratives by embedding largely American cultural content patterns as defaults.

Consider poetry. A model trained primarily on English-language poetry learns specific metaphorical conventions, rhythmic patterns, and cultural references. When a poet whose first language isn’t English uses this tool, they’re subtly pushed toward anglophone poetic conventions. Multiply this across domains and cultures: choreography tools that encode Western concert dance aesthetics, visual art tools that privilege certain compositional rules, music tools that assume particular harmonic frameworks.

We’re not just losing individual creative diversity. We’re risking a form of algorithmic cultural colonialism where dominant patterns in training data suppress alternative creative traditions.

Share

Understanding the full impact of creative homogenisation requires seeing creativity not as individual genius but as fundamentally collaborative and network-based/ meshwork. Research on creative collaboration (Eds Kaufman & Sternberg, 2019) treats the local communities that spark novel ideas and the broader fields that diffuse to recombine ideas in a unified fashion. Innovation emerges from diverse people bringing different perspectives into dialogue, from unexpected combinations across disciplinary or cultural boundaries.

Network effects on novelty, creative networks show that cross-cutting or cross-functional teams (bridging different backgrounds, disciplines, and perspectives) produce work that is both more creative and more commercially successful (monks.com; Reiter-Palmon & Leone, 2018). The magic happens at the intersections, where different ways of thinking collide and recombine.

But what happens when everyone in the network uses similar AI tools, trained on similar data? The diversity that makes creative networks productive diminishes. Even if individuals come from different backgrounds, they’re all channelling their creativity through the same algorithmic bottleneck, which smooths out difference in favour of learned patterns.

Human-AI hybrid creative networks reveals a troubling dynamic: whilst AI-only networks initially produce more creative outputs, over time human-AI collaborations maintain greater diversity because humans preserve continuity and cultural context that AI agents discard. This suggests that human judgment and cultural specificity are essential for sustaining creative diversity. But only if humans maintain genuine agency rather than deferring to algorithmic suggestions.

The tacit knowledge problem: Much creative expertise is tacit, embodied, difficult to articulate, transmitted through practice rather than explicit instruction. A dancer’s proprioceptive knowledge of how their specific body moves through space. A jazz musician’s sense of when to break the rules for expressive effect. A novelist’s intuition about when a character’s behaviour feels psychologically true.

This tacit knowledge develops through thousands of hours of deliberate practice, cultural immersion, and learning from diverse examples and mentors. Research on expert creative practice shows that expertise involves building extensive knowledge structures, pattern recognition capabilities, and evaluation heuristics developed over years of deep engagement (Doherty et al., 2025).

When AI tools offer instant competence-’good enough’ outputs without the slow work of developing expertise. They short-circuit this process. Students get better results faster, which feels like winning. But they’re not building the deep tacit knowledge that enables genuine innovation. They’re learning to prompt effectively rather than developing the embodied understanding that lets them recognise when to break patterns rather than follow them.

Measuring collective creativity-a difficult task: Computational social science has emerged as crucial for understanding creativity as a social phenomenon, using methods that would be challenging for other approaches. By analysing large-scale patterns (semantic networks, collaboration structures, diffusion pathways) researchers can see dynamics invisible at the individual level.

These methods reveal something important: individual creativity gains from AI assistance are real and measurable, but they come at a collective cost that’s equally measurable. We’re trading aggregate diversity for individual efficiency. Whether that’s a good trade depends on what we value, and crucially, whether we’re making the trade consciously or by default.

The question is what happens to the creative ecosystem when everyone uses similar tools optimised toward similar patterns. Do we lose the weird outliers and culturally specific approaches that often spark the most significant innovations?

Working with body-data in dance and movement offers a unique lens for understanding what gets lost in algorithmic mediation of creativity. When you capture movement through sensors (motion capture, accelerometers, EMG) you immediately confront the gap between quantifiable and qualitative data.

Motion capture records position, velocity, acceleration with high precision. But it misses intention, why the dancer made that choice. It misses quality, the difference between mechanically correct movement and movement imbued with meaning. It misses cultural specificity, the embodied knowledge of how bodies move in particular traditions. The data is accurate, but something essential to creative movement eludes datafication.

This isn’t a limitation unique to body-data. It’s fundamental to understanding creativity through computational methods. Research on computational creativity acknowledges that whilst algorithms can generate novel combinations, they struggle with the intentionality, meaning-making, and ‘cultural situatedness’ that characterise human creativity at its most significant (The Lancet Digital Health, 2023).

The embodied cognition insight reveals that thinking isn’t just a brain process, it’s distributed across brain, body, and environment. When dancers improvise, they’re not executing pre-planned sequences but responding in real-time to proprioceptive feedback, spatial relationships, other dancers’ movements, even audience energy. This distributed, embodied intelligence resists reduction to algorithmic patterns.

Exploring creative computing through embodied practice, learning algorithmic concepts through choreography, understanding data structures through physical movement or synthetic data, exploring conditionals through improvisation games, cultivates a different relationship to computational thinking. Students develop felt sense of algorithmic logic rather than abstract understanding. They learn that algorithms are tools serving human purposes, not authorities dictating correct outputs.

This matters for creative AI use. When creators have embodied understanding of their domain (whether dance, writing, visual art, or music) they can potentially maintain critical distance from algorithmic suggestions. They can recognise when AI outputs feel ‘off,’ when they’re getting generic rather than genuinely novel results. They have expertise to evaluate rather than simply accepting algorithmic authority.

The friction factor: Embodied creative practices inevitably involve friction, the gap between intention and execution, the need for interpretation, the time required for skill development. In live coded performance, there’s always latency between writing code and seeing results. In choreography, instructions require interpretation by dancers whose bodies, experiences, and cultural backgrounds shape execution. In improvised music, gaps between players’ responses create space for creative dialogue.

This friction isn’t inefficiency to be optimised away. It’s where creativity lives. Psychological research shows that constraints and friction in creative processes can enhance rather than limit novelty by forcing deeper engagement with problems and preventing premature convergence on familiar solutions Doherty et al., 2025.

Generative AI collapses friction. The speed from prompt to output (seconds rather than hours or days) feels like pure efficiency gain. But it eliminates the space for slow thinking, for trying multiple approaches, for ‘staying with the trouble’ (Haraway, 2016), until genuinely novel solutions develop. When ‘good enough’ results come instantly, why struggle toward something genuinely different?

Preserving agency in human-AI collaboration: Research suggests that if creators maintain autonomy and distinctive voice in AI interactions, homogenisation effects can be counteracted Salehzadeh Niksirat et al., 2024. But maintaining autonomy requires metacognitive awareness that most users haven’t developed. They need to understand their own creative process well enough to recognise when AI is doing the thinking for them versus genuinely augmenting their capabilities.

This is where embodied creative practices become pedagogically crucial. When students learn to be creative through physical, data practice. Where they directly experience the relationship between constraints, exploration, and novel outcomes, they develop intuitions about creativity that transfer to other domains. They’re better equipped to use AI critically rather than deferring to algorithmic suggestions by default.

The cultural preservation stake: When movement gets datafied and fed into generative systems, these cultural specificities risk being smoothed into statistically average patterns. The algorithm doesn’t ‘understand’ the cultural meaning of particular gestures or postures, it only knows their frequency in training data. Rare patterns get treated as noise to be eliminated rather than valuable cultural knowledge to be preserved.

This pattern repeats across creative domains. Language patterns specific to particular dialects or communities get smoothed into dominant standard forms. Visual aesthetics outside Western art historical traditions get marginalised as ‘unconventional.’ Musical traditions based on different harmonic or rhythmic systems get distorted toward Western norms embedded in training data.

Embodied creative practices (which are inherently situated, culturally specific, and resistant to full datafication) offer crucial counterweight to algorithmic homogenisation. They preserve spaces where creativity remains irreducibly human, where cultural difference is valued rather than smoothed away, where novelty emerges from lived experience rather than pattern matching.

How we teach computing in the age of generative AI will shape creative culture for decades. If students learn creative problem-solving primarily through AI-mediated tools, they’re being subtly trained in particular patterns. Ones that reflect biases and limitations of training data. But if we cultivate practices that maintain human agency, cultural specificity, and critical consciousness, we might produce technologists capable of building genuinely different futures.

The embodied computing pedagogy: Teaching computational thinking through embodied practice (choreography, physical computing, tangible interfaces) does more than make abstract concepts concrete. It cultivates a fundamentally different relationship to computation. Instead of computers as authorities that provide correct answers, students experience them as collaborators requiring interpretation and critical judgment.

When students choreograph using conditional logic (if dancer A raises arm, then dancer B turns), they’re learning control flow through culturally and individually specific movement. The algorithm serves their creative vision rather than dictating outputs. When they design wearables that give ambiguous instructions requiring interpretation, they’re building technologies that preserve human agency rather than eliminating it.

Research on computing education equity demonstrates that embodied and culturally responsive approaches significantly improve engagement and outcomes for students from underrepresented groups (Merkouris et al, 2017, Lachney et al., 2019). These aren’t just pedagogical niceties, they’re essential for building a computing workforce that reflects human diversity rather than perpetuating narrow technical cultures.

Teaching critical AI literacy: Students need metacognitive awareness about their own creative processes before they can effectively evaluate AI-generated suggestions. This means cultivating practices that slow down and make visible the normally invisible work of creative thinking.

Practical approaches include: having students document their creative process before using AI assistance, then comparing their initial thinking to AI suggestions; explicitly analysing whose aesthetic or cultural patterns AI outputs privilege; designing projects where students must intentionally resist AI suggestions to maintain distinctive voice; teaching prompting as critical practice rather than just technical skill, understanding how prompt choices shape what becomes thinkable.

The goal isn’t rejecting AI tools (which students will use regardless). It’s developing sophisticated users who maintain creative agency rather than deferring to algorithmic authority.

The data ethics imperative: Every technical choice is a cultural and political choice. When students select training data, they’re choosing whose creativity becomes template. When they optimise algorithms, they’re encoding values. When they deploy systems, they’re reshaping social possibilities.

Creative computing education must make these choices visible rather than treating them as neutral technical decisions. This means teaching data ethics not as separate module but integrated throughout: whose movement data are we capturing, and do they control how it’s used? What cultural patterns does this model privilege? Who benefits from this tool’s deployment, and who is disadvantaged?

Students need to understand that ‘debiasing’ algorithmic systems isn’t just technical challenge. It’s fundamentally about confronting structural inequalities that technical solutions alone cannot address. Research on algorithmic bias (such as Zhang et al, 2025; Taka, 2023)shows that systems trained to be ‘neutral’ invariably encode dominant cultural assumptions as defaults, marginalising alternative perspectives.

Cultivating alternative epistemologies: Computing education has traditionally privileged certain ways of knowing: abstract, formal, disembodied. But innovation often comes from bringing different epistemologies into dialogue. Embodied knowledge from dance and physical practice. Qualitative understanding from humanities and arts. Cultural knowledge from diverse communities.

By exploring and teaching creative computing through multiple modalities (not just text and symbols but movement, sound, physical materials) we cultivate students who can think across epistemologies. They’re better equipped to recognise what algorithmic approaches miss, to value forms of knowing that resist datafication, to build technologies that genuinely serve human flourishing rather than just optimising narrow metrics.

This matters urgently as AI tools become ubiquitous. We’re training the people who will build the next generation of creative technologies. If we teach only technical competence without ethical consciousness, we’re producing sophisticated perpetuators of homogenisation. But if we cultivate critical consciousness, embodied understanding, and commitment to cultural diversity, we might produce technologists capable of building genuine alternatives to Big Tech monoculture.

Share

Understanding creative homogenisation as structural problem rather than individual failing points toward systemic solutions. Individual creators can develop sophisticated AI literacy, maintain distinctive voice, resist algorithmic suggestions. But when the entire creative ecosystem is shaped by a handful of companies controlling foundation models, individual resistance isn’t sufficient.

Data sovereignty and cultural self-determination: Communities should control how their creative traditions and cultural knowledge are documented, shared, and computationally represented. This applies to Indigenous knowledge systems, minority cultural practices, regional artistic traditions, any creative domain at risk of extraction and homogenisation.

Practical implementations include: community-controlled datasets that require meaningful consent and benefit-sharing for use in model training; licensing frameworks that prevent cultural knowledge extraction without accountability; infrastructure supporting communities to build their own AI tools optimised for their aesthetic values rather than dominant patterns.

The goal isn’t isolating cultural practices from computational tools, but ensuring communities maintain agency over how their creativity is datafied and algorithmically mediated. This is data ethics applied to cultural preservation. It is recognising that creative knowledge is heritage deserving protection, not raw material for extraction.

Algorithmic alternatives from the margins: Rather than trying to make dominant AI systems “fair” through debiasing interventions, we need genuine alternatives built by and for marginalised communities. Critical data studies scholars argue that responsible AI development requires ongoing engagement with affected communities, transparency about system limitations, and mechanisms for contestation and redress Mone & Shakhlo, 2023.

For creative AI, this means: funding open-source alternatives to proprietary models; supporting development of models trained on diverse cultural corpora; building tools optimised for maintaining rather than smoothing cultural difference; creating infrastructure allowing communities to train models reflecting their aesthetic values.

This isn’t about technological separatism. It’s about refusing monoculture, recognising that diverse creative futures require diverse technical infrastructures, not just diverse content created through identical algorithmic intermediaries.

Preserving spaces of productive friction: Not everything should be optimised for efficiency. Some creative processes benefit from slowness, interpretation, ambiguity. We need to deliberately preserve spaces where these qualities are valued rather than treated as inefficiency to be eliminated.

This includes: educational contexts that teach slow, deep expertise-building rather than rapid prompting; collaborative creative processes that maintain human dialogue and negotiation; performance contexts where liveness and improvisation resist algorithmisation; cultural practices where transmission through apprenticeship preserves tacit knowledge.

These aren’t nostalgic retreats from technology. They’re intentional preservation of creative modalities that algorithms cannot fully replicate, not due to technical limitations we’ll eventually overcome, but because some forms of creativity are irreducibly embodied, culturally situated, and emergent from human sociality.

Regulatory frameworks for creative diversity: Just as we have policies protecting biodiversity, we might need frameworks protecting creative diversity. This could include: transparency requirements about training data sources and model capabilities; impact assessments for large-scale creative AI deployments; support for alternative technical infrastructures; protections against cultural knowledge extraction without consent.

The challenge is crafting regulation that protects diversity without stifling innovation, that acknowledges legitimate concerns about homogenisation without falling into technophobic rejection of useful tools.

Conclusion: Maintaining Collective Creative Capacity

The paradox remains: generative AI enhances individual creativity whilst reducing collective diversity. This isn’t a technical problem with technical solutions. It’s a social, cultural, and political challenge requiring responses at multiple levels.

Individual creators can develop sophisticated AI literacy, maintaining distinctive voice and critical distance from algorithmic suggestions. Educators can teach computing through embodied, culturally responsive practices that cultivate agency rather than deference to algorithmic authority. Communities can assert control over their cultural knowledge and build alternative technical infrastructures. Policymakers can create frameworks protecting creative diversity whilst enabling beneficial innovation.

But underlying all these responses must be recognition that creativity is fundamentally social phenomenon, not just individual attribute. Innovation emerges from diverse perspectives colliding, from unexpected combinations, from cultural difference preserved rather than smoothed away. When algorithms mediate our imagination, trained on patterns extracted from existing work, they subtly push toward convergence.

The question isn’t whether to use AI tools, they’re here, they’re useful, they’re not disappearing. The question is whether we use them consciously, maintaining the spaces and practices that preserve genuine novelty. Whether we optimise purely for individual productivity or also value collective diversity. Whether we accept algorithmic monoculture by default or deliberately cultivate alternatives.

We’re witnessing not just technological change but risk of technological colonialism at computational scale where a few entities’ cultural values and biases shape global creative norms Pincus, 2016. This is not a new phenomenon. This affects what stories get told, what movements feel ‘natural,’what ideas seem thinkable, whose ways of being in the world get encoded as normal.

Maintaining collective creative capacity means preserving the messy, embodied, culturally specific practices that resist algorithmic smoothing. It means teaching students to dance with algorithms without letting them choreograph everything. It means building technologies that amplify difference rather than erasing it. It means choosing diversity over efficiency when they conflict.

Because creativity isn’t just about how good individual outputs are. It’s about maintaining the collective capacity to imagine genuinely different futures. And that requires preserving the full spectrum of human creative possibility. Not just the patterns that fit easily into algorithmic templates.

The algorithmic monoculture isn’t inevitable. It’s a choice we’re making right now, largely by not choosing. We can choose differently. We must.