Doshi and Hauser, in a study published in 2024 in Science Advances, showed that AI assistance in creative tasks tended to homogenize productions: stories created with AI assistance proved more similar to each other than those produced without it. Technical progress increased individual productivity, but in this writing experiment, access to AI-generated ideas reduced the collective diversity of stories.
The Essential Points
- In a short-story writing task with fixed output, AI assistance increased the similarity of produced stories (Doshi and Hauser, Science Advances, 2024).
- AI assistance tends to homogenize creative productions, according to Doshi and Hauser’s study of creative writing tasks.
- Recommendation algorithms amplify concentration by directing demand toward already-dominant median styles, at the moment when creation tools are manufacturing them at scale.
- UNESCO reports projected global revenue losses of 24% for music creators and 21% for audiovisual creators by 2028, due in part to the impacts of generative AI.
- By 2030–2035, the possibility of redirecting these tools toward diversity depends on design and regulatory choices still to be determined.
The Vendi Score and What It Measures
This drift toward the center of the distribution is not explained solely by the choices of creators. Generative models are trained on large corpora to learn statistical distributions, whose composition depends on data selection and preprocessing. The Vendi Score captures this result precisely because it measures diversity at the scale of the whole collection rather than at the scale of each individual work. A work may appear original viewed alone, and yet contribute to a collection that is narrowing. It is this tension between the creator’s individual experience and the collective outcome that the indicator makes visible.
Without a measure operating at this scale, the phenomenon remains invisible to each actor, since each creator evaluates their own production, never the full range of what their peers are producing simultaneously with the same tools. Convergence thus takes hold without anyone deciding it or perceiving it from their individual position.
The Vendi Score, proposed by Friedman and Dieng, is a measure that assesses how much the works in a collection resemble each other. A high score means varied works; a low score means a collection that repeats itself. Experimental data on short stories show a convergence of productions under AI assistance.
Under AI assistance, works produced in the same task tend to resemble each other more, even if each work considered in isolation may be technically satisfactory. Human portfolios contain failures, aborted attempts, unusual directions, and it is precisely this irregularity that feeds collective diversity. AI, designed to optimize satisfaction criteria, can reduce the unpredictable results that characterize certain creative processes.
This mechanism is not accidental. Generative models are trained on massive corpora weighted toward what has worked historically. This training structure can orient generated outputs toward the most frequent forms in the data, which risks bringing creators closer to similar solutions.
The Recommendation Algorithm Closes the Loop
The trend toward production homogenization would already be concerning if demand remained plural. Yet the recommendation algorithms of Spotify, YouTube, or TikTok optimize for short-term engagement and direct users toward what resembles what they already liked.
Algorithms impoverish culture at each iteration, an observation that predates the massive arrival of generative tools. The interaction between AI-assisted production and algorithmic recommendation can create a loop in which the most common forms in training data are amplified.
Doshi and Hauser, in a study published in 2024 in Science Advances, showed that AI assistance in creative tasks tended to homogenize productions: participants using the tool produced ideas more similar to each other than those working alone.
Guerouaou’s Contribution to Measurement
Nadia Guerouaou, in Our Brain Under Influence, describes how generative AIs reconfigure emotional expressivity in real time: voice, face, and text. According to her, these tools shape the way individuals express what they feel, even generating what she calls a “technomorality”—a progressive affective normalization.
According to Guerouaou, when a creator uses a generative tool to complete a production, they tend to adapt their expressivity to what the tool proposes. This adaptation, repeated by many creators simultaneously, can contribute to observable convergence in collective productions.
Guerouaou emphasizes that expressivity has never been a fixed given: the printing press, photography, and sound recording each transformed what creators produced and the manner in which they produced it.
Generative tools are distinguished by their speed and scale: their deployment operates in real time, across billions of simultaneous users.
Revenue Concentration Widens the Gap
This dynamic also acts on creative behaviors upstream of production. When a creator anticipates that only a narrow fraction of works will capture a significant share of revenues and attention, they may adjust their stylistic choices accordingly. Revenue concentration in cultural industries can orient creators toward similar stylistic choices.
Escaping this dynamic would require simultaneous modification of several variables: tool architecture and revenue distribution. Acting on one without the other leaves other pressures intact. Creators who seek to depart from dominant prototypes assume an economic cost that the current structure does not offset.
UNESCO, in its 2026 report Reshaping Policies for Creativity, documents the impact of generative AI on cultural industries, particularly projected revenue losses. Content abundance makes discovery harder and reinforces reputation effects, concentrating attention on certain works and artists.
Culture weighs more than before, the creator is more isolated: this concentration of revenues and attention predates AI. The lowering of technical barriers is accompanied by intensified competition for attention. Creators who already have an audience benefit from increased productivity; those trying to build one drown in a flow that grows faster than audiences’ capacity to pay attention.
Tyler Cowen, whose work on the economics of culture examines how attention markets distribute gains among creators, offers a reading complementary to Guerouaou’s. For Cowen, revenue concentration in cultural industries is a structure inherent to markets where reproducing a good is free and reputation signals quality in a noisy space. AI can accelerate these concentration dynamics. These dynamics raise the question of whether spontaneous market adjustment will suffice to preserve cultural diversity.
What Can Still Change by 2030
Nothing in current data allows projecting a stable trajectory by 2035. But available signals allow examining plausible scenarios.
The first scenario is one of continued concentration. Generative tools become more powerful, cheaper, more integrated into platforms. Creators who do not use them lose ground in productivity. Those who do might converge toward the same prototypes. Cultural diversity could continue to decline, and revenues could concentrate further.
In this scenario, regulation of recommendation algorithms becomes the only available lever in the short term, by imposing, for example, diversity quotas in recommendation feeds, along the lines of what the European Union is exploring in applying the Digital Services Act to large platforms.
The second scenario is one of rebalancing through the tool itself. Generative models could be explicitly trained to maximize the diversity of outputs rather than their statistical likelihood. Research in this direction exists; it amounts to replacing the function optimizing toward the center of the distribution with a function that penalizes repetition and rewards divergence. If this type of architecture enters mainstream tools, creators could have assistants that pull them toward originality rather than toward consensus. This scenario depends on design choices that technology companies have no spontaneous incentive to make, and which would require regulatory or competitive pressure.
The third scenario is one of market bifurcation. Part of cultural demand shifts toward what is explicitly produced without algorithmic assistance, something certain craft markets experienced after industrialization. Certification labels, dedicated platforms, audiences willing to pay a premium for human singularity could emerge. This movement already exists at small scale in certain communities of readers, listeners, and collectors. Its capacity to expand depends on signal clarity—being able to distinguish what was produced with or without generative tool, something current technologies make difficult.
Administration by algorithms gains speed but loses legitimacy: the same trade-off between efficiency and legitimacy runs through cultural creation. Public decision-makers, platforms, and creators have tools that increase individual productivity; their redirection toward preserving collective diversity remains a choice to be made. Doshi and Hauser’s data show that AI assistance tends to homogenize productions in the observed tasks.
Sources
- Youngblood, Olson et al., Scientific Reports, 2026, Diversity of AI-assisted portfolios: https://arxiv.org/pdf/2609.02620
- Nadia Guerouaou, Our Brain Under Influence: How Generative AIs Shape Our Emotions, Éditions Eyrolles, https://www.editions-eyrolles.com/livre/notre-cerveau-sous-influence
- UNESCO, Reshaping Policies for Creativity, 2026, Concentration of creative revenues and impact of AI on cultural industries
- Doshi, A. & Hauser, O., Science Advances, 2024, Homogenization of creative productions with AI assistance
