Cultural homogenisation
If a growing share of what is read, seen and heard passes through a few models, the available repertoire narrows.
- Severity
- Severe
- Horizon
- 1–3 years
- Evidence
- Projected
- Consensus
- Low
The hypothesis is easy to state and hard to measure: if a growing share of what is read, watched and heard is produced with a handful of models, trained on similar data and tuned to similar criteria, the available cultural repertoire narrows even though nobody decides it.
The two observable inputs are there. About half of new indexable articles in English are mostly AI-generated [472]AI Now Writes as Many Online Articles as Humans DoView source ↗, and use of those models is concentrated by geography: high-income countries are over-represented relative to their working-age population [71]Anthropic Economic Index report: Uneven geographic and enterprise AI adoptionView source ↗, with more than 50% of the working-age population using AI in some states against less than 10% in many low-income economies [538]International AI Safety Report 2026View source ↗. And there is a laboratory mechanism that points in the same direction: Shumailov and co-authors show in Nature that training a generative model recursively on synthetic data produces irreversible defects in which the tails of the original distribution disappear [967]AI models collapse when trained on recursively generated dataView source ↗.
What this does not demonstrate. Practically nothing, and that has to be said first: there is no series of cultural diversity comparable over time that would allow a narrowing to be detected, let alone attributed to AI. This is the worst-supported risk in the observatory. The mechanism usually lent to it does not hold up as it stands either: Shumailov’s model collapse depends on the assumption that the new data replaces the old, and Gerstgrasser and co-authors show that accumulating synthetic data alongside real data avoids it, across a range of sizes, architectures and hyperparameters [450]Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic DataView source ↗. The web accumulates: it does not delete the articles of 2019 when it publishes those of 2026. Besides, the phenomenon is about the model, not about culture; the bridge between the two is an argument, and it is held up by a literature that declares itself conceptual [618]Gradual Disempowerment: Systemic Existential Risks from Incremental AI DevelopmentView source ↗.
Chain of materialisation
PreconditionObserved
Mediation runs through few models and their use is concentrated
About half of new indexable English-language articles are primarily AI-generated, and use of those models is geographically concentrated: high-income countries are overrepresented relative to their working-age population, with over 50% usage in some states and under 10% in many low-income economies.
TriggerLab
Generative models lose the tails of the distribution
Shumailov and co-authors show in Nature that repeatedly training a generative model on recursively generated data produces irreversible defects in which the tails of the original content distribution disappear. The effect appears in language models, variational autoencoders and Gaussian mixtures.
Observed and demonstrated evidence ends here. What follows is projection.
CascadeProjected
What the model loses, the available repertoire loses too
The jump from a training phenomenon to observable cultural narrowing is precisely this risk's weak point. Kulveit and co-authors place it in the cultural domain of their argument -economic power shapes narratives and narratives alter economic behaviour- but provide no measurement.
ImpactSpeculative
Less variety than anyone kept count of
There is no time-comparable cultural diversity series that would allow detecting narrowing or attributing it to AI. This is the worst-supported risk in the observatory and that should be said before anything else.
Scenarios where it appears
See on the map →Report a mistake in this entry →
Sources
- [967] AI models collapse when trained on recursively generated data · University of Oxford / University of Cambridge / Imperial College London / University of Toronto 2024
- [450] Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data · Stanford University / Harvard University / MIT 2024
- [472] AI Now Writes as Many Online Articles as Humans Do · Graphite 2026
- [71] Anthropic Economic Index report: Uneven geographic and enterprise AI adoption · Anthropic 2025
- [618] Gradual Disempowerment: Systemic Existential Risks from Incremental AI Development · Kulveit, Jan 2025
- [538] International AI Safety Report 2026 · International AI Safety Report (panel con representantes nominados por más de 30 países) 2026
- [773] AI as Normal Technology · Knight First Amendment Institute at Columbia University 2025
- [493] The levers of political persuasion with conversational artificial intelligence · UK AI Security Institute / University of Oxford 2025 verified through Crossref