ai, money and the economics of sameness

The debate about using AI to write is often framed as though there are only two possibilities: either a person writes every word unaided, or the machine writes for them. Reality is less binary. There are situations where unaided writing is entirely appropriate. Examinations, assessments and other contexts that explicitly require independent work should remain exactly that. Outside those settings, the more useful question is not whether AI participates in writing, but how it participates in the organisation of thought.

Large language models are often treated as machines for generating ideas. That is not their greatest strength. They are built from the accumulated patterns of human communication and tend, when given little direction, towards forms that have already achieved statistical persistence. Ask one for a business idea, a content strategy or a way to become wealthy and it will usually return a competent variation on structures that already circulate widely. The answer may appear original within a single conversation while remaining entirely ordinary across the wider communicative field.

That does not mean people are foolish to ask. Economic life is insecure, opportunity is unevenly distributed and these systems appear to provide inexpensive access to forms of analysis, advice and productive capacity once available only through education, capital or professional networks. Generative AI has produced measurable productivity gains in some forms of work, particularly for less experienced workers, so the expectation that it can confer an economic advantage is grounded in evidence (Brynjolfsson, Li and Raymond, 2023). The cold water arrives when millions of people seek that advantage through the same systems, asking much the same questions in pursuit of much the same markets.

The advantage then begins to disappear through use. AI lowers the cost of producing articles, advertisements, videos, applications, courses, consultancy and countless other forms of digital production, but it does not proportionally expand the money, time or attention available to consume them. When productive capacity grows faster than demand, competition intensifies. Models of AI-mediated content platforms consequently predict oversupply, information overload and increasing concentration of attention around already successful creators, even as overall production continues to expand (Zhang, 2024).

This is not a field organised around cooperation. It is increasingly organised around competition. Every participant is encouraged to outperform every other participant for a limited supply of attention, income and recognition. Platforms continuously measure, rank and reward visibility, transforming communication into an ongoing contest of output, optimisation and strategic self-presentation. Individually, this behaviour is understandable. Collectively, it produces a system in which everyone is encouraged to generate more because everyone else is generating more, while the resource they are competing for remains fundamentally limited.

The same dynamics apply to the companies building these systems. They do not stand outside the competitive field; they are immersed within it. Each faces pressure to acquire users, attract investment, release new capabilities, increase engagement and establish market dominance before competitors do. Under those conditions, convergence is not merely a property of the models but of the industry itself. Firms increasingly optimise against one another, drawing upon similar research, benchmarks, commercial incentives and expectations of growth. The systems they produce therefore reflect not only the statistical persistence of language, but the statistical persistence of markets. AI is shaped by competition at precisely the same time as it intensifies competition among those who use it.

This symmetry matters because it reveals that the technology is not an external force acting upon society. It is one expression of the same economic and communicative dynamics already reorganising society more broadly. Users optimise against one another. Companies optimise against one another. Platforms optimise against one another. Each adaptation alters the conditions under which the next adaptation becomes rational. The result is a recursively coupled system in which sameness is continually reproduced, not because anyone intends it, but because the field increasingly rewards those trajectories that have already demonstrated persistence (Bommasani et al., 2024).

Research on creative work already reflects this dynamic. AI assistance can improve the assessed quality of individual outputs while simultaneously reducing their collective diversity. AI-assisted stories have received higher quality ratings when assessed individually, while also becoming more similar to one another. Comparable effects have been observed during idea generation, where outputs from different users become less semantically distinct (Doshi and Hauser, 2024; Anderson, Shah and Kreminski, 2024). The technology can improve individual performance while quietly reducing variation across the population.

Nor is this process likely to stabilise once people realise that effortless content generation rarely produces effortless wealth. Saturation becomes the starting point for another cycle of adaptation. New tools promise better prompts, automated distribution, synthetic audiences, optimisation, ranking, verification and escape from the congestion created by earlier tools. New narratives promise that success remains one workflow, subscription or platform away. Failure does not terminate the process. It becomes the raw material from which the next layer of technological and commercial infrastructure emerges.

The result is a recursive economy of competitive escalation. As returns become harder to secure, creators produce more material, automate more workflows and optimise more aggressively for visibility. Platforms respond with new recommendation, moderation and ranking systems, which creators immediately begin optimising against in turn. Evidence from Pixiv already suggests that increasing AI-generated supply can displace human creators from audience attention, particularly in categories most exposed to machine production (Kim, Jin and Lee, 2026). AI may democratise production without democratising attention, influence or income.

Its deeper value lies elsewhere. Large language models are powerful tools for composition. They organise fragments, expose hidden relationships, restructure arguments and help partially formed ideas acquire coherent expression. They are most valuable when the writer contributes something the statistical field cannot generate on its own: lived experience, sustained observation, specialised knowledge, judgement, contradiction, uncertainty and a genuine reason for writing. The model can help organise those differences. It cannot substitute for them.

A better way to use AI is therefore to begin before the prompt. Bring it an argument that refuses to stabilise, two ideas whose relationship remains unclear, notes accumulated across years, a pattern that resists articulation or a draft whose structure is failing. Ask it to compare, test, reorganise and challenge. Revise repeatedly. Reject whatever feels generic, convenient or suspiciously complete. The objective is not to receive ideas from the machine, but to use interaction with the machine to reorganise your own thinking.

The more consequential development is that language has begun interacting with itself through technological recursion. Models trained on accumulated human communication now generate communication that shapes future writers, markets and models. As people repeatedly ask what will generate attention, money or approval, those communicative patterns acquire increasing statistical and economic persistence. Sameness does not merely reproduce itself. It reorganises the conditions from which future communication emerges.

The challenge is not deciding whether AI is good or bad for writing. It is preserving the conditions under which genuinely new organisation can emerge. Used carelessly, these systems compress thought towards forms that have already achieved statistical, cultural and economic persistence. Used thoughtfully, they become instruments for exploring relationships, refining ideas and extending human thought beyond its current organisation. The difference is not whether AI is involved. It is whether the writer remains the source of direction, or simply asks the accumulated machinery of language what everyone else should do next.

References

Anderson, B.R., Shah, J.H. and Kreminski, M. (2024) ‘Homogenization effects of large language models on human creative ideation’, in Proceedings of the 16th Conference on Creativity & Cognition. New York: Association for Computing Machinery, pp. 413–425. doi: 10.1145/3635636.3656204.

Bommasani, R. et al. (2024) The Foundation Model Transparency Index. Stanford University.

Brynjolfsson, E., Li, D. and Raymond, L.R. (2023) Generative AI at Work. NBER Working Paper No. 31161. Cambridge, MA: National Bureau of Economic Research. doi: 10.3386/w31161.

Doshi, A.R. and Hauser, O.P. (2024) ‘Generative AI enhances individual creativity but reduces the collective diversity of novel content’, Science Advances, 10(28), eadn5290. doi: 10.1126/sciadv.adn5290.

Kim, S., Jin, G.Z. and Lee, E. (2026) Does Generative AI Crowd Out Human Creators? Evidence from Pixiv. NBER Working Paper No. 34733. Cambridge, MA: National Bureau of Economic Research.

Zhang, Y. (2024) ‘The influence of generative AI on content platforms: Supply, demand, and welfare impacts in two-sided markets’, arXiv preprint, arXiv:2410.13101.