The Homogenisation Effect: Convergent Outputs in AI-Driven Workflows

Advances in tools, computation, and access to information have been expected to expand the range of what humans can create, enabling faster experimentation, broader participation, and more efficient execution.

Today, however, a different pattern is beginning to draw attention across research, design, and technology communities.

AI-mediated systems have dramatically increased the efficiency of knowledge work. Teams generate more concepts, more prototypes, and more polished outputs than at any previous moment. At the same time, emerging empirical research and industry observation point to a quieter tension: as output increases, variation may be decreasing. Across domains, independently produced solutions appear to cluster around similar structures, aesthetics, and problem framings.

Standardisation and convergence are not inherently negative. Historically, shared tools and conventions have enabled scale, coordination, and even periods of accelerated innovation.

But how innovation changes when the cognitive systems shaping solutions are primarily optimised for statistical convergence rather than exploratory divergence?

Two Different Things Called "Thinking"

Contemporary language models are, at their core, high-dimensional probability estimators. They don’t explore unknown solution spaces; they redistribute mass over learned distributions. As Melanie Mitchell and others in cognitive AI argue, this produces systems with remarkable surface fluency but limited capacity for what researchers call "out-of-distribution generalisation" — reasoning through genuinely new situations that fall outside the training manifold. For professionals tackling complex, ambiguous, or socially situated problems, this matters immensely.

From a neuroscientific perspective, human innovation relies on something radically different. Beaty et al. (2016) show that creativity emerges from the interaction between brain networks that usually compete: the Default Mode Network (imagination and mental simulation), the Executive Control Network (critical evaluation), and the Salience Network (detecting what matters). When these three networks operate in simultaneous coupling — measurable via fMRI and characteristic of highly creative individuals — the brain can evaluate critically while still in generative mode. It’s this state that allows for cross-domain associations — combinations no model trained on the past would produce as a high-probability output, because they’ve never existed before. This cooperation enables something no AI architecture possesses: generating ideas that are statistically unlikely, yet cognitively significant.

The question isn’t whether humans and AI can work together. Clearly, they can. The question here is what this partnership does, silently, to the diversity of what we produce. And that’s where the data gets uncomfortable.

Studies on AI-assisted creative writing (Gómez-Rodríguez & Williams, 2023; Doshi & Hauser, 2024) found that while individual outputs were rated higher in quality when AI-assisted, the variance between outputs dropped significantly. AI raised the floor while compressing the ceiling. In practical terms: fewer failures, but also fewer outliers. Fewer breakthroughs.

Doshi & Hauser’s (2024) study in Science Advances is particularly telling. When writers used AI assistance, their stories became more similar—converging on narrative patterns that were statistically safe, structurally familiar, and emotionally predictable. The individual diversity of the corpus dropped by a measurable margin, even as average quality rose.

The risk is not that AI produces bad work. The risk is that it produces competent, indistinguishable work — at scale.

The Homogenisation Effect

The implications for UX and product design are direct. When design teams use the same AI tools — trained on the same datasets, with the same interface paradigms encoded as high-probability outputs — the available design space shrinks. Not because the tools are bad, but because they are too good at producing what has already worked.

The potential result isn't a creative explosion, but the emergence of a "culture of the median" — safe, predictable solutions optimised for statistical consensus rather than conceptual disruption.

This is the homogenisation problem, and it operates (at least!) on three levels:

• At the Artifact Level: Generated solutions — interfaces, product strategies, technical decisions—cluster around established conventions. The tools have absorbed what’s already been done and reproduce it faithfully.

• At the Problem-Framing Level: AI systems optimise toward defined goals. They don’t question if the problem is framed correctly. In product and engineering, reframing the problem is often the most valuable work — and it requires exactly the kind of motivated ambiguity and contextual judgment that current systems can’t replicate.

• At the Organisational Level: When AI becomes the default starting point for ideation, the divergent, uncomfortable, and inconclusive stages of the process — where new insights actually emerge — are compressed or skipped. The process accelerates, but the solution space narrows.

Neuroscience offers a precise explanation here. Research on creative cognition (Beaty et al., 2016; Kounios & Beeman, 2014) shows that insight and novel ideation are linked to periods of defocused attention, incubation, and productive ambiguity. These are the states the Salience Network flags as relevant when it hits an anomaly it can't easily resolve. Workflows that prioritise rapid output structurally reduce time spent in these states. By removing cognitive friction, they eliminate the very signal that directs attention toward problems worth deeper investigation.

Furthermore, Immordino-Yang and Damasio (2007) demonstrated that emotion and cognition are structurally inseparable — somatic markers guide judgments of relevance and drive sustained exploration. A PM frustrated with a solution that "almost" works is, neurologically, in a highly productive state. That frustration activates the Salience Network. AI doesn't have this state. It doesn't get frustrated; it has no drive to understand why. The emotional substrate that sustains genuine inquiry is absent by design.

Extended Cognition, Not Replacement

The most defensible framework positions AI as a form of cognitive extension, not a replacement. Clark and Chalmers’ (1998) "extended mind" thesis anticipated this: when a tool couples tightly with cognitive processes, the boundary between tool and thinker becomes practically irrelevant. What matters is whether the integrated system — human plus tool — produces outcomes that neither could achieve alone.

Evidence suggests it can — but only under specific conditions: when the human professional brings genuine domain expertise, maintains active critical judgment, and uses AI to expand problem exploration rather than taking a shortcut to a solution.

The question is not whether to use AI. It is whether the human in the loop is contributing judgment, or merely providing approval.

The distinction is operational, not philosophical. Using AI to generate 40 variations and applying expert judgment to pick the two with real potential is cognitively demanding — it actively engages the Executive Control and Salience Networks. Using AI to generate the "optimal" version and giving it a superficial "looks good" bypasses this process. The first workflow amplifies human intelligence. The second gradually replaces it — and you’ll hardly notice it happening because the outputs still look great.

What This Means for Anyone Building Products with AI

If the homogenisation effect is real, then competitive differentiation in product and tech will increasingly depend on the quality of the problem, not the quality of the output.

Key skills are shifting:

• From Execution → to the ability to question which problem is worth solving.

• From Individual Ideation → to expert curation and judgment at scale.

• From Process Efficiency → to the deliberate preservation of conditions that allow for non-obvious insight.

  
  

Organisations that optimise solely for AI-assisted speed will eventually converge in the same design space as everyone else using the same tools. The sustainable advantage belongs to teams that use AI to explore more of the problem space — not just to reach the same destination faster.

The Question(s) Worth Asking

The most honest assessment of AI’s current role in innovation is this: it is an extraordinary amplifier of execution. It brings the cost of producing competent work down to near zero. This is a massive shift that reshapes the economics of design and team structures.

But amplifying execution is not the same as generating the insight that makes the execution worthwhile. The ability to sit with a poorly defined problem, resist the urge for premature closure, and arrive at a reframing that changes the entire solution space — that remains stubbornly human. Not for romantic reasons, but for structural ones.

The organisations winning with AI are not the ones that automated creativity. They are the ones that used AI to free up the cognitive bandwidth for the kind of thinking that cannot be automated.

The practical questions for any professional are clear: In your current workflow, where is human judgment actually happening? Is that the right place for it? Is there still room for the Default Mode Network — for imagination, mental simulation, and non-obvious connections — to run uninterrupted? For incubation, for productive ambiguity, for the insight that doesn’t come on demand?

And if not—what are you assuming is filling that gap?

References

Beaty, R.E., Benedek, M., Silvia, P.J. & Schacter, D.L. (2016). Creative cognition and brain network dynamics. Trends in Cognitive Sciences, 20(2), 87–95.

Clark, A. & Chalmers, D. (1998). The extended mind. Analysis, 58(1).

Doshi, A.R. & Hauser, O.P. (2024). Generative AI enhances individual creativity but reduces collective diversity. Science Advances, 10, eadn5290.

Gómez-Rodríguez, C. & Williams, P. (2023). A Confederacy of Models: a Comprehensive Evaluation of LLMs on Creative Writing. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 14504–14528, Singapore.

Kounios, J. & Beeman, M. (2014). The cognitive neuroscience of insight. Annual Review of Psychology, 65. Mitchell, M. (2020). Artificial Intelligence: A Guide for Thinking Humans. Picador.

Immordino-Yang, M.H. & Damasio, A. (2007). We feel, therefore we learn. Mind, Brain & Education, 1(1).