Design for Difference: Practical Levers to Prevent Homogenisation in Sociotechnical Systems

Homogenisation need not be inevitable. Treating software and AI as sociotechnical systems surfaces practical levers—social expertise, preserved human judgement, diversified technical choices and infrastructure planning—that allow teams to keep shared efficiencies while protecting local fit, opportunity and resilience.

Why homogenisation is a sociotechnical problem, not just a technical one

When people worry that AI or software makes everything look and feel the same, they are often tracking a pattern with deep social and infrastructural roots. Design trends and platform incentives push interfaces, interactions and even opportunities toward similar outcomes (Sessions College describes this long arc in design history). But research shows the mechanics matter: it isn’t solely the model or the algorithm that flattens difference, it is how systems are built, presented and governed.

Four practical levers to design for difference

Below I expand on the core claim: homogenisation is not inevitable. Here are concrete levers teams can use, grounded in recent work on design interfaces, AI personas and the effects of frictionless design.

  • 1. Involve social expertise early and often

    Technical choices have social consequences. Inviting sociologists, ethnographers, designers and local stakeholders into design conversations early helps reveal where a single default will harm local fit or opportunity. The literature on design homogenisation and sociotechnical practice emphasises that context matters: results from one setting (for example educational essay-writing studies) do not automatically generalise to others, and prematurely broad claims can mislead product direction (see the critique of overgeneralisation in the frictionless design discussion).

    Practical steps: set up short cultural-probe sessions before major defaults are set; run participatory requirement workshops; document non-functional expectations such as dignity, equity and interpretability alongside performance metrics.

  • 2. Preserve meaningful human judgement and configurable friction

    Sometimes “frictionless” design accelerates convergence because it hides choices and removes opportunities for local tailoring. The argument that frictionless interfaces, not AI per se, drive homogenisation suggests a useful countermeasure: reintroduce purposeful decision points where humans provide judgement.

    Practical steps: expose provenance and stylistic options (not just a single default), make human review easy and efficient, and design checkpoints where operators can override automated outputs. These friction points should be lightweight but meaningful—enough to support local adaptation and guard against one-size-fits-all behaviour.

  • 3. Diversify technical choices—treat model style as an interactional option

    Research on generative interfaces suggests we can structure systems so diversity is an input requirement rather than a hoped-for emergent property. The concept of “style subfloors” or exposing the choice of model weights as a primary interactional choice allows users to anchor outputs in distinct aesthetic or cultural lineages instead of a single homogenised mode.

    Practical steps: offer multiple curated model personas or stylistic presets (research indicates diverse personas can mitigate homogenisation), let teams select or combine backend models deliberately, and provide clear labels explaining what each persona or model tends to do. Design the UI so selection is visible and reversible, and keep model provenance traceable in logs and metadata.

  • 4. Plan for infrastructure effects and incentives

    Infrastructure—defaults, pipelines, monitoring, third-party integrations—shapes behaviour at scale. The same set of integration choices across many teams produces convergence in outcomes. If you want diverse outcomes, consider how infrastructure steers decisions and where you can introduce variability while maintaining safety and efficiency.

    Practical steps: make sensible defaults that are intentionally plural rather than singular; document trade-offs for each default; instrument systems to surface convergence metrics (e.g. how often one preset is used across contexts); and align platform incentives so that local customisation is not penalised by extra cost or maintenance burden.

Operational checklist for teams

  • Run a quick stakeholder mapping: who will benefit or lose from a single default?
  • Prototype at least two contrasting presets or personas before shipping a single option.
  • Expose choices and provenance in the interface; treat model selection as meaningful UX, not a backend detail.
  • Add lightweight human checkpoints for high-impact decisions and keep override logs for auditability.
  • Monitor for convergence: collect metrics on usage patterns, diversity of outputs and downstream effects on opportunities.

What the evidence says and what it doesn’t

Recent studies and analyses make a few clear points we can act on. Work on interface designs that provide style subfloors shows one practical mitigation strategy: make diverse styles an explicit part of the interaction rather than leaving variety to chance. Empirical research on diverse AI personas demonstrates that design choices in the model layer and persona layer can reduce homogenisation effects. And critical commentary on frictionless design reminds us to be cautious about overgeneralising narrowly scoped studies to all professional contexts.

None of this proves that a single technical tweak will fix everything. Instead, the sociotechnical framing asks teams to combine social insight, interface design and infrastructure planning so that efficiency and shared tools do not come at the cost of local fit, resilience and opportunity.

Expanded from

This post expands on a point from Sociotechnical Systems in Software and AI: The Homogenisation Risk.

Sources