Design systems

A design language is the set of choices that gives a family of interfaces one look and behaviour: colours, type, spacing, components, patterns and the principles behind them. A design system is usually the maintained, documented form of it. Reviews of the literature describe design systems as component libraries, design guidelines and style guides (Lamine & Cheng, 2022), and distinguish them from style guides and pattern libraries by their usually carrying the identity of the organisation (Lamine & Cheng, 2022). Practitioners dispute whether to treat consistency as the goal and how much freedom a system should leave. Controlled evidence for its effects is thin: a few small studies find faster building, studies of adoption report that mature systems are used less than their owners hoped, and almost nothing measures what end users gain. Generated interfaces add a new question, whether a model given a design language follows it, and whether its output looks alike regardless.

History

The research did not cover the origin of style guides, pattern libraries and design systems. The sources describe the present practice, not how it arose.

Core concepts

Terms and contents

The Nielsen Norman Group defines a design system as a living set of standards and a style guide as documentation of their implementation (Gordon, 2024). A design token is information associated with a human-readable name, at minimum a name and value, according to the W3C Design Tokens Community Group draft format (Chenais et al., 2026).

A multivocal mapping of 23 sources, 17 of them gray literature, identified four types of elements, three resources, seven developer needs, four solutions and seven impacts of using a design system (Vendramini et al., 2021).

Deriving a language from what an organisation wants to express

A design language is meant to express something: a brand, a tone, values. Some mappings from attribute to visual choice have support. Colour-emotion associations show a universal pattern across 30 nations (average r = .88), with local differences larger between nations that are linguistically or geographically distant (Jonauskaite et al., 2020). Typefaces are perceived to carry personality, mostly bold display and script faces, while serif versus sans serif is neutral (Shaikh & Chaparro, 2016), and Arabic typefaces each produce a distinct pattern of personality ratings from skilled readers, consistently across participants (Jordan et al., 2017).

Whole designs are less well tested. In one test with manipulated web pages, simplicity and cohesion established a bold personality, and contrast, density and regularity related to an analytical one, but the expected effects on a sophisticated personality were not observed (Park, 2006). Design look was the feature mentioned most often (46.1% of comments) when 2,684 people judged the credibility of two live sites; this is a frequency of self-reported comments, not a measured effect size (Consumer Reports, 2002).

Appeal is partly predictable. Websites with low visual complexity and high prototypicality (looking typical for their kind) were rated highly appealing (Tuch et al., 2012). Perceived complexity and colourfulness, combined with demographic variables such as education and age, explain about half the variance in appeal ratings given after 500 ms (Reinecke et al., 2013), and judgments formed that fast later affect opinions of usability and trustworthiness (Reinecke et al., 2013). The preferred level of colourfulness and complexity differs by group, with females liking colourful sites more than males and higher education lowering the preference, in 2.4 million ratings from nearly 40 thousand participants (Reinecke & Gajos, 2014). Japanese participants were slower than US participants at finding information on websites, more so on complex sites, though the hypothesised East Asian advantage for contextual information was not supported (Baughan et al., 2021).

Constructing a palette

Evidence on palettes comes mostly from charts, not interface palettes. Accuracy of reading categorical colour encodings falls as the number of categories rises from 2 to 10, and palette discriminability also matters (Tseng et al., 2023). Redundant colour-and-shape encoding improves accuracy, most for 5 to 8 categories, though colour and shape interact so the pairing matters (Tseng et al., 2026). Measured colour-difference perception varies by mark type, so a perceptual space does not by itself guarantee that small interface marks are discriminable (Szafir, 2018).

For colour vision deficiency, most tools only simulate it and few help designers generate accessible palettes, while guidelines stress contrast ratio and overlook other advice on colour use (Geddes et al., 2025). The two main contrast measures, the WCAG 2 ratio and APCA, disagree on a large share of random colour pairs, neither had been validated against user legibility data in that article, and both ignore the page background (Waller, 2022). The WCAG 4.5:1 threshold is derived by multiplying a 3:1 baseline by a contrast-sensitivity loss factor of about 1.5 for 20/40 acuity, not from a test of reading performance on interface palettes (W3C Web Accessibility Initiative, 2026b). In a user study of colour schemes for coloured text at fixed contrast, non-analogous and yellow-inclusive schemes improved information-finding performance and yellow-inclusive ones were preferred (Ng et al., 2024).

Light and dark variants have conflicting evidence. In a study of 173 participants cognitive test scores were higher in light mode than dark mode (Gazit et al., 2025), whereas on dashboards dark mode improved accuracy and confidence for medium-complexity tasks and lowered perceived workload (Ettling et al., 2025). Measured search times on coloured backgrounds did not track what participants said they preferred (Hill & Scharff, 1999).

Constructing a type scale

A review of 42 studies finds serifs are not a significant legibility factor, no single size or typeface suits every reader and situation, and familiarity may matter (Ho Sang & Petrarca, 2025). A second review likewise concludes there is no most legible typeface; it depends on the reading situation (Beier & Thiessen, 2025). Size matters more: fluent reading covers a tenfold range of print size, and below a critical print size reading speed declines sharply (Legge & Bigelow, 2011). Evidence does exist for effects of letter design: letter width raised acuity for all groups in a low-vision experiment (Mangas, 2024), and a font designed for peripheral vision reduced perceptual errors in peripheral crowded letter recognition (Bernard et al., 2016). Across individuals, different fonts gave faster reading, with about a 35% gap between a person’s fastest and slowest font, and personal preference did not predict speed (Cannon, 2022); results varied so much that no single font suited all readers (Readability Matters, 2022).

For hierarchy, size is the strongest cue to a heading’s level, and level differences of about 20% in relative size are more discriminable than absolute differences (Williams & Spyridakis, 1992). Headings that combine bold with a larger size or a sans serif typeface helped readers find information faster and more accurately, in familiar or unfamiliar text (Timpany, 2018). A paper on automatic font pairing reports only that experiments and user studies demonstrate the effectiveness of its methods (Jiang et al., 2018).

Spacing and layout

Wider line spacing improved the accuracy and speed of visual search in web pages, and left-aligned text performed better though users preferred justified text (Ling & van Schaik, 2007). In mock-up experiments uniform density improved and variable density worsened performance, while alignment did not improve performance (Parush et al., 2005).

Documentation

Documentation is where a system meets its users, and the studies are about failures. At Xero, designers and developers had trouble finding their way around the design system guide, and the navigation based on atomic design did not match how they searched, so the team flattened it and added search and related links (Sanderson-Harris, n.d.). Among 321 issues about documentation in open-source design systems, contributors mostly reported missing documentation, errors or requests to improve a component’s documentation; documentation made up only about 10% of issues (Lamine & Cheng, 2022). A system vendor describes “most visited pages” analytics but publishes no usage figures by section (Mansour, 2026). The nearest measured evidence is on API documentation: in an observation of 11 developers the API reference took about 18% of time, with recipes and samples next (Meng et al., 2019).

Following a design language: people and AI agents

Models given design requirements often miss them. Over a quarter of the design rationales that generative interface tools state are not implemented in the generated interface, and four of five tools implemented 6% or fewer functional UX principles stated in prompts (Imteyaz et al., 2026). Good visual reconstruction by multimodal models does not imply code that meets implementation requirements, and eight models showed substantial gaps in generating UI code that satisfies them (Shi et al., 2026). A structured, parameterised specification captured reference design intent more faithfully than prompt-based baselines (Chen et al., 2025b). Machine checking helps in one case: using automatic WCAG audit results as a reward in post-training cut the inaccessibility rate of generated pages by 87.5% and 58.1% under two auditors, in a single generation pass (Yoon et al., 2025). In a controlled industry experiment, design-system-aware AI assistance cut delivery time compared with manual or design-system-only work (Silva et al., 2026).

Questions, controversies, and new developments

Does a design system speed work and improve consistency?

Developers coded a simple form faster with the Carbon design system than from scratch, with a median of 2 hours versus 4.2 hours in an eight-developer study (Sparkbox, n.d.); the study calls itself small and of excellent developers, and its consistency and accessibility rankings were subjective (Sparkbox, n.d.). A thesis found that efficiency, consistency and learning vary together with how firmly design knowledge is externalised in a maintained, shared artefact (Gjidoda, 2026).

Adoption and the cost of maintenance

Having components does not produce consistent use. At SAP, adoption of the mature Fiori design system stayed limited across teams because of knowledge gaps, slow evolution, lack of flexibility and inefficiencies, in 76 stakeholder interviews; capacity and skill gaps in maintaining components remained a challenge even after shared ownership improved motivation (Richter et al., 2025). A 2026 vendor survey reports adoption as the top challenge for five years, and that most design system teams are understaffed, with buy-in satisfaction falling from 42% to 32% in a year (zeroheight, 2026). In a 2018 vendor survey, the 22 respondents who rated their system average or unsuccessful most often named a lack of an executive champion, and half named adoption and staffing difficulties (Sparkbox, 2018). Interviewed system leaders reported considerable difficulty evolving their systems (Lamine & Cheng, 2022). Production sites also drift from their tokens: in a vendor audit of 375 sites, average token coverage was 40.4% (median 38%) and 7.5% of sites reached 90% or more (OverlayQA, 2026), though the audit’s selection and detection methods are not described.

Consistency against flexibility

Grudin (1989) argues that consistency is an unreliable guide for interface design and that designers would often do better to focus on users’ work environments; this is an argument, not a measurement Grudin (1989). Practitioners interviewed by Lamine and Cheng report that balancing customisability and standardisation, and keeping a system relevant as products evolve, are difficult (Lamine & Cheng, 2022), and a project leader reported complaints from people who thought the system limited them (Lamine & Cheng, 2022). In an interview study one designer said a corporate design system limited them to predefined buttons and styles, and another argued with the design system manager to add a component (Shokrizadeh et al., 2025).

Do generated interfaces all look alike?

Across five generative UI tools, interface outputs converge in visual appearance and layout, with more variation in colour (Imteyaz et al., 2026). In a benchmark of 15 models on 80 briefs, generated UIs were substantially more repetitive than a human-designed baseline while meeting nearly all brief requirements (Rusia et al., 2026). In text design problems, human-crowdsourced solutions were more diverse than those of language models in all topics tested (Ma et al., 2024b). Against this, a study of 73 vibe-coded homepages found residual variety and some builders escaping the dominant styles, without a human-made comparison (Boussioux et al., 2026), and human-made web design also homogenised over a decade, with layout distance down 44% from 2010 to 2019 in a secondary summary of the original study, and visual similarity strongly correlated with use of similar libraries (Kolko, 2025). Sameness is not unique to generated interfaces. A paper argues generative web design risks homogenised output and proposes productive friction to help creators challenge defaults (Shin et al., 2026).

What reduces sameness is barely tested on interfaces. In the non-UI study, a critique-based prompt gave the highest diversity among prompt methods, and higher temperature and top-P did not reliably raise it (Ma et al., 2024b). The one UI benchmark ran all models on the same prompt with default sampling, so it tested no diversity-seeking prompt (Rusia et al., 2026).

How reliable is the colour-emotion link?

A systematic review of 132 studies finds people reliably link colours to emotions but not everybody shares the same correspondences (Jonauskaite & Mohr, 2025); individual differences explain variation beyond culture, and correspondences would likely change in specific contexts (Jonauskaite & Mohr, 2025). Many studies do not specify colours with perceptually accurate models, so findings cannot be reproduced exactly (Jonauskaite & Mohr, 2025). In the 30-nation study the participant’s country predicted associations beyond the universal pattern, in a task pairing 12 colour terms with 20 emotion concepts rather than interface colours (Jonauskaite et al., 2020). Implicit measures show a culture difference in the red-green valence association that explicit ratings did not predict, but none in a red-white pair (Kawai et al., 2022).

Do designers and laypeople agree on appeal?

In a small EEG study, design experts rated as unaesthetic more often than laypersons the webpages that a separate group had rated aesthetic (Bölte et al., 2017). A painting study (not of interfaces) found that preference for novel over familiar items was stronger for complex items and modulated by art experience (Song et al., 2021).

Reviewing and critiquing designs

Critique studies come from visual design education and crowdsourcing. A rubric of expert criteria made novice critiques rated as valuable as expert critiques (Yuan et al., 2015); expertise changed critique style only by increasing length, though experts may add clearer justifications (Yuan et al., 2015). Organising critique around design principles let novices produce more critiques than freeform, but two variants of the system did not differ in quality or agreement (Wu et al., 2013). Student designers found feedback most helpful when emotional, positive and specific (Yuan et al., 2015). Aggregated crowd critique matched 45 to 60% of expert critiques but did not significantly change design quality compared with generic feedback, in a course discussion summary of the original paper (CS 6724 course discussion post, 2015).

For model critique, GPT-4 heuristic feedback is useful for subtle errors, text and semantics but its utility fell over iterations (Duan et al., 2024a). Human evaluators had higher precision on guideline violations with similar F1 (0.478 for humans against 0.466 for GPT-4) (Duan et al., 2024a). Of 100 distinct violations, GPT-4 found 9 alone, 29 together with humans, and humans alone 62; humans found more global violations and understood context better, while GPT-4 was more thorough and specific (Duan et al., 2024a). Current LLM critique techniques do not match human evaluators; few-shot and visual prompting with a critique dataset gave a 55% gain (Duan et al., 2024b).

Automated audits against a guideline set find violations at scale. In 60,756 Android UIs from 9,286 apps, 7,497 UIs from 2,587 apps had at least one Material Design guideline violation, found by a detector with precision 0.81 and recall 0.90 (Yang et al., 2021a). A study of mock-ups against implementations found 82 design violations in 71 screen pairs, mostly layout, resource and text, and its checker is used by over a thousand designers and developers at Huawei (Moran et al., 2018).

Broader connections

Design systems set the constraints within which interface design choices are made, and the measured effects of layout, colour and type reported there bear on what a system should fix. Design process covers how an interface is built, where a design language is one input among others. Whether an interface is usable, as opposed to consistent, is settled by methods such as usability testing. The question of how closely AI agents follow a design language and whether their output converges is taken further in AI-generated interfaces.

Further reading

• Lamine and Cheng’s study of design systems practice, an interview and issue study of how systems are made and where they struggle (Lamine & Cheng, 2022).
• Gordon’s Nielsen Norman Group note on design systems versus style guides, for definitions (Gordon, 2024).
• Richter, Rüssel and Karczewski on adoption of SAP’s Fiori system, a case of a mature system with limited uptake (Richter et al., 2025).
• Waller on whether contrast ratios predict legibility, for the evidence behind accessibility thresholds (Waller, 2022).
• Duan and colleagues on GPT-4 feedback on UI mock-ups, for model critique compared with human evaluators (Duan et al., 2024a).
• Jonauskaite and Mohr’s review of colour and emotion research (Jonauskaite & Mohr, 2025).