I’m seeing more bad design in the AI era, but that is an observation—not proof that design quality is declining across the industry. AI use is widespread in surveyed product teams; that alone says nothing about whether the resulting interfaces are usable, accessible, faithful to a design system, or ready to ship. A Figma-to-code comparison is most useful when it tests those qualities separately rather than asking whether the output simply looks right.
AI adoption does not show that bad design is increasing
Figma’s 2024 survey polled nearly 1,800 designers and developers across four continents. Its FAQ says 59% of respondents were already using AI at work. But fewer than half of the respondents who used AI said they had launched anything, and only one-third of respondents who reported shipping an AI feature said they were proud of it. These are survey responses about use and sentiment—not an audit of interface quality or evidence of a year-over-year rise in bad design. Figma’s 2024 survey FAQ
As an Amazon Associate I earn from qualifying purchases.
Figma’s 2025 report landing page describes a second annual survey of 2,500 product builders across seven countries, covering agentic AI, differences in designer and developer workflows, and design principles. The detailed findings are behind a form on the page, so the sample size alone cannot support claims about what those respondents concluded. Figma’s 2025 report landing page
That distinction matters when describing a personal impression. More AI-generated screens may be visible, and faster generation may increase the volume of work people see, but neither fact establishes that the share of poor design has risen. To demonstrate a trend, one would need comparable quality measures over time—not just anecdotes, adoption rates, or examples of especially rough output.
#1 Best Overall
“Looks good” is only one measure of interface quality
A polished screen can still fail at its job. Visual or hedonic appeal is different from pragmatic quality: whether a person can understand the interface and complete the task. Other distinct checks include legible text, adherence to the brief, structural fidelity to a reference, accessibility, interaction behavior, and consistency with a design system. Combining these into a single “good design” judgment can hide meaningful trade-offs.
A study summarized by the Chartered Institute of Ergonomics & Human Factors explored burger-ordering interfaces generated with Midjourney, DALL-E 3, and Stable Diffusion 3. The systems initially struggled with legible text and following prompts; after prompts were adjusted, DALL-E 3 and Stable Diffusion 3 produced viable designs that met the brief. In a survey of 32 participants comparing the AI designs with commercial products and work by eight competent human UI designers, the researchers found no difference in pragmatic quality. The AI-generated designs received higher hedonic ratings than the commercial and human examples, while the commercial apps scored lowest on all measures. That is an interesting result for this task, not a general verdict on AI interfaces. CIEHF’s summary, published May 23, 2025
The same study tested AI evaluation of interfaces using UEQ-S prompts and found little correlation between AI ratings and human ratings. An evaluator that confidently calls a screen effective is not a substitute for observing people use it. Human evaluation and model-generated evaluation should be treated as separate evidence, not interchangeable scores.
What a Figma-to-code comparison should actually test
A Figma frame and a rendered implementation are not equivalent artifacts. The frame describes a visual state; code has to express layout rules, responsive behavior, components, and interactions across states and screen sizes. A useful comparison starts by specifying which of those properties matter and how they will be checked.
- Visual fidelity: Do spacing, hierarchy, typography, color, and assets match the reference at the same viewport and state?
- Structural fidelity: Does the implementation preserve the intended information order, component relationships, and interaction model?
- Responsive behavior: What happens when the viewport narrows, text wraps, or content grows?
- Usability and task support: Can a person understand what to do and complete the intended task?
- Accessibility: Are text contrast, keyboard access, focus states, labels, and semantic structure adequate? A screenshot cannot establish these.
- Design-system adherence: Are the right components, tokens, and rules used, rather than approximated by visually similar one-offs?
- Refinement cost: How much human correction is needed before the result meets the defined bar?
Keep observations separate from interpretation. If the comparison only checked a static screenshot, call the result a visual comparison; do not imply that it tested accessibility, usability, or production behavior. Likewise, a close match at one viewport does not show that an implementation will remain faithful across other sizes or states.
Generation speed does not remove the need for refinement
A 2026 peer-reviewed study by John Bustamante-Orejuela, Xavier Quiñonez-Ku, and Pablo Pico-Valencia examined undergraduate IT engineering students recreating mobile interfaces modeled on Duolingo’s interaction approach. It reported System Usability Scale scores of 82.86 for Figma, 67.14 for Uizard, 78.57 for Visily, and 80.36 for Stitch. Those scores belong to that study’s participants and task; they are not universal product rankings.
The study found that all four tools enabled rapid generation, but differences emerged in usability, structural fidelity, and perceived control. Uizard and Visily automated initial generation quickly, yet required further manual refinement for higher fidelity and customization. The authors emphasize that output quality depends in part on responsiveness, user control, iteration, and foundational design knowledge. Bustamante-Orejuela, Quiñonez-Ku, and Pico-Valencia, 2026
PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteThis helps explain why an impressive first pass is not the same as a finished interface. The relevant comparison is not just time to first output, but the output’s quality after a clearly described amount of correction—and what that correction involved.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Design-system compliance is a separate engineering problem
Google Research’s 2024 case study addresses the challenge of identifying and correcting violations of design-system guidelines. It describes a hybrid UI-linting pipeline that combines deterministic heuristics with the flexibility of large language models. Its conclusion is explicit: “Our case study demonstrates that AI alone is not sufficient for practical adoption and highlights the importance of a deep understanding of AI capabilities and user-centered design approaches.” Google Research’s UI-linting case study
The practical implication is not that every AI-generated interface will violate a system. It is that a visual resemblance to the Figma reference does not prove compliance. Checking component choices, tokens, and guidelines requires explicit rules and inspection; a hybrid approach can combine repeatable checks with human judgment where rules are ambiguous.
Connecting design context to AI can help, but it is not production validation
Google Research’s PromptInfuser study looked at a Figma widget that connected interface elements to LLM prompt inputs and outputs to create semi-functional mockups. Fourteen professional designers compared that connected workflow with their disconnected workflow. Participants felt the connected version communicated a product idea better, stayed closer to the envisioned artifact, and helped them anticipate UI issues and technical constraints. Google Research’s PromptInfuser study
Recommended Free Tools
That finding supports the value of keeping product context connected to interface work. It does not show that a Figma-connected workflow automatically produces production-ready code, preserves all design-system rules, or replaces implementation and user testing.
How to make a Figma ↔ Code experiment informative
For the result to mean more than “this screenshot looks off,” describe the inputs and the checks. A reader should be able to tell what was compared, under what conditions, and what “bad” meant in this instance.
Quick Recap
- Record the input: Identify the Figma frame or flow, relevant design-system constraints, target viewport, and any states or interactions included.
- Describe the route: Name the code-generation method, tools, prompts, and context supplied. Distinguish generated code from code edited by a person.
- Set criteria before judging: Separate appearance, fidelity, task support, accessibility, responsive behavior, design-system adherence, and refinement effort.
- Compare like with like: Use the same screen size, content, and state for visual comparisons; test additional sizes or states separately.
- Report what was checked: Say whether accessibility or usability was actually tested, and how. Do not claim either based on visual inspection alone.
- Show the corrections: Note what failed, what worked, and what a human had to change. This reveals whether speed at generation came with a substantial refinement burden.
- Limit the conclusion: State whether the result describes one artifact, one workflow, or a broader pattern. A single experiment can illuminate a process; it cannot by itself establish an industry-wide trend.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
