Published research on using language models to design dashboards is still early. The best-known open example, LIDA, presented at ACL 2023 by Victor Dibia, can summarize a dataset, propose chart goals, and generate chart code, and its author reported that 3.5 percent of generated charts failed to run in a benchmark of 57 datasets. That figure measures broken code, not whether a chart is accurate, readable, or accessible, so dashboard quality still depends on established design patterns and on checks such as the WCAG 2.2 contrast rules.
Key takeaways
- LIDA splits chart creation into four modules: a summarizer, a goal explorer, a visualization generator, and an infographic stage.
- The reported 3.5 percent visualization error rate counts code that failed to execute; it says nothing about chart correctness.
- A 2022 study of 144 dashboards grouped design choices into eight pattern groups, which gives a vocabulary for reviewing generated layouts.
- WCAG 2.2 success criterion 1.4.11 asks for a 3:1 contrast ratio for graphics needed to understand content.
- Generated dashboards still need human review of data, labels, and accessibility.
What LIDA does
The paper, LIDA: A Tool for Automatic Generation of Grammar-Agnostic Visualizations and Infographics using Large Language Models, was posted on arXiv on March 6, 2023, and accepted to the demonstration track of ACL 2023. It treats visualization as a chain of subtasks rather than one prompt. A summarizer turns a table into a compact text description. A goal explorer proposes questions worth charting. A visualization generator writes, runs, and filters chart code. An optional infographer stage uses image generation to produce stylized graphics.
The design is "grammar-agnostic," meaning the same pipeline can target different charting libraries. It ships as a Python library with an interface that mixes direct manipulation and natural-language requests. That is a research tool for generating individual charts and stories; the paper does not present a finished dashboard builder.
What the 3.5 percent figure measures
For evaluation, the author ran the pipeline on 57 datasets from the Vega datasets collection, asking for five goals per dataset. The reported visualization error rate was 3.5 percent, defined as the share of generated charts whose code failed to compile or run. The benchmark mainly used OpenAI's gpt-3.5-turbo models, and GPT-4 scored visual quality on a ten-point scale. The same paper notes that the 3.5 percent figure was measured without a self-repair loop that could lower failures further, but which was not evaluated.
A chart can run without error and still mislead: a truncated axis, a pie chart with 15 slices, or an average that hides a skewed distribution all execute cleanly. The metric therefore answers whether the code works, and a separate review has to answer whether the chart is right.
A vocabulary for reviewing a layout
Bach and colleagues, in Dashboard Design Patterns (IEEE VIS 2022, with a journal version in 2023), reviewed 144 dashboards and ran a two-week workshop with 23 participants. They started from 83 dashboards collected in earlier work, dropped three, and added 64, 36 of them related to Covid-19. The patterns fall into eight groups. Three describe content: data information, meta information, and visual representation. Five describe composition: page layout, screenspace, structure, interaction, and color. The paper also describes six dashboard genres, from static and analytic dashboards to magazine-style, infographic, repository, and embedded mini dashboards.
These groups give a reviewer concrete questions to ask of an AI-proposed layout. A reviewer can check whether labels and units are present (meta information), whether the page layout fits the intended genre, and whether filters behave as the interaction patterns suggest.
Accessibility checks that generated charts must still pass
The W3C's explanation of WCAG 2.2 success criterion 1.4.11, Non-text Contrast, sets a Level AA requirement: graphical objects needed to understand content, and the visual parts of user interface components, need at least a 3:1 contrast ratio against adjacent colors. For charts, the lines, bars, or slices needed to read the data should contrast with the background and, where they touch, with each other. If values are also given as visible text, the graphic itself may be exempt. A computed ratio of 2.999:1 fails.
A related criterion, 1.4.1 Use of Color, covers information carried by hue alone. A legend that distinguishes two lines only by red versus green fails for many users even when contrast is high. Neither criterion is checked by a language model unless a tester runs a contrast tool on the output.
Who checks what
| Dashboard task | What a language-model pipeline can propose | What a person verifies |
|---|---|---|
| Choosing questions to chart | Candidate goals from a data summary | Whether the goals match the audience's decisions |
| Writing chart code | Runnable code in a chosen library | Correct filters, aggregations, and axis ranges |
| Layout | An arrangement of charts | Fit with the genre and the pattern groups above |
| Color and contrast | A palette | 3:1 contrast and non-color cues |
| Data freshness | Nothing reliable | Source, refresh date, and definitions on the page |
The table is an illustrative division of work drawn from the sources above, not a result reported in any study.
Limits the LIDA authors state
The paper acknowledges that performance may fall for charting grammars that appear less often in training data, including graphical tools such as Tableau and Power BI. It also notes limits on tasks that need complex data transformations, the computing cost of running large models, the need for a sandbox when executing generated code, and the interpretability problems that come with language models. The author calls for broader benchmarks across more datasets and grammars.
FAQ
Can AI design a whole dashboard from a spreadsheet?
Research tools can generate individual charts and propose goals from a table. The sources reviewed here do not show a pipeline that produces a finished, reviewed dashboard without human checking.
Does a low error rate mean the charts are correct?
No. In the LIDA paper the rate counts code that failed to run. Correctness of values, labels, and scales needs a separate review.
Which standard covers chart accessibility?
The W3C's Web Content Accessibility Guidelines 2.2. Success criteria 1.4.11 (non-text contrast), 1.4.1 (use of color), and 1.4.3 (contrast for text) are the most relevant to charts.
Are Tableau and Power BI covered by these studies?
Not directly. The LIDA authors list graphical tools of that kind as a case where performance may be weaker, and their evaluation used code-based charting libraries.
Related reading
- How Netflix uses AI for personalization
- AI help with designing multiple-choice questions
- Artificial intelligence design and its implications
- AI in interior design
More in AI Tools and Industry.
Last verified: October 2026
Sources
- LIDA: A Tool for Automatic Generation of Grammar-Agnostic Visualizations and Infographics using Large Language Models, Victor Dibia, ACL 2023 demonstrations, arXiv:2303.02927
- Dashboard Design Patterns, Bach, Freeman, Abdul-Rahman, Turkay, Khan, Fan, and Chen, IEEE VIS 2022, arXiv:2205.00757
- Understanding Success Criterion 1.4.11: Non-text Contrast, WCAG 2.2, World Wide Web Consortium