03 · What You Need to Know
How AI Can Support Accessibility in Scientific Figures
What Makes a Research Figure Accessible?
An accessible research figure communicates its essential information to readers who cannot perceive or interpret the visual presentation in the usual way.
This includes readers who use screen readers, have low vision, experience color-vision deficiencies, or encounter figures through alternative presentation formats.
Accessibility is not achieved merely by attaching a text description to an image. The description must communicate information relevant to the figure's purpose, while the surrounding document should provide appropriate structure and context.
For example, a scatterplot may require an explanation of the variables, the general pattern of observations, and any scientifically important features that are not otherwise available in the text.
Some figures also benefit from accessible data tables, high-contrast visual encodings, or alternative representations.
The World Wide Web Consortium's Web Content Accessibility Guidelines (WCAG) provide a general framework for text alternatives to non-text content. The appropriate implementation depends on the publishing format and the complexity of the figure.
What Is the Difference Between Alt Text, a Caption, and a Long Description?
These elements have overlapping purposes but should not automatically contain identical information.
| Element |
Primary purpose |
Typical content |
| Short alternative text |
Provides a concise text alternative identifying the image's essential purpose or content |
Figure type, subject, and principal information not already conveyed nearby |
| Figure caption |
Identifies and explains the figure within the scholarly document |
Variables, methods, abbreviations, statistical information, and relevant context |
| Long description |
Communicates complex visual information that cannot be adequately represented in brief alt text |
Graph structure, axes, series, relationships, patterns, and important numerical details |
| Accessible data table |
Provides structured access to numerical observations or estimates |
Relevant values organized with meaningful row and column headings |
A caption does not automatically replace an accessible description. Likewise, repeating the entire caption in an image's alt text may create unnecessary duplication for screen-reader users.
Researchers should consider what information is already available in the surrounding text and what the figure contributes beyond that information.
Why Is Short Alt Text Often Insufficient for Complex Scientific Figures?
A simple image may be adequately described in one sentence. A complex scientific visualization may contain several variables, multiple panels, statistical estimates, and important relationships.
Attempting to include all this information in one alt-text attribute can produce an excessively long description that is difficult to navigate.
For example, a forest plot containing twelve regression estimates may require explanations of the reference line, effect measure, confidence intervals, and relationships among predictors.
A short alternative can identify the figure and direct readers toward a detailed description available in the document.
The detailed description can then organize the information into meaningful sections or paragraphs.
The appropriate solution depends on the publishing platform and its support for accessible content. A detailed description should be programmatically available or clearly linked through an accessible mechanism rather than existing only as visually displayed text that assistive technologies cannot reach.
What Can AI Contribute to Writing Accessible Descriptions?
Multimodal AI can examine a figure and propose text describing its visible structure.
It may identify chart types, axis labels, legends, color categories, panel arrangements, and apparent trends.
For example, AI might recognize a scatterplot and draft a description explaining which variables appear on each axis and whether the observations show an upward, downward, or unclear relationship.
It can also help reorganize a long technical description into a logical reading sequence.
These capabilities may reduce the initial drafting effort, especially for researchers who are unfamiliar with accessibility writing conventions.
However, visual recognition and scientific interpretation are not equivalent. A model may correctly identify a graph type while misreading a numerical value or misunderstanding the statistical meaning of an interval.
AI-generated descriptions should therefore be treated as drafts rather than independently verified accessibility content.
How Should an Accessible Graph Description Be Organized?
A useful description generally begins by identifying what the figure represents, then explains the structure necessary to understand the information.
For a graph, this may involve identifying the horizontal and vertical variables, measurement units, graphical series, and important patterns.
More detailed descriptions may explain notable outliers, uncertainty intervals, reference lines, or differences between groups.
The order should support comprehension rather than mechanically list every visual element.
For example, describing a regression coefficient plot may be clearer when the effect measure and reference line are explained before individual estimates are discussed.
Researchers should avoid turning accessible descriptions into exhaustive verbal inventories of decorative details that contribute little to understanding.
The objective is equivalent access to meaningful information, not a word-for-word conversion of every visible feature.
How Can AI Describe Multi-Panel Research Figures?
Multi-panel figures require careful organization because different panels may represent separate outcomes, samples, or analytical procedures.
AI can help produce a structured description that introduces the overall figure before explaining each panel in sequence.
However, it may confuse panel labels, combine results from different analyses, or overlook differences in axis scales and statistical measures.
Researchers should verify the description of each panel separately.
Where the figure contains substantial information, a long description with distinct paragraphs or headings may be more accessible than one uninterrupted block of text.
The relationship between panels should be explained only when supported by the figure caption or research context.
Why Must Accessible Descriptions Preserve Scientific Meaning?
An accessible description should communicate the evidence represented in the figure without exaggerating, minimizing, or changing its meaning.
Suppose a graph shows an association between online learning engagement and assessment performance. An AI-generated description stating that engagement causes higher performance would introduce a conclusion not established by the graph alone.
Similarly, a description claiming that one group performed significantly better than another would require appropriate statistical evidence.
Researchers should distinguish observed graphical patterns from inferential interpretations.
The limitations involved in AI explanations of complex research figures are relevant here, but accessible descriptions have an additional responsibility: readers may depend on the description as their primary access to the visual information.
Can AI Misrepresent Numerical Values in Accessible Descriptions?
Yes. Exact values can be difficult to extract from rendered figures, particularly when labels are small, scales are transformed, or observations overlap.
AI may estimate a value visually and present the estimate as an exact measurement.
This is problematic when numerical precision affects the scientific interpretation.
For example, a coefficient of 0.08 should not be reported as 0.8 because of a misread decimal point. Likewise, a confidence interval should not be described as excluding zero unless that conclusion is supported by the actual interval.
Researchers should obtain exact numerical information from the underlying data or verified analytical outputs whenever possible.
For detailed scientific descriptions, supplying AI with a verified table of plotted values may be more reliable than asking it to reconstruct every number from an image.
How Should AI Describe Statistical Uncertainty?
Scientific figures often communicate uncertainty through error bars, shaded regions, or intervals around estimates.
Accessible descriptions should identify these elements and explain their meaning when that information is relevant.
However, the description must use the correct terminology.
Standard deviations, standard errors, confidence intervals, credible intervals, and prediction intervals are not interchangeable.
AI may incorrectly label an interval or overinterpret whether two estimates differ.
Researchers should verify the interval definition against the figure caption and statistical methods before incorporating it into an accessible description.
Where exact interval values are central to interpretation, an accessible table may provide a clearer alternative to a lengthy verbal enumeration.
Can AI Create Accessible Descriptions of Heatmaps?
Heatmaps can be difficult to describe because they encode many values through color.
An accessible description may need to explain the row and column variables, the meaning of the color scale, and the most important patterns.
For example, a correlation heatmap might show positive and negative relationships among several questionnaire dimensions.
However, the description should not assume that a particular color represents a specific numerical meaning without examining the legend.
It should also avoid implying statistical significance unless significance is explicitly represented and supported by the analysis.
For dense heatmaps, a structured data table or machine-readable matrix may be necessary to provide access to individual values.
AI can help summarize the overall pattern, but a summary alone may not provide equivalent access when readers need to inspect specific cells.
Can AI Describe Scientific Diagrams and Conceptual Frameworks?
Yes. AI can help identify the components of a diagram and describe their spatial or logical relationships.
For example, an educational technology framework may show learner characteristics, instructional design features, engagement, and academic outcomes connected by arrows.
An accessible description should identify the major components and explain what the connections represent.
However, arrows can have different meanings. They may indicate hypothesized causal relationships, information flow, temporal order, or conceptual associations.
AI should not infer that every arrow represents an empirically demonstrated causal effect.
Researchers should provide the diagram's intended interpretation and verify that the description distinguishes proposed relationships from established findings.
How Can AI Help With Accessible Tables and Alternative Representations?
Some visual information is more effectively communicated through structured data than through prose alone.
For example, a figure containing numerous model estimates may be accompanied by an accessible table listing the variables, estimates, and uncertainty intervals.
AI can help organize verified values into a proposed table structure, but the numbers and labels must be checked against the analytical output.
Tables should use meaningful headings and a logical reading order so assistive technologies can communicate their structure.
Researchers should avoid replacing every visualization with a table automatically. Graphs can communicate patterns and relationships that are cumbersome to reconstruct from numerical lists.
Accessible research communication may require complementary representations rather than a single universal format.
Should AI Describe Every Visual Detail?
No. An effective description prioritizes information necessary to understand the figure's purpose and scientific content.
Decorative features may not require description unless they carry meaning.
For example, the precise shade of a line may be less important than identifying which experimental condition the line represents.
However, color distinctions must be explained when they encode categories or numerical values.
Researchers should consider what a reader needs to know to understand the evidence, compare relevant quantities, and follow the surrounding discussion.
Excessive detail can make a description difficult to navigate, particularly when important findings are buried among minor visual observations.
Can AI Make a Figure Accessible Simply by Generating Alt Text?
No. Alt text is only one component of accessible figure presentation.
Other considerations include readable typography, sufficient contrast, meaningful color choices, accessible document structure, clear captions, and appropriate alternatives for complex visual information.
A figure may have accurate alt text while remaining difficult for readers with low vision because its labels are too small or its colors are indistinguishable.
Likewise, a detailed description may be unavailable to screen-reader users if it is inserted into an inaccessible document or disconnected from the figure.
Researchers should evaluate both the content of the description and how readers access it.
The more specific task of generating accurate alt text for scientific figures requires attention to brevity, purpose, and the distinction between alternative text and longer explanations.
How Should Researchers Verify AI-Generated Accessible Descriptions?
Verification should involve comparing the description with the actual figure, caption, and relevant analytical evidence.
Researchers should confirm that axes, variables, groups, units, patterns, and statistical quantities are represented correctly.
Claims involving exact numerical values or statistical conclusions should be checked against the source data or verified outputs.
Accessibility review should also consider whether the description is understandable without seeing the figure and whether its organization supports navigation.
Where feasible, feedback from people who use screen readers or other assistive technologies can reveal practical accessibility problems that a visual review may overlook.
Automated accessibility checks are useful, but they cannot fully establish the quality of a complex scientific description.
Does AI Accessibility Assistance Need to Be Disclosed?
Disclosure requirements depend on the journal, publisher, institution, and role of AI in preparing the material.
Some publishers have policies addressing generative AI assistance in manuscript preparation or image-related content. These policies may differ in scope and may change over time.
Researchers should consult the current requirements applicable to their publication rather than assume that all AI-assisted accessibility writing is treated identically.
Regardless of disclosure requirements, the authors remain responsible for ensuring that the accessible description accurately represents the research figure.
Accessibility should not become an exception to ordinary standards of scientific accuracy.
