How to develop and assess visual explanation in AI-enabled work

Start from the reader's comparison

The Office for National Statistics advises visualisation authors to decide which comparison matters, choose data and a chart type that reveal it, keep the display simple, and place the chart near related text. Its chart-text guidance also calls for concise titles, specific subtitles, sources, footnotes, and alternative text. W3C guidance for complex images recommends a short description plus a longer textual representation of essential information, including structure when it matters.

These principles shift the task from “make a chart” to “make this evidence understandable to this reader.” A visual is useful when its encoding, labels, context, and text alternative all serve the same explanation.

Use the CLEAR visual brief

Create a CLEAR brief before opening a chart tool or asking an AI assistant to design the visual:

| Field | Decision | Evidence to retain | | --- | --- | --- | | Conclusion | What should the reader be able to see, without implying more than the data supports? | One bounded takeaway sentence | | Lens | Which comparison, sequence, relationship, or distribution answers the reader's question? | Comparison choice and rejected alternatives | | Encoding | Which visual form represents the values honestly and with the least decoding effort? | Chart type, units, scale, and ordering | | Annotation | Which context, qualification, or event must appear near the evidence? | Title, subtitle, note, source, and limitation | | Reader test | Can a non-expert explain the takeaway and retrieve the values or logic without relying on colour or guesswork? | Test questions, text alternative, and revisions |

The brief can justify a table or short paragraph instead of a chart. That is a successful decision when the evidence contains only a few exact values or when the visual would add decoration rather than understanding.

AI can suggest chart types, labels, layouts, or alternative text. The worker should verify every value against the source, inspect the scale and ordering, and decide which comparison is appropriate. A fluent visual description is not evidence that the plotted data or implied conclusion is correct.

Worked example: a synthetic service backlog review

Imagine an operations team reviewing a synthetic twelve-week service backlog. The evidence contains weekly opened and closed case counts, a change in intake rules during week seven, and three service categories. This example is invented for instruction and does not describe Two Prune or a customer.

An AI assistant proposes a three-dimensional pie chart showing the latest week's category shares. The chart is attractive but does not answer the manager's question: did the backlog start falling after the intake change, and was the pattern broad or concentrated?

The worker completes CLEAR:

The final artifact has a takeaway title, a subtitle defining the measure and twelve-week period, a zero-based and labelled count scale, direct labels, a source note, and a concise annotation at week seven. The text alternative identifies the line chart and summarizes the fall in total backlog and the category exception. A linked data table preserves the exact weekly values.

The visual supports a discussion about where to investigate next. It does not claim that the rule change caused the decline. The separation between visible pattern and causal explanation is part of the skill.

• Conclusion: total backlog declined in the final four weeks, while one category remained above its week-six level.
• Lens: compare weekly backlog levels over time, with category detail separated rather than combined into shares.
• Encoding: use a line for total backlog and a compact companion table for the three category values at weeks six and twelve.
• Annotation: mark the rule change at week seven, state that the chart is descriptive, and note that other operational changes were not controlled.
• Reader test: ask a colleague to identify the direction, the persistent category, the event date, and the causal limitation using the visual and text alternative.

Assess the decisions behind the image

A useful assessment gives the participant a decision question, a small evidence pack, and permission to use AI support. Include tempting but unsuitable options, such as a prebuilt chart with a truncated axis, a data series with a different denominator, or a visual that relies only on colour. Ask for a visual explanation and a short design note.

Observe and score behaviors such as:

Retain the CLEAR brief, source extract, calculation or transformation notes, chart specification, text alternative, reader-test answers, and revision history. A polished image alone is insufficient because it hides whether the participant selected the right comparison or verified the data.

• the participant frames the reader question before selecting a format;
• the chosen comparison corresponds to the decision rather than the easiest available chart;
• values, units, dates, categories, and denominators match the source data;
• scale, ordering, colour, and annotations do not exaggerate or conceal important differences;
• title and surrounding prose distinguish observation from interpretation;
• essential qualifications appear near the relevant evidence;
• the text alternative communicates the visual's purpose and essential information;
• feedback from a reader test leads to a meaningful revision.

Verify at data, display, and interpretation levels

First, reconcile the data. Recalculate plotted values from the supplied source, confirm filters and missing values, and check that the chart's labels use the correct unit and period. Second, inspect the display. Verify the axis baseline where area or bar length encodes magnitude, test contrast and non-colour cues, and compare the visible order with the intended reading order.

Third, test the interpretation. Ask a fresh reader what the visual shows, what it does not establish, and where exact values can be found. Compare that response with the CLEAR conclusion. If the reader infers causation, certainty, or a population beyond the data, revise the title, annotation, scale, or surrounding explanation.

Visual explanation is task-dependent. One successful artifact does not establish broad design expertise or data-science competence. In a Two Prune assessment, the brief, workflow evidence, final visual, and verification steps can contribute bounded evidence about Evidence Navigation, Verification & Risk Control, Judgment & Synthesis, and Communication & Delivery. They do not produce a decision about employment or prove future performance.

Sources

Office for National Statistics data visualisation principles

• W3C Web Accessibility Initiative guidance for complex images

Sources: [1] [2]