TL;DR
Paste your chart (image or table) into ChatGPT, Claude, or Gemini and ask for three captions — descriptive, interpretive, actionable — plus the single most likely misreading. Pick the caption that matches your audience, keep the data caveats visible, and never let the model invent a trend that is inside the noise. The copy-ready prompt below does all of this in one pass.
One thing to fix first: when you upload an image of a chart, the model estimates values from pixel positions — it does not read them exactly. So feed it the underlying numbers when you have them, and treat any value it “reads” off an image as approximate.
The task
You have a chart or table the audience could misread. You want three caption variants so you can pick the right register for the context, and a “common misreading” callout that stops the wrong conclusion before it forms. Charts without that guard turn into arguments; charts with it turn into decisions.
When AI helps — and when it does not
AI is genuinely good at naming the dominant pattern, varying caption registers, and predicting the most likely wrong reading. It is poor at two things: deciding which register your audience needs (descriptive for analysts, actionable for execs, interpretive for board decks), and pulling exact numbers off a chart image.
That second limit is the one people forget. Every current frontier model reads a chart picture by estimating positions on the axes, so a bar that is “about 47” can come back as 45 or 50. For headline accuracy, paste the table or the source numbers as text. Reserve image upload for when you only have a screenshot.
Which model to use (June 2026)
All three flagships accept an image plus text and return solid prose captions. They differ on raw chart-reading skill and on how much underlying data you can paste in one go.
| Model | Best for here | Image input | In-app context | Plan |
|---|---|---|---|---|
| Gemini 3.1 Pro | Charts and mixed media (leads CharXiv / MMMU-Pro chart benchmarks) | Native multimodal | 1M tokens | Google AI Pro $19.99/mo |
| Claude Opus 4.7 | High-res financial charts and dense tables | Image + text | 1M tokens | Claude Pro $20/mo (Max for volume) |
| GPT-5.5 | Default in-chat workflow, OCR on messy screenshots | Image + text | ~320 pages on Plus (full 1M on $200 Pro) | ChatGPT Plus $20/mo |
Practical read: if the input is a picture and accuracy matters, Gemini 3.1 Pro has the edge on chart understanding. If you are pasting a large table as text and want careful interpretation, Claude Opus 4.7 or GPT-5.5 are both fine. The prompt below is identical across all three.
What to feed the AI
- Chart description: axis, units, time period, segmentation
- The data points as text (preferred) or a screenshot
- Audience and their decision context
- What “good” and “bad” look like in your business
- Any known data caveats (missing weeks, methodology change, segment exclusion)
- Length target (one sentence vs one paragraph)
Copy-ready prompt
Explain this chart / table in three caption variants.
Chart description: [axis, units, time period, segmentation]
Audience and decision context: [line]
"Good" vs "bad" in our context: [line]
Data caveats: [list]
Length target: [one sentence / one paragraph]
Data (paste the numbers as text when you have them):
"""
[paste data or table]
"""
Return:
1. Three captions:
- Descriptive — what the chart shows
- Interpretive — what it means in business terms
- Actionable — what to do with it
2. The single most likely wrong reading — with one sentence debunking it
3. A "watch out" — caveats that should appear near the chart
4. A confidence rating per caption (1-5) based on data quality
5. The caption I should ship to this audience
If any number is read from an image rather than the pasted data, flag it as approximate.
Do not re-state cell values. Do not say "as you can see."For executive decks, add: “Include an ‘if challenged’ line — the one-sentence defense if a stakeholder pushes back on the interpretation.”
Recommended output structure
Three labelled captions (descriptive / interpretive / actionable), then the misreading, the watch-out, and the recommended caption for this audience. Keep the data caveats glued to the chart — readers quote charts without them.
How to check the output is usable
- Descriptive caption says what the chart shows without interpreting it
- Interpretive caption ties to your business “good vs bad”
- Actionable caption names a specific action, not “monitor closely”
- The “common misreading” is a real risk, not boilerplate
- Confidence rating tracks data quality (small sample → lower)
- The recommended caption matches the audience
- Any image-read number is flagged approximate, and you spot-checked it against the source
Common mistakes
- Captions that re-state the data (“Q1 was 12; Q2 was 13”). Useless
- No caveats. Your reader gets quoted on the number including the gap you knew about
- Missing the “so what”. Interpretation with no action
- One caption only. Different audiences need different registers
- Trusting an image-read value as exact. Paste the numbers when accuracy matters
- Letting AI call noise a trend. Ask for a confidence rating
FAQ
- Can I just upload a screenshot instead of typing the numbers? Yes, but values come back approximate — the model reads pixel positions, not data. Paste the underlying numbers whenever the exact figure matters; use image-only for quick reads.
- Which model is best for chart images? As of June 2026, Gemini 3.1 Pro leads published chart-reading benchmarks (CharXiv / MMMU-Pro). Claude Opus 4.7 and GPT-5.5 are close and fine for text tables.
- Why does the model “lose” my chart on follow-up questions? After the first answer, vision models sometimes stop referencing the uploaded image and answer from memory. Re-state “re-read the chart above” in the follow-up to pull it back.
- Should I include the chart as well as the caption? Yes — caption above, chart below. Most readers stop at the caption.
- What if a table cell contradicts the caption? Surface it. A visible contradiction is far better than a hidden one a reader finds later.
- Can AI fact-check the chart? It can flag internal inconsistencies; it cannot verify against your source data. That part is on you.
Related
- Chart takeaway — broader chart takeaway pattern
- Table explanation — for tables specifically
- Spreadsheet error diagnosis prompts — when the table itself looks wrong
- Excel analysis prompts — deeper analysis on the underlying data
- AI table explainer tutorial — full workflow
- Business data analysis AI — when the chart drives a business decision
- Experiment interpretation — for A/B chart readouts