Writing figure captions and lab-meeting slide narratives with AI

A quick primer before the walkthrough: A good caption lets someone understand a figure without the talk. AI is useful for turning your rough description into that caption — as long as you supply the facts and it invents none.

In plain English: You tell the assistant what the experiment was and what the figure shows. It writes the caption and a short spoken narrative. You check that it did not add a conclusion you did not make.

Who this is for: Students and scientists preparing their first lab meeting or group slides, and anyone whose captions currently read "Figure 3. Results."

What you need: Any general AI assistant, and your own description of the result. You do not upload the figure or any unpublished image — you describe it in words.

Comfort level: No coding.

Key terms (defined again at the end): n, error bars, replicate, effect size.

Step 1 — Write the five facts a caption needs

What this step does: Forces you to have the information before you ask for prose.

Every caption needs:

1. What was measured and by which method.
2. The system — cell line, strain, tissue, or model.
3. The conditions compared, including the control.
4. n — how many independent replicates, and what the error bars represent.
5. What the reader should see in the figure.

If you cannot fill one of these in, the caption is not the problem — the figure is.

Step 2 — Draft the caption

What this step does: Converts your five facts into standard caption form.

Role: You are a scientific editor writing figure captions.
Task: Write a figure caption from the facts below.
Rules:
- Use only the facts I provide. Do not add interpretation, mechanism, or significance.
- State the method, the system, the comparison, n, and what the error bars show.
- If a fact is missing, end with "Missing: ..." instead of guessing.
- Two to four sentences, past tense.

Facts:
- Measurement: [ ]
- System: [ ]
- Conditions compared: [ ]
- n and error bars: [ ]
- What the figure shows: [ ]

The "do not add interpretation" rule is what keeps a caption from quietly becoming a claim.

Worked example (illustrative data)

Facts supplied: absorbance-based viability assay; HEK293 cells; compound X at 0, 1, 10 uM versus vehicle control; n = 3 independent experiments, error bars are standard deviation; viability decreases at the highest concentration.

Caption produced:

Figure 2. Viability of HEK293 cells after 24 h exposure to compound X. Cells were treated with vehicle or compound X at 1 or 10 uM and viability was measured by absorbance-based assay. Bars show the mean of three independent experiments; error bars show standard deviation. Viability was lower at 10 uM than in the vehicle control.

All numbers here are illustrative. Note what the caption does not say: it does not say the compound is cytotoxic, and it does not claim significance, because neither was supplied.

Step 3 — Get the three-sentence slide narrative

What this step does: Gives you words to say while the slide is up.

Using only the caption above, write three sentences to say aloud:
1. Why we did this experiment.
2. What the figure shows.
3. What we will do next.
Rules: no claims of significance or mechanism unless I stated them.
Add the question a reviewer is most likely to ask.

That last line is the useful part — it surfaces the "what about the vehicle control at 24 h?" question before someone asks it in the room.

Step 4 — Run the overstatement check

What this step does: Catches language creep, the most common problem in AI-drafted science writing.

Ask directly:

List any word in the caption and narrative that implies causation,
significance, or a mechanism that I did not state. Suggest a neutral
replacement for each.

Typical catches: proved, significantly, induced, demonstrates, confirms. Unless you ran the statistics, "significantly" does not belong in the caption.

If something goes wrong

• The caption sounds impressive but says more than your data. That is the failure mode this tutorial exists for — run Step 4 and cut every unsupported word.
• It invents n or a p-value. Never accept numbers you did not supply. Re-prompt with the "missing facts" rule.
• The caption is too long for the slide. Ask for a short slide version and keep the full version for the write-up.
• The wording does not match your group's conventions. Paste two captions from your lab that you like and ask it to match their style.
• The result is unpublished and sensitive. Describe it generically, or keep it in-house — see Safer AI use at work: what to share and what to keep private.

What to try next

• Explaining your research in plain language to non-scientists
• Your first ChatGPT prompt for lab and office work

Key terms

• n — the number of independent replicates behind a data point.
• Replicate — a repeat; independent (separate experiments) is not the same as technical (repeat readings of one sample).
• Error bars — the spread shown around a mean; always state whether they are standard deviation, standard error, or a confidence interval.
• Effect size — how large a difference is, as opposed to whether it is statistically detectable.
• Vehicle control — the treatment condition without the active compound.

Captions and narratives drafted this way are first drafts. The scientist who ran the experiment owns every claim on the slide.