A Midjourney cover illustration, a DALL·E graphic in the methods chapter, a Stable Diffusion diagram in the appendix — at many departments this was still the exception in 2023 and has become the norm in 2026. It comes with a question no style guide answers cleanly: how do you cite an image you generated from a prompt? This guide shows how APA 7, Chicago and Harvard handle it, what belongs in the caption — and when you should not use an AI image at all.
Do AI images need a citation?
Yes — but for different reasons than a book citation. A book citation attributes a claim. An AI image citation covers two other things: traceability (which tool, which prompt, which version) and authorship (you were not the sole creator, the model participated). Both need to be visible in your thesis.
In practice: every AI image needs a caption that makes clear it is AI-generated, with which tool, in which version, from which prompt. What lands in the running text, what in the caption, and what in the appendix depends on the citation style and your department’s rules.
Who is the author of an AI image?
Legally this is contested in Europe; academically it is clearer: you are not the sole author. Neither is the model — Midjourney and OpenAI do not claim authorship in the scholarly sense. Academic practice has settled on declaring the image as AI-generated under human supervision. That matches what major publishers have required since 2023: Nature, Elsevier and Springer do not accept AI-generated images in manuscripts (with narrow exceptions for documented methods). Student theses have softer rules, but the disclosure requirement remains.
Concretely: the caption should not say “own illustration” — that misrepresents authorship. Use “Generated with Midjourney v6, prompt see Appendix C” or one of the forms below.
AI images in APA 7
APA published official guidance on generative AI in July 2023 (Style Blog: “How to cite ChatGPT”). It treats text and image the same: the model is the “author”, the year is the year of generation, the prompt belongs in the running text or an appendix.
In-text citation
The figure was generated with Midjourney (Midjourney, 2026).
Caption
Figure 3. Visualisation of a sustainable city district. Generated with Midjourney v6.1 on 12 May 2026. Prompt: “aerial view of a sustainable city district, solar panels, green roofs, cinematic lighting, photorealistic”. Prompt log see Appendix C.
Reference list
Midjourney. (2026). Midjourney (Version 6.1) [Text-to-image AI]. https://www.midjourney.com/
If you generate several images with the same tool in one chapter, cite the tool once and list the individual prompts compactly in an appendix.
AI images in Chicago (Notes & Bibliography)
Chicago treats AI images like personal communication or unpublished material: they appear in a footnote and in the figure credits, not necessarily in the bibliography.
Footnote (first mention)
¹ Image generated with DALL·E 3 (OpenAI), prompt: “watercolor illustration of a Swiss mountain village at dusk”, 12 May 2026.
Figure credits at the end of the chapter
Fig. 3: DALL·E 3 (OpenAI), 2026. Prompt details in appendix.
If your department uses Chicago Author-Date, the APA-style form above works the same.
AI images in Harvard
Harvard has no central authority — every university defines it slightly differently. The prevailing UK/Commonwealth consensus (following Cite Them Right / Anglia Ruskin 2024):
In-text
(OpenAI 2026)
Reference list
OpenAI (2026) DALL·E 3 [AI image generator]. Available at: https://openai.com/dall-e-3 (Accessed: 12 May 2026).
Prompt and date belong in the caption, as with APA.
What always belongs in the caption
Regardless of style, six pieces of information should always appear:
- Tool and version. “Midjourney v6.1”, not just “Midjourney”. Models change substantially between versions — a v5 image looks nothing like a v6 image.
- Generation date. Month and year at minimum, ideally the date.
- Prompt (in full). Short prompts in the caption itself, long ones in the appendix with a cross-reference.
- Post-processing. If you retouched, cropped or recoloured the image in Photoshop, note it briefly.
- Seed or job ID (Midjourney/Stable Diffusion). Nice-to-have for reproducibility; some regulations require it.
- Rights chain. For models with unclear training data provenance (Stable Diffusion base models, Midjourney), a short note — more below.
Prompt log in the appendix
For a thesis with more than two or three AI images, a dedicated appendix is worth it. Layout:
| No. | Fig. no. | Tool + version | Date | Prompt | Seed / Job ID | Post-processing |
|---|---|---|---|---|---|---|
| 1 | Fig. 3 | Midjourney v6.1 | 2026-05-12 | “aerial view of …” | 1234567890 | Cropped in Photoshop |
| 2 | Fig. 7 | DALL·E 3 | 2026-05-14 | “watercolor illustration of …” | — | None |
That makes your work checkable for supervisors and in the oral defence — they could enter the prompt into the same tool and see whether something similar comes out. Some departments now do exactly this on a spot-check basis.
Legal pitfalls — briefly
- Training data dispute. Midjourney and Stable Diffusion were trained on copyrighted images without a licence. Ongoing cases in the US and EU are unresolved. For a bachelor thesis in an academic context this is usually unproblematic (fair dealing / research exception), for a later publication it can become one.
- Tool terms of use. Midjourney allows commercial use from the base plan onward, DALL·E likewise, free tools often not. For theses, “personal / academic use” is normally covered.
- Personality rights. Photorealistic AI images of identifiable real people are risky — personality rights apply to generated depictions too. Rule of thumb: no recognisable real people without consent.
None of this is legal advice, but it will keep you out of the most common traps.
When you should not use an AI image
Not every occasion justifies a generated image. Skip it when:
- The image illustrates your results. An AI image that “looks like” your survey results is a data integrity problem. Use charts (Excel, R, Python) based on the real data.
- It concerns methodological steps. Process graphics, flowcharts, study designs belong in Draw.io, Lucidchart or LaTeX — not Midjourney. Reproducibility beats aesthetics.
- The scholarly value is unclear. A cover image “because it looks nice” weakens the impression your thesis makes. Supervisors read that as a focus signal — rarely a good one.
- Your department prohibits AI images. More and more regulations have done so explicitly since 2024/25. Before you start: check the FAQ, the thesis handout, and — if unsure — email your supervisor.
Better alternatives for the common cases
- Diagrams / processes → Draw.io, Lucidchart, Mermaid (in LaTeX/Markdown), TikZ.
- Data visualisation → matplotlib, ggplot2, Datawrapper, Excel (not glamorous, but honest).
- Illustration without licence worries → Unsplash (CC0), Wikimedia Commons (CC-BY), NounProject (icons), scanned own sketches.
- Photorealistic reference → take a photo. Sounds trivial, is often more honest than an AI rendering.
Pre-submission checklist
- Every AI image is declared as AI-generated in the caption.
- Tool + version, date and prompt appear either in the caption or in the appendix.
- The appendix prompt log is complete and linked to the figure numbers.
- Rights questions (personality rights, terms of use) have been considered.
- The declaration of original work has been extended to cover AI image use (if the template requires it).
- No AI image sits in a place where real graphics / real photos / real data belong.
AI images are a disclosure question, not a citation question
The key difference from text citations: for a book citation, a citation check verifies whether the claim actually appears in the source. For an AI image, the point is not what the image “says” — the point is that its origin is disclosed honestly. An AI image is honest when the caption makes you the prompter and the tool the generator visible.
The mirror version applies to text: a nicely formatted citation is not honest yet if the source does not support the claim, or does not exist. That is exactly what Acurio covers — for the text part of your thesis it checks every citation against the source. For images, the disclosure remains your job; the checklist above is your tool.
AI images are a legitimate device in theses when they are made transparent. Take the ten minutes for a prompt log and three lines of caption and neither correction nor defence will have anything to catch you on. Hide them, and both will.