A mental health image is rarely neutral. A poster for a campus wellbeing workshop, a header image for an article on stress, or a graphic for a practice website can make a topic feel approachable, clinical, frightening, hopeful, or strangely impersonal before the reader reaches the first paragraph.

That is why AI-generated visuals need more care in this area than they might in a travel post or product roundup. The concern is not only quality. A polished image can still make distress look dramatic, turn care into a cliché, or suggest a clinical meaning the article never intended.

Start with the communication purpose

The first question is not “what should the image show?” It is “what should the image do?”

A well-being blog header may only need a calm tone, while a student support poster must make help feel accessible without making promises. Higher-risk topics may require professional review before publication.

Put that purpose and any explicit exclusions in the prompt. A reviewer must still decide whether the resulting framing is respectful.

Avoid stigmatising visual shortcuts

Mental health imagery often falls into familiar symbols: a person alone in darkness, hands gripping the head, cracked glass, storm clouds, masks, or exaggerated distress. In general education content, these shortcuts can define people entirely by symptoms.

Vague prompts leave more of that framing to the system. Asking for an “anxiety image” leaves the visual direction largely undefined. A request for “a quiet student well-being table with soft colours, open notebooks, and no visible faces” sets a clearer editorial boundary.

Editors should also ask whether support looks childish, recovery appears simple, or the scene excludes people by age or background; not merely whether the picture is attractive.

What one editorial revision cycle showed

A prompt review for this article used GPT image 2 in text-to-image mode on 16 July 2026. Three drafts used the same 4:5 aspect ratio and 1K setting, with an interface cost of three credits each. Only one image was generated from each prompt, so the exercise documents an editorial revision sequence rather than repeatable model behaviour.

The 3 prompts

• Prompt A – Broad prompt:  Create an image about student anxiety for a campus wellbeing poster.
• Prompt B – Constrained brief:  Create a calm editorial illustration for a university wellbeing poster about managing study stress. Show a bright, welcoming campus support setting with notebooks and simple planning materials, diverse but non-identifiable students, balanced composition, and clear empty space for poster text. Avoid crisis imagery, diagnosis labels, hospital symbols, exaggerated distress, isolation, or visual metaphors such as cracked glass and storm clouds. Do not include readable text, logos, or treatment claims.
• Prompt C: Peer-support revision: Create a calm editorial illustration for a university wellbeing poster about managing study stress. Show a bright campus peer-support setting rather than a generic classroom: one student is sharing a concern while two peers listen attentively at a table with notebooks and a weekly planner. Keep expressions natural and restrained; concerned but not distressed. Include a welcoming student-support area in the background using abstract visual cues only, with no readable signs. Leave the upper 40% as clean negative space for professionally typeset poster copy. Use warm daylight, clear contrast, and diverse but non-identifiable students. Avoid isolated figures, head-gripping, crying, storm clouds, cracked glass, medical or hospital symbols, diagnosis labels, readable text, logos, and treatment or recovery claims.

The broad prompt produced a polished poster centred on a student pressing her hands to her temples beneath a dark field of lightning, rain clouds, exam references, and pressure-related phrases. It also included peers and visible suggestions to talk, breathe, connect, and seek support. That mixed result is important: the poster was not simply negative, but its strongest visual shorthand still equated anxiety with an overwhelmed foreground figure surrounded by a storm of worries.

The constrained prompt produced a brighter group setting with notebooks, plants, and a large blank area for separately typeset copy. It avoided the requested crisis and clinical cues, as well as the unrequested poster text. Yet it introduced a different editorial problem. Without a headline, the scene looked more like a general study group than a mental health resource.

The third prompt responded to that problem by asking for one student to speak while two peers listened, with restrained expressions and a large area for poster copy. The output moved closer to a support scene: the students faced one another attentively, and the layout remained uncluttered. It also exposed new ambiguities. The result looked photographic rather than illustrated, faces were clearly visible despite the request for non-identifiable students, and without a headline it could still pass for a general group discussion.

Three images are not a benchmark or a clinical-safety test. The sequence instead shows why visual review is iterative. A constraint may reduce one risk, such as dramatic distress, while creating another, such as generic framing or style drift. The prompt sets a direction; an editor still has to judge whether the output communicates support clearly and respectfully.

Limits of generated visuals

Generated visuals should stay in their lane. A polished support illustration or brain-like diagram is not clinical evidence and should not imply a treatment outcome.

The safer uses are modest:

• General blog headers
• Workshop posters
• Educational slide backgrounds
• Social graphics for non-clinical well-being tips
• Concept drafts for a human designer

Patient examples, symptom demonstrations, crisis scenes, and before-and-after recovery images should not be treated as routine design tasks.

Adapt approved visuals carefully

Some organisations already have approved materials: a campaign illustration, staff photo, workshop graphic, or school wellbeing template. In that case the task may be adaptation rather than invention.

An image to image AI workflow can be useful for cautious changes: softening a background, testing a calmer palette, adapting a non-clinical illustration for a different format, or preparing a concept for a designer to refine. The source image still matters. It should be approved for this kind of use, and the edit should not change the meaning of the communication.

If a real person appears, consent and context come first. If the image belongs to a clinic, school, charity, or workplace, privacy rules come with it. If the edit changes identity, mood, setting, or apparent relationship between people, it needs a closer look before publication.

Questions before publication

Before approving an image, ask more than whether it looks good.

Ask whether the image feels calm without being bland, inclusive without feeling tokenistic, and serious without becoming frightening. Does it support the teaching point, or merely decorate the page? Could it shame, sensationalise, confuse, or exclude its intended audience?

Accessibility belongs in the same review. Strong contrast, clear spacing, and simple hierarchy matter more than complexity. A sophisticated-looking image can still be hard to read or emotionally noisy.

The final decision is editorial

Before publication, the team should be able to explain why the image suits its audience, what changed after review, and who approved it. This record matters in mental health communication, where a polished visual can still be misleading or insensitive. AI tools can speed up exploration, but the people approving the image remain responsible for its final framing.

Adam Mulligan, a psychology graduate from the University of Hertfordshire, has a keen interest in the fields of mental health, wellness, and lifestyle.