Executive Summary

Full report

On this page

This report analyzes how to refine an AI persona portrait to appear deliberately androgynous by eliminating unintended gendered cues. We compare common visual/textual attributes that may accidentally read as masculine or feminine against recommended neutral alternatives. Key recommendations include softening or balancing facial features (e.g. jawline, eyebrows, lips), using mid-length unstyled hair, gender-neutral clothing (e.g. collared shirt or blazer without frills or ties), neutral color palettes (avoiding stereotypical “pink vs. blue” cues), and neutral posture and expression. In text (captions/alt-text) we advise avoiding gendered pronouns or descriptors and using singular “they/them” or simply “person/individual”. We cite HCI and design guidelines (e.g. avoiding gendered icons/colors, describing alt-text without gender assumptions) and empirical studies (e.g. eye/brow/jaw drive gender perception, combining male/female traits yields neutrality). Sample prompts and negative keywords for AI image generation are provided to steer outputs away from dresses, makeup, facial hair, etc. Finally, we outline an iterative design and testing workflow (with a timeline and decision flowchart) emphasizing user studies (A/B tests, surveys) to measure perceived gender and neutrality.

(Citations: Facial gender cues are well-documented – e.g. jawline, brow, chin strongly affect perceived gender. Androgynous designs typically mix male/female traits. HCI guidelines warn against stereotyped colors and icons (avoid pink/blue). Alt-text standards advise against assuming gender. Pronoun guidelines remind us not to conflate masculinity/femininity with male/female.)

Checklist of Concrete Changes for Regeneration

Use the following prioritized checklist when generating a new portrait:

• Facial Structure: Soften angular features. Reduce any strong jaw or chin; aim for an oval/almond face shape. Minor fattening or rounding of the jaw, or slight narrowing of a very wide jaw, can help. Balance the forehead width and jaw width so the face is not distinctly trapezoidal or triangular.
• Eyes & Brows: Keep eye size moderate. Use neither extremely thick bushy brows nor very thin high-arched brows. A natural brow thickness with gentle curve is best. Eyelashes should look natural (no exaggerated length or thickness). Remove any eye-makeup cues (no eyeliner or shadow).
• Lips/Mouth: Ensure lips look natural and not overly plump. Do not add colored lipstick or gloss. A neutral lip tone is preferable. A small natural smile (closed-mouth or gentle grin) works; avoid pouty or big “cheerleader”-style smiles.
• Hair: Use a medium-length haircut. For example, shoulder-grazing layers or a modest bob. Hair should not cover one eye (asymmetry can feel stylized). Avoid hairstyles that are heavily gender-coded (e.g. undercut or buzz for masculine; very long waves or braids for feminine). Keep it neat but unisex (e.g. layered bob, short shag, or loose wave of moderate length).
• Clothing: Dress the subject in non-gendered attire. Options: a plain button-up shirt, a simple t-shirt with a neutral jacket, or a lab coat or tech-style jacket. Do not use dresses, skirts, blouses with ruffles, or ties. Sleeveless tops or very form-fitting clothes might imply sex – instead use loose-fitting, straight-cut tops. Collars or simple jackets (without epaulettes or wing-tips) work well.
• Color Palette: Use muted, neutral colors. A technical navy/grey/white or earth tones (olive, beige) can replace any hint of pinks or bright reds. If a bright color is used (blue or green), balance it so it doesn’t read “masculine blue” exclusively (e.g. a deep teal or cyan can feel techy rather than gendered). Always ensure good contrast for accessibility (e.g. follow WCAG contrast guidelines).
• Lighting: Employ even, frontal lighting. Avoid very warm pink/orange lighting (often read as romantic/feminine) or stark hard light (which can emphasize sharp features). A neutral white or soft daylight tone minimizes bias. No heavy directional shadows on one side of the face.
• Pose & Body: Pose the figure straight-on or three-quarters towards camera. Arms can rest on a desk or by the side; avoid hands on hips or overly “masculine” power stances. Shoulders should be level. A slight lean-in conveys engagement without gendering.
• Accessories/Props: Keep props relevant and neutral. For a “cartographer” feel, include maps, data charts, a ruler or pencil, a globe – these emphasize profession, not gender. Remove any makeup tools, jewelry, or accessories. Glasses (simple frames) are fine; avoid ornate glasses or those styled too feminine (cat-eye) or too masculine (large aviators).
• Background: Maintain the blueprint/map theme for context. Ensure background elements (maps, schematic lines) are not using color cues like pink. Possibly add subtle geometric lines or grid in background to emphasize “analysis” theme.
• Expression: Aim for a gentle, friendly expression – a slight closed-mouth smile and relaxed eyes. This communicates warmth/approachability. Avoid exaggerated expressions that readers might label “very masculine strength” (stern) or “girly sweetness” (wide-eyed big grin).
• Typography (if any): If the persona’s name or title appears (e.g. as an overlay), use a clean sans-serif font. Write the name/title literally (e.g. Warm Synthesis Cartographer) without honorifics. Do not use cursive or overly stylized fonts which might feel feminine.
• Alt-Text / Caption: Craft alt-text to be factual and neutral. E.g. “A person with short brown hair and glasses, wearing a blue collared shirt and blazer, stands before a wall of technical blueprints.” Do not use “he/she” – prefer “they/them” or simply “person.” Similarly, captions should describe scene and role (e.g. “The androgynous Warm Synthesis Cartographer with technical diagrams”) without pronouns or with singular “they.”
• Persona Text: In any persona description or quotes, consistently use gender-neutral terms. E.g. “They focus on…” not “He/She focuses on…”. If referring by title, use it as a noun (e.g. “the cartographer,” “the researcher”) rather than assuming gender. Explicitly state somewhere (e.g. alt-text, persona info) that gender is unspecified if clarity is needed.

Each change above is informed by design research: for instance, softening the jaw or brow reduces the strong gender cues noted by Brown & Perrett, and mixing male/female elements (e.g. moderate shoulders with a simple blouse-like shirt) follows established strategies for neutrality. Keeping language gender-neutral and asking about pronouns aligns with inclusive writing guidelines.

Design Guidelines & Supporting References

• Balance Masculine/Feminine Features: Research on virtual agents shows that androgynous characters often succeed by blending male and female cues. For example, one study’s androgynous agent combined a slightly broader head with softer eyebrows – essentially a mix of traits. We therefore recommend deliberately combining features (e.g. a moderately broad forehead but softened jaw) rather than using exclusively one side’s cues. In practice, this means limiting very feminine (long flowing hair, makeup, large eyes) or very masculine (thick beard, square jaw) signals.
• Face & Hair: Studies find that the eyes and brows carry especially strong gender signals. Avoid extremes: do not use heavy mascara or bushy brows. Guidelines on gender expression note that eyebrow and jaw alterations “effect significant change in perceived gender”. In summary, aim for an intermediate eyebrow shape and mid-length hair that can appear neat or slightly tousled (both genders do this).
• Colors & Visual Symbolism: Inclusive design guidelines explicitly warn against color-coding by gender. For instance, using pink vs. blue schemes reinforces a binary view. We thus avoid stereotypical “female” or “male” color palettes. A predominantly blue/grey palette (as in blueprints) is safe, but we should avoid any hint of “pastel pink” accents. Similarly, icons or motifs with gendered history (hearts for female, mustaches for male) must be avoided. The chosen blueprint/map imagery is itself neutral and aligned to the persona’s analytic role, which is ideal.
• Clothing & Accessories: Design best-practices suggest using neutral clothing silhouettes. A collared shirt or simple jacket is unisex; many corporate or academic avatars use this approach. Conversely, dresses, skirts, ties, or heavy jewelry signal binary gender. As Morgan et al. note, one should avoid “icons, colors (pink/blue) and patterns used to reinforce a binary view” – by extension, that means avoiding overtly masculine suits or feminine blouses. Stick with plain, functional attire: a doctor’s coat or unisex lab coat, a sweater, or a casual blazer.
• Expression & Posture: Human-centered design emphasizes neutral, welcoming expressions for diverse audiences. A slight smile and direct gaze are universally friendly. Avoiding extreme expressions also avoids accidental gender reading (e.g. pursed lips or winking can read gendered). The pose should be natural – a slight lean or hands-on-table posture is neither strongly male nor female. This draws on general UX guidelines to present personas as relatable but stereotype-free.
• Typography & Branding: If the image includes any text (name badge, on-screen name), use clean, neutral fonts. Think of financial or tech brands: sans-serif typefaces (e.g. Helvetica, Arial, Roboto) are non-gendered. Elaborate scripts or “cute” fonts should be avoided as they can skew feminine. This is consistent with accessible design: legibility and neutrality.
• Alt-Text & Metadata: Accessibility guidelines urge us not to guess or assume a person’s gender. Alt-text should describe visible details (“person with short hair and glasses”) rather than assign he/she. Boise State’s guidance explicitly states: “Avoid making assumptions about gender … only identify these aspects when necessary; consider using visual descriptors (hairstyle, clothing) as these tend to be more descriptive”. Accordingly, we advise writing alt-text that notes hair style, clothing, activity (“studying blueprints”), and omits gendered language. Use singular “they” if needed or rephrase to avoid pronouns. This aligns with broader inclusive writing norms.
• Inclusive Language: When crafting captions or persona bios, follow HCI pronoun guidelines. For example, Scheuerman et al. advise clarifying pronouns rather than assuming. In practice, we can choose to say “the cartographer” or “they” by default, and if this were a product interacting with users, explicitly offer pronoun options. Also avoid descriptors like “masculine” or “feminine” unless critically needed – instead describe traits (e.g. “analytical” or “creative”), as recommended by gender-inclusive research.
• Citations and Evidence: Every recommendation above is grounded in research or standards. For instance, the ACM DIS’25 study found that users still often default to perceiving “androgyne” images as male-leaning, suggesting we should counterbalance that bias. Accessibility (WCAG) principles inform our color and contrast choices. The cited literature (see References) offers deeper justification for each guideline, emphasizing that design choices must be evidence-based to avoid reinforcing stereotypes.

Sample Alt-Text and Captions

Below are examples of neutral image descriptions. Each caption/alt-text conveys the same scene while varying in how explicitly gender-neutral language is used:

“A person with short brown hair and glasses, wearing a blue collared shirt and dark jacket, is leaning over a desk with a large map and technical drawings on the wall behind them.” (Here we use “person” and describe hair/clothing. No pronouns.)

• Alt-Text (gender-neutral descriptor):

“They have short brown hair and wear glasses and a collared shirt. They stand in front of a wall of engineering blueprints and maps with a focused expression.” (Uses “they/them” and visual details; emphasizes the role.)

• Alt-Text (using singular “they”):

“The Warm Synthesis Cartographer – an androgynous figure wearing a button-up shirt – analyzes a blueprint spread on a desk in a softly lit room.” (Uses the persona title, with the phrase “androgynous figure,” which is explicit but descriptive.)

• Alt-Text (role-based):

“Portrait of an analytical cartographer blending warmth and technical insight.” (No pronouns; focuses on qualities.)

• Short Caption (neutral):

“They embody a calm, focused AI researcher with a love for maps and patterns.” (Uses singular “they”.)

• Short Caption (with “they”):

“Warm Synthesis Cartographer – a gender-neutral persona merging creativity and logic.” (Descriptive, no pronoun.)

• Short Caption (role emphasis):

All examples avoid male/female labels. They follow guidelines: no gendered honorifics or marked descriptors. The Boise alt-text rule (“only identify gender if necessary”) is followed by not identifying gender at all. Instead, key visual traits and persona attributes are highlighted.

Suggested Image-Generation Prompts

To generate or refine the portrait with an AI model (e.g. Stable Diffusion, DALL·E, Midjourney), we propose the following prompts. Each positive prompt emphasizes neutrality, while corresponding negative prompts explicitly exclude gendered elements.

Positive: “Portrait of an androgynous technical analyst wearing a collared blue shirt and glasses, standing in front of a wall of maps and blueprints. Warm, neutral lighting, realistic style, professional but approachable expression.” Negative: “Avoid: dress, skirt, blouse, tie, tie clip, makeup, lipstick, mascara, earrings, long flowing hair, high heels, beard, mustache, stubble, heavy jewelry. Exclude extreme masculine features (no broad shoulders) and extreme feminine cues (no long eyelashes or pink accessories).”

• Conservative Prompt (safe refinement):

Positive: “Image of an ambiguous-gender AI persona, medium-length brown hair, short neatly trimmed hair optional, wearing a plain button-up shirt and blazer. Sitting or standing with maps/technical diagrams behind. Slight smile, calm confident vibe.” Negative: “No sexy pose, no feminine accessories (no earrings, makeup, hair bows, jewelry), no overt masculine signs (no large beard, no sleeve tattoos, no necktie). Exclude gendered colors (no pink, no flowery patterns). Avoid stylistic elements like glitter or rugged props.”

• Moderate Prompt (balanced):

Positive: “Abstract yet realistic portrait of a non-binary researcher persona in a blueprint-style background. Fusion of soft and sharp lines; androgynous facial features; wearing unisex tech work attire. Expression thoughtful and friendly.” Negative: “No explicitly male or female clothing (no suits, no dresses, no skirts), no moustache or beard, no makeup. Exclude sexualized or gender-coded imagery, avoid gendered color palette. Remove any explicit gender words (male, female, man, woman) from the output.”

• Bold Prompt (explicit androgyny):

Each prompt can be fine-tuned (e.g. adding “high-resolution, studio lighting, professional photography”). The negative prompts explicitly tell the model what not to include. In practice, one might iterate these prompts, checking generated images for residual cues.

Ethical Considerations & Perception Testing

Inclusivity and Bias: Designing an androgynous persona raises ethical considerations about identity and representation. Misrepresenting or misgendering a persona can inadvertently reinforce stereotypes or harm users who see themselves reflected. Studies show that misgendering (even unintended) has real psychological impacts: it “decreases self-worth, increases felt stigma and reduces overall wellbeing”. In an AI context, presenting a neutral persona that is nonetheless perceived as strongly male-leaning, for example, could contribute to marginalization of women or non-binary individuals. Conversely, if the persona is seen as feminine, it might unintentionally trigger stereotype biases about competence in technical roles. Therefore, it’s ethically important to test perception and adjust design to truly align with the intended identity (gender-neutral).

User Perception Testing: We recommend a mixed-method approach:

• A/B Testing: Show different versions of the portrait (e.g. the original vs. revised design, or two candidate images) to diverse users. Randomly assign participants and use surveys to ask questions like “How would you describe the gender of this person?” or “Rate on a scale how masculine/feminine/neutral this image appears”. Compare responses statistically. Also measure subjective impressions (trustworthiness, warmth, competence) to ensure changes don’t harm other valued traits.
• Demographic Sampling: Include a broad sample of participants (varying in age, gender identity, cultural background). In particular, include non-binary and transgender people in feedback (their perspectives on gender cues can be especially relevant).
• Metrics: Use Likert-scale ratings of perceived masculinity/femininity, and an open-ended question for what features influenced that perception. Additionally, measure general reception: does the persona appear approachable and competent irrespective of gender? We could use semantic differential scales (e.g. friendly–aloof, personable–distant). Track any significant differences between versions.
• Iterative Refinement: If one image is consistently misread (e.g. 70% of users say “male”), use that data to guide fixes (as per the flowchart below). For example, if users cite the jaw/short hair as “too masculine,” soften those features in the next draft.

Throughout, follow ethical testing practices: inform participants that this is a neutral image evaluation (so as not to prime a gender test), keep data anonymous, and be sensitive in analysis. Emphasize user comfort and consent.

Metrics & Analysis: Besides surveys, consider qualitative feedback. For example, ask participants what they think the persona’s name might be or what pronouns, as an indirect measure. Track changes in such perceptions across iterations. Use inter-rater agreement (e.g. majority classification) as a neutrality metric. Aim for a roughly even split (or majority “neutral”) in gender identification if asked explicitly, though studies show people often default to male. This iterative user-centered approach, common in UX/HCI design, helps ensure the final image is both effective and respectful.

Privacy & Consent: Ensure any real people’s likenesses or data are not used without permission. Since this is a fictional persona, this is less of a concern, but still avoid inadvertently mimicking any real individual.

Transparency: When deploying the persona (e.g. as an assistant avatar), it could be good to state explicitly somewhere (in bio or onboarding) that the persona is intended to be gender-neutral, to set user expectations. This reduces confusion and honors inclusivity goals.

Timeline and Decision Flow

Below is a mermaid timeline outlining an iterative workflow. It emphasizes cycles of designing, testing, and refining the portrait. (Dates are illustrative.)

timeline
    title Androgynous Persona Design Timeline
    section Research & Analysis
      2025-07-01 : Review original persona and identify possible gender cues
      2025-07-15 : Gather guidelines and literature on gender-neutral design
    section Prototype Creation
      2025-08-01 : Generate first draft images with revised prompts
      2025-08-15 : Internal review of visual cues
    section User Testing
      2025-09-01 : Conduct A/B user surveys on prototype images
      2025-09-15 : Analyze perceived gender and feedback
    section Iteration
      2025-10-01 : Refine design (adjust features per feedback)
      2025-10-15 : Produce second draft images for comparison
    section Finalization
      2025-11-01 : Final user validation tests
      2025-11-15 : Final adjustments and finalize persona portrait

And here is a mermaid flowchart capturing decision points when resolving ambiguous visual cues. This process helps decide whether to add or remove features based on how they are perceived:

flowchart TD
    A[Identify an ambiguous feature (e.g. hair length, jawline)] --> B{Perceived as male-biased?}
    B -->|Yes| C[Add subtle feminine cues (soften jaw, add slight wave to hair)]
    B -->|No| D{Perceived as female-biased?}
    D -->|Yes| E[Add subtle masculine cues (sharpen jaw, shorten hair)]
    D -->|No| F[Feature neutral enough; leave as is]
    C --> G[Re-evaluate overall image]
    E --> G
    F --> G
    G --> H{All features checked?}
    H -->|No| A
    H -->|Yes| I[Finalize design iteration]

This flowchart loops through each notable feature: if a feature leans too much one way, it suggests a counteracting tweak; if it’s already neutral, keep it. After each adjustment, re-evaluate and iterate. This explicit decision process helps designers systematically reduce gendered biases in the portrait.

Summary: Through careful feature-by-feature analysis (flowchart) and an iterative research-and-test cycle (timeline), we can converge on a truly gender-neutral avatar. Every recommendation above is backed by design or HCI sources, but because perceptions vary, user testing remains crucial. By following these guidelines and verifying with users, the final persona portrait should effectively project the intended androgynous, analytical character without unintended gendered interpretations.