Profile picture ideas: what research actually says about your photo, and how to use the real findings

01Specific poses for anyone with a profile pictures

• Direct gaze, face centred, eyes in the upper third of the frame: 100ms is below conscious-attention threshold; the viewer's brain decides 'trustworthy' or 'not' before the rest of the page even loads. A direct, centred gaze biases that snap-judgment positively.
• A genuine smile (Duchenne) over a posed one: The visible-eye-corner crinkles distinguishing a real smile from a posed one register at the same speed as the trait judgment, and the difference is detectable to viewers consciously and unconsciously.
• Crop the photo using a stranger's recommendation, not your own: Self-selection bias is a real measured effect: people consistently pick worse photos of themselves than uninvolved strangers do. Ask three people to pick from your shortlist.

02Anyone with a profile picture wardrobe guide

Match register to platform: formal for LinkedIn, smart-casual for messengers, expressive-stylised for creator platforms. The best wardrobe choice across the studies is one that does not pull attention away from the face: a solid colour that contrasts with the background, no busy patterns, no logos that compete with facial features.

03What you should expect to pay

A professional studio session typically ranges from to . The AI route provides a comparable result for $15.

01The 100-millisecond first-impression finding (Willis and Todorov, 2006)

The original paper is Willis and Todorov, "First Impressions: Making Up Your Mind After a 100-Ms Exposure to a Face," Psychological Science, 17(7), 592 to 598. The study was conducted at Princeton (not Cornell, which is the misattribution that circulates).

What the experiment actually did:

• Five experiments, each isolating one trait: attractiveness, likeability, trustworthiness, competence, and aggressiveness.
• Participants were shown faces for either 100ms, 500ms, or 1000ms, then for as long as they wanted.
• The trait ratings made after a 100ms exposure correlated highly with the ratings made after unlimited viewing time.
• Of the five traits, trustworthiness showed the strongest correlation between fast and slow judgments. People decide "trustworthy or not" almost instantly, and longer viewing does not change the judgment.

What the paper did not say:

• It did not measure the "halo effect" (a related but distinct phenomenon). The halo claim sometimes attached to this paper comes from later, separate studies.
• It did not study profile pictures specifically. The faces were neutral-expression photos used for psychophysics research, not social-media images.
• It did not measure how durable the judgment is across actual interactions. The "sticky first impression" claim comes from a different research thread (Asch, 1946; modern replications by Rule and others).

The accurate practical reading: a profile picture has roughly 100 milliseconds to do most of its work. The features the brain attends to in that window are the eye region, the mouth, and the overall facial-feature symmetry. A profile picture that hides any of those (sunglasses, far-away crop, profile-angle pose) leaves the viewer with less signal in the window where the judgment is made.

Fig. 01

02The Big Five image study (Liu et al., 2016)

The original paper is Liu, Preotiuc-Pietro, Samani, Moghaddam, and Ungar, "Analyzing Personality through Social Media Profile Picture Choice," Proceedings of the ICWSM 2016 conference. The study was conducted at the University of Pennsylvania, with collaborators across multiple institutions.

What the experiment actually did:

• Analysed the Twitter profile pictures of over 66,000 users.
• Estimated each user's Big Five personality scores from the language of their tweets (using validated text-based prediction models, not self-reports for all users).
• Extracted aesthetic and facial features from the profile pictures: image saturation, composition, sharpness, presence of a face versus an object, presence of multiple people, expression-related features, and several others.
• Tested which image features correlated significantly with which personality traits, controlling for demographic confounds.

The headline correlations the paper reported (paraphrased; for the full statistical detail see the paper):

• Openness: higher-quality, sharper, more aesthetically composed photos. More likely to feature non-face content (objects, scenes), less likely to show a clear face.
• Conscientiousness: brighter photos, faces clearly visible, more conventional composition. Slight tendency toward older or more professional-looking subjects.
• Extraversion: highly saturated, colourful photos. More likely to include other people. Visible-teeth smiling more common.
• Agreeableness: warm, expressive faces. Some correlation with lower technical quality (the friends snapping casual photos pattern).
• Neuroticism: less consistent visual signature; some correlation with the absence of a face in the profile picture.

What the paper did not say:

• It did not say these correlations are strong predictors at the individual level. They are statistical patterns in aggregate; an individual user's photo cannot reliably predict their personality.
• It did not say these findings cause anything. A bright saturated photo does not make someone extraverted; the correlation is between two outcomes of the same underlying trait.
• It did not study LinkedIn, Instagram, TikTok, or WhatsApp. The data was Twitter only.

The accurate practical reading: image-feature choices are visible signals of personality, and the signals are detectable enough that a viewer's snap impression draws on them. Picking image features that match the impression you want to give is rational; expecting any single photo to carry strong individual-level personality information is not.

03The self-selection bias finding (White et al., 2016)

The original paper is White, Burton, Kemp, and Jenkins, "Not looking yourself: The cost of self-selecting photographs for identity verification," British Journal of Psychology, 107(2), 359 to 369. The team is led by researchers at the University of New South Wales (UNSW, Australia) and the University of Aberdeen.

What the experiment actually did:

• Participants picked the photo they thought was most representative of themselves from a personal photo collection.
• Strangers (with brief familiarisation to the participant's appearance) picked photos from the same collection.
• Both sets were tested against a face-matching task: how reliably could a third party verify that the chosen photo matched the same person in a separate verification photo?
• The strangers' chosen photos consistently outperformed the self-chosen photos at the verification task.

The robust finding: people are reliably worse than uninvolved strangers at choosing photos that look like them. Self-selectors over-attend to features only they care about (a particular bad-hair-day they remember, a perceived flaw nobody else notices) and underweight features strangers use to recognise them (overall face shape, characteristic expression, lighting that matches typical viewing conditions).

A subsequent paper, Ritchie, Kramer, and Burton, "Choosing face: The curse of self in profile image selection," extended the finding specifically to social-media profile-picture selection: when the task is "pick the photo that should be your profile picture," self-choosers still consistently underperform strangers.

What the papers did not say:

• They did not say strangers always pick a more flattering photo. They pick a more accurate, more identifiable photo, which is not always the most flattering.
• They did not test AI tools as the "stranger" alternative. Whether an AI image-rating model produces stranger-quality recommendations is unstudied.

The accurate practical reading: shortlist five candidate photos and ask three people who do not know you intimately (a stranger or an acquaintance, not a close friend or family member) to pick. The crowd-of-strangers verdict is reliably better than your own.

04What does not work, sourced

A few patterns that pSEO content keeps recommending despite weak or contrary evidence:

• "Take a power-pose selfie." The original power-pose research (Carney, Cuddy, Yap, 2010) failed to replicate in multiple subsequent attempts; the lead author Cuddy partially retracted the strongest claims in 2016. The visual cue of "confidence in posture" still reads as confidence to viewers, but the hormone-and-behaviour mechanism behind the original study did not hold up.
• "Use Photoshop to enhance your symmetry." The "more symmetric face is more attractive" finding (Rhodes et al., 1998 and others) is real but is about ratings of attractiveness, not perceptions of trust or competence. Symmetry-enhanced photos can read as uncanny rather than attractive once the symmetry is unnaturally high.
• "Wear blue for trustworthiness." This claim is widely repeated and weakly sourced. The colour-trust literature is mixed; the much stronger predictor of perceived trust is facial expression, not clothing colour.

Fig. 02

05The platform-specific spokes

The cross-platform research is one input; the per-platform display constraints are the other. For platform-specific design rules:

• The LinkedIn profile picture spoke covers the 400 by 400 spec, the recruiter-trust shift on AI-generated headshots, and the corporate-formal register.
• The Discord profile picture spoke covers the 16-pixel voice-channel render, the Nitro animated-avatar rules, and the post-2023 dark UI redesign (the #313338 background, not the old #36393f).
• The TikTok profile picture spoke covers the 40-pixel comment-row render, the (mythological) link between PFP and FYP placement, and the consistency rule that genuinely affects creator brand recognition.
• The WhatsApp profile picture spoke covers the phone-number-visibility problem, the March 2024 screenshot-block, and the cross-context dual-use constraint that does not exist on any other platform.

06Where AI generation fits

The honest reading of the research: AI portrait generators are useful for the production half of the problem (lighting, composition, framing, background, wardrobe) but they do not solve the selection half (which photo of yourself to use). The self-selection-bias finding is unaffected by whether the candidate photos were taken with a phone or generated by an AI; the curse-of-self problem is in the chooser, not the source.

The most useful workflow that pairs AI generation with the research:

1. Generate or photograph 10 to 20 candidate photos in the appropriate register for the platform you are targeting.
2. Make a shortlist of 5 candidates.
3. Ask three uninvolved people (not close friends, not family) to pick the most representative one for the specific context.
4. Use their pick, even if you would have picked differently. The research is consistent on this; your own judgment underperforms.

The MyPhotoAI starter plan is $15 for 5 portraits, which produces enough variants to populate the shortlist for the strangers-pick step.

07One-line version

Three peer-reviewed findings shape profile pictures: 100ms-trait judgments (Willis and Todorov, Princeton 2006), Big-Five image-feature correlations (Liu et al., UPenn ICWSM 2016), and self-selection bias (White et al., UNSW/Aberdeen 2016); the practical move is to generate or shoot a shortlist and let strangers pick.

Try a research-aligned profile picture. Generate a shortlist for $15, ask three strangers to pick.

Fig. 03