Am I attractive? How accurately people judge their own faces, how much raters agree, and what the numbers actually look like

1. Why your own judgment is the least reliable instrument

The single most replicated finding in this literature is that self-rated attractiveness barely tracks other-rated attractiveness. Feingold's meta-analysis of 21 studies reported "weighted mean correlations ... of .24, .25, and .24 for the male, female, and all samples" (1992, Psychological Bulletin) — a self-rating explains about 6% of the variance in how others see you.

The error has a direction. In six studies totalling 1,180 people, "unattractive participants considerably overestimated their attractiveness compared to ratings by strangers. In contrast, attractive participants were more accurate. If anything, they underestimated their attractiveness" (Greitemeyer, 2020, Scandinavian Journal of Psychology); self–other correlations in the individual studies ran from 0.13 to 0.32, and mean self-ratings exceeded stranger ratings in every one. People also literally see themselves as better-looking than they are: shown a line-up of their own face morphed more and less attractive, participants picked the enhanced version as the real one, and did so faster (Epley and Whitchurch, 2008, Personality and Social Psychology Bulletin). The inflation is larger in men and tracks narcissism (Gabriel, Critelli and Ee, 1994).

Practical consequence: if you are asking this question because you suspect you are below average, the research says your self-estimate is probably high, not low — and if you suspect you are above average, it is probably about right.

2. How much other people agree — a panel is reliable, one person is not

The figure usually quoted here is misleading. The classic meta-analysis found that adult raters agree on adult faces at r = 0.90 within a culture and 0.94 across cultures (Langlois and colleagues, 2000, Psychological Bulletin) — but that is the effective reliability of an aggregated panel. The same table reports the mean correlation between two individual raters as 0.47. Hönekopp's experiments concluded that "private taste is about as powerful as shared taste" (2006, JEP: Human Perception and Performance), and in a twin study of 547 identical and 214 fraternal pairs rating 200 faces, two people picked at random agreed on only 48% of face preferences — with individual preferences shaped mostly by personal experience, not genes (Germine and colleagues, 2015, Current Biology).

So: your friend's opinion, a date's, a forum thread's, or one app's is roughly half shared standard and half that rater's private taste. Sixty strangers on a fixed scale are a different instrument — a panel of that size rates faces with a reliability above 0.97, whether the scale runs 1–7 or 1–10 (Kramer and colleagues, 2024, Perception).

3. What the numbers actually look like

Two sources, one shape. In the normed Chicago Face Database, faces are rated on a 1–7 scale and the norms for its White subset sit near or below the midpoint — women around 3.4, men around 3.0, with roughly a one-point standard deviation (Ma, Correll and Wittenbrink, 2015; subset figures as reported by Kramer and colleagues, 2024). On SoftMaxx AI's own 0–10 scale, the median, the middle half and the share of scans at 7.0 or higher are published, dated and with their caveats, on the score-distribution page; the middle of that distribution sits near the middle of the scale, not at 7.

Read those together and the folk scale collapses. "A 5" is the middle of measured reality; "a 7" is the top fifth; "a 10" does not occur. If you have been grading yourself against a mental scale where 7 is average, every honest instrument will feel harsh.

4. Ratings move with things that are not your face

The same face is rated more attractive in a group photo than alone — the "cheerleader effect" — a small but consistent boost across five experiments (Walker and Vul, 2014, Psychological Science). Faces become more attractive with repeated exposure (Peskin and Newell, 2004). Composite "average" faces are rated more attractive than almost every individual face that went into them, and more so as more faces are averaged (Langlois and Roggman, 1990). A rating of one photograph is therefore a rating of that photo's lighting, crop, expression, company and familiarity as much as of the person in it — which is why single-shot verdicts, including a single scan, deserve less weight than a consistent pattern across photos.

5. How to get an honest read

1. Do not use yourself. Your estimate carries a known upward bias, largest exactly where you most want accuracy.
2. Do not use one person. Half of any single opinion is private taste. Friends add a courtesy bias; anonymous forums add a hostility bias.
3. Use a fixed scale with a published distribution. A number only means something against the distribution it came from. Ask what the median is before you ask what you scored.
4. Use level, consistent photos. Same lighting, no group shots, no filters; compare trends across several photos, not one verdict.
5. Separate the measurable from the human. Proportion, symmetry and averageness are measurable and explain part of attractiveness; expression, grooming, posture, style and presence are not scored by any photo instrument and move how you read in person.

A measurement tool such as the SoftMaxx AI free scan is useful for exactly one thing: a fixed scale with a published distribution that removes your own bias from the loop. It is a photograph measured, not a verdict on you — the methodology page says precisely what it does and does not measure. If you were about to pay for a web analyzer such as FaceIQ Labs to get that scale, compare the free FaceIQ Labs alternatives first.

6. What none of this can tell you

• Consensus is a panel property, not a person property. The 0.90 agreement figure describes an aggregated group; two individuals agree at 0.47, and random pairs share only about half their preferences. Any single opinion — including an app's — carries roughly half private taste.
• Self-report is biased in a known direction, and the bias is largest where feedback matters most. Self–other correlations of 0.13–0.32; overestimation concentrated at the bottom of the distribution; larger in men and in narcissism.
• A rating of a photo is a rating of that photo. Group context, exposure and averaging all move scores independently of the face.
• The SoftMaxx AI distribution is self-selected. People who scan their own face on a looksmaxxing site skew young, male and appearance-invested; a population-representative median could sit elsewhere. The sample grows with every scan, so the as-of date on the distribution page is part of the data.

7. Sources

• Feingold A (1992). Good-looking people are not what we think. Psychol Bull 111(2):304–41. doi:10.1037/0033-2909.111.2.304
• Greitemeyer T (2020). Unattractive people are unaware of their (un)attractiveness. Scand J Psychol 61(4):471–83. doi:10.1111/sjop.12631
• Epley N, Whitchurch E (2008). Mirror, mirror on the wall: enhancement in self-recognition. Pers Soc Psychol Bull 34(9):1159–70. doi:10.1177/0146167208318601
• Gabriel MT, Critelli JW, Ee JS (1994). Narcissistic illusions in self-evaluations of intelligence and attractiveness. J Pers 62(1):143–55. doi:10.1111/j.1467-6494.1994.tb00798.x
• Langlois JH et al. (2000). Maxims or myths of beauty? A meta-analytic and theoretical review. Psychol Bull 126(3):390–423. doi:10.1037/0033-2909.126.3.390
• Hönekopp J (2006). Once more: is beauty in the eye of the beholder? J Exp Psychol Hum Percept Perform 32(2):199–209. doi:10.1037/0096-1523.32.2.199
• Germine L et al. (2015). Individual aesthetic preferences for faces are shaped mostly by environments, not genes. Curr Biol 25(20):2684–9. doi:10.1016/j.cub.2015.08.048
• Ma DS, Correll J, Wittenbrink B (2015). The Chicago Face Database: a free stimulus set of faces and norming data. Behav Res Methods 47(4):1122–35. doi:10.3758/s13428-014-0532-5
• Kramer RSS et al. (2024). The psychometrics of rating facial attractiveness using different response scales. Perception 53(9):645–60. doi:10.1177/03010066241256221
• Walker D, Vul E (2014). Hierarchical encoding makes individuals in a group seem more attractive. Psychol Sci 25(1):230–5. doi:10.1177/0956797613497969
• Peskin M, Newell FN (2004). Familiarity breeds attraction. Perception 33(2):147–57. doi:10.1068/p5028
• Langlois JH, Roggman LA (1990). Attractive faces are only average. Psychol Sci 1(2):115–21. doi:10.1111/j.1467-9280.1990.tb00079.x

8. Common questions

How do I know if I'm attractive?

Not by asking yourself: self-ratings correlate only 0.24 with how strangers rate you, and people rated least attractive overestimate the most. A fixed scale with a published distribution, or a panel of many strangers, is the only honest instrument; a single friend or forum is half private taste.

What is an average attractiveness score?

On SoftMaxx AI's 0–10 scale the middle of the scan distribution sits near the middle of the scale; the dated figures are on our score-distribution page. Normed research face sets rated 1–7 also sit near or below the midpoint. "A 5" is the middle of measured reality, not below average.

Do people agree on who is attractive?

As a group, strongly: aggregated panels reach reliabilities of 0.90–0.97. As individuals, weakly: two raters agree at about 0.47, and random pairs share roughly half their preferences. Consensus is a panel property.

Why do I look worse in photos than in the mirror?

Partly because you have been seeing an enhanced version: people recognise attractively morphed versions of their own face as the real one. Photos also strip context — group, movement, familiarity — that raises ratings in person.

Is a 7 out of 10 attractive?

Yes. Only about 17% of completed SoftMaxx AI scans score 7.0 or higher; a 7 is the top fifth of real results, not the middle that folk usage implies.