How the palm scan works: what's measured and what's tradition

What's measured

The free scan runs in your browser, on your hand and nothing else. Here is every measurement it makes, how it makes it, and what comes out. No trained line model is involved: line tracing is plain code that anyone could read and check. The scanner page walks through the same steps with pictures.

The 3D hand and the star map add no new measurements. The hand's proportions come from your 21 points, and its thickness, skin and the slight curl of the fingers are a stylised model. The bright stars are real points on your hand, and the faint ones are sampled from your crease map.

What isn't measured

• The sun, marriage, children and money lines. The search for them is off. With it switched on, it made 6 traces in 288 chances across 72 photos, and all 6 were wrong by eye: 3 ran along the palm's edge, 1 along pen ink and 2 on fold shadows. You can place these lines yourself on the line tracer, and placing them does work: the line snaps to the crease between your points and gets measured.
• Stars, islands, mystic crosses and other tiny marks. Not attempted. Palm skin has crossing wrinkles everywhere, and a filter that sees creases a few pixels wide can't defend a claim about a small cross.
• How raised a mount is. A photo can't tell height from shadow.
• You. Your age, sex, health and future aren't measured or estimated. No free tool sends your hand to an AI to be interpreted.

How well it works, as far as we've tested

We graded the engine once on 72 held-back photos, 38 clear and 34 poor. The numbers below are from that run, and we have not re-tuned since. "Correct" means the trace sits on the right crease and covers at least 80% of its visible length. "Partly" means anyone looking would still recognise the right line. "Wrong" means it is mostly on another crease or none. "Missed" means the scan said not found, and the crease could be seen.

In plain words: on a clear photo of an open hand, the three main creases are usually found, and getting all three fully right on one photo is less common than getting any one of them right. There is a spread here, and the lower figures are the ones to quote: 55 to 57% for all three correct on the second grading's grades, against 68% on the first grading's. Line by line the two are close, 80 to 84%, but one trace that stops short flips a whole photo, so the whole-photo figure is fragile. The real-world row is the closest to what you'll upload, at 53 to 62% for all three correct, on only 13 to 15 photos.

The fate line doesn't work, and these two grades show why. Our first grading called 1 of 22 traces on clear photos fully correct, and the second grading called 65% or 62% correct, under the same written rule. It counted any lengthwise crease up the middle of the palm, and many hands have several, so its figure shows the trace is on a real crease and not that it's the one a palm reader would point to. Two graders that far apart means the rule can't settle it, so we treat fate as unmeasured. That is why the scan asks you to check or place it whatever its label says.

We set out aiming for "major lines right on 80% or more of clear photos". Whether we reached that depends on how you read it. Line by line, heart, head and life reach 84% in the first grading and 83% or 80% on the second grading's, right at the edge, on a set that is about 60% one lab rig. Whole photo, all three at least partly right, reach 84%, or 91% or 87%. All three fully correct is 68%, or 57% or 55%, which misses. Real-world photos miss on all three correct, and so does anything that counts the fate line as major.

Why you should doubt those numbers

• The samples are small: 72 photos, or 139 with the tuning set. Treat a few points of difference as noise.
• None of these came through the live scan on a real phone in a real person's hand. They are photos other people took. We have not measured real phones.
• We chose photos where the hand model found an open hand, from 626 licensed candidates, most of which weren't palms at all. A photo with no detected hand isn't in the numbers.
• Grades are judgments, made by eye on small sheets.

Does the label mean anything?

Mostly, yes. A high label was rarely wrong, though it means a strong line was followed and not that it was the right crease: on real-world clear photos the second grading found 2 of 19 high traces wrong, both a mix-up of the heart and head lines. A low label is a real warning overall, and on clear photos it is too cautious, which we prefer: better to ask you to check a right line than to assert a wrong one.

When the scan said not found, on 68 calls, our first grading counted 56 as fair (the crease really couldn't be seen, or something covered it) and 12 as misses, 7 of them fate. The second counted 54 as fair and 14 as misses.

Skin tone

The hand model's own card says it found no error pattern by tone. Its table of landmark error runs from about 4.5% to 6.2% of palm size for the full model across six skin types, highest at types 6 and 1. That covers the points and says nothing about palm lines, which are our code. For the lines, here are the three main ones on clear held-out photos, in the first grading, because the second grade wasn't split by tone. Tone and source are tangled, so they are shown split.

Held-out has no clear real-world photo of a dark palm. The tuning set has 5, 15 lines in all: 67% correct, 87% at least partly right and 13% wrong. Dark palms are under-tested. Nothing here shows a big gap, and nothing here rules one out. Palms are lighter than the back of the hand, and these classes describe only the palm.

The rest of what the scan reports

• Forks and breaks are weak. On 72 photos, 17 forks were reported: 5 looked real, 10 were wrong and 2 we couldn't tell. Of 13 breaks, 3 looked real and 10 were the trace straying off its crease. So your reading never mentions a break, and hints at a fork only on a clear line, in words that allow for being wrong.
• The single-crease candidate fired on 6 of 139 photos, 3 of them hands whose source calls them single-crease. Of 6 known single-crease hands, it found 3. Most of those were in the tuning set, so these figures flatter it. It is a shape, never a diagnosis.
• The M shape fired on 3 of 72 photos, since it needs a fate line at medium. It says nothing about how rare an M is.
• Which hand and which way it faces came out right on 99.3% and 99.2% of 752 flat hands with known labels. That wasn't tested on labelled real-world photos.

Hand type: the splits and how steady they are

The tradition gives no numbers for its four types, so the 1.60 and 0.92 splits are ours. They sit at the middle of 375 palm photos from 11k Hands, a research dataset of hand photos: palm ratio median 1.60, spread 0.14, finger ratio median 0.92, spread 0.09. Those photos are mostly of people in their twenties, from one study, so the splits are a fair split of that sample and not a law of hands. The four types aren't equally common under these splits: 146 Fire, 138 Air, 46 Earth, 45 Water.

The catch is repeatability. Across 50 hands and 663 photos in a fixed rig, one hand got the same type 73% of the time. When both photos sat at least half a spread clear of the lines, that rose to 98%. That is why a close call is shown as a lean, and why your first result is kept for the session.

Finger ratio: how far it moves

Our 2D:4D figure is an estimate. It has a median of 0.976, and the spread within one hand across photos is about 0.015, about half the spread between people, 0.031. Don't read anything into small differences.

What's tradition

Everything that says what a line means. None of it is measured, and nothing in the code can check it. The words come from a fixed library of sentences written in advance, drawing on old palmistry books (Cheiro, Markun, Williams, Dale) and modern pages, quoted and named where possible. Your measurements choose which sentences you get. One scan always gets the same wording, and moving a line changes only what that line changes.

The label decides how firmly the library may speak:

The fate line is an invitation until you place it. Children lines read nothing even when placed. Automated tests run every combination of features through the library and fail on a list of words covering death, illness, lifespan, pregnancy, curses and doom. The free reading on an example hand, and the offer that follows it, are on the sample reading.

Does any of this show palm reading works?

No. Measuring a crease accurately says nothing about whether the crease means anything. There is no good evidence that palm reading predicts anything, and it is widely viewed as a pseudoscience. Three studies have tested the best-known claim, that the life line tells you how long you will live: one reported a link in 1990, and two others, in 1974 and 2019, found none. Biologically, the 2019 authors write, the line of life "is merely another flexion crease". A reading can still feel true, for well-studied reasons: in 1949 Forer gave 39 students the same description, taken from a newsstand astrology book, and they rated it about 4.3 out of 5 for accuracy.

We keep the tradition because people find it interesting and useful for thinking, and we label it as tradition wherever it appears. The full evidence is on is palm reading real.

The photos behind the test

139 photos: 67 for tuning and 72 held out, split by a hash of the person or photographer before anyone looked at them. 55 are from 11k Hands, which is for research and isn't shipped with the site or shown on it. 36 are from Wikimedia Commons, and 48 from Flickr and the WordPress Photo Directory, all under their own open licences. The one hand shown on the site is a CC BY 2.0 photo, credited wherever it appears. Thresholds and code were tuned on the tuning set only.

What happens to your own photo is on the privacy page. If you find a mistake in any of this, or in a guide, tell us and we'll correct it.