Skip to content
human.ai

About the reports

Method and limits

For each report: what it is built on, how it is scored, who it compares you with, and what it cannot tell you.

Shared ground rules

Instruments as published
The questionnaire reports use the items, response options and scoring of published scales as their sources give them. Any departure is stated in that report's method chapter, and Korean items we translated ourselves are labelled as an in-house translation.
Scored and kept on your device
Your answers are scored on this device by fixed rules. The same answers always give the same scores. No sign-up is needed: your answers and reports stay on this device, and your answer to a question never leaves it.
Every reading is anchored
Each interpretive sentence carries the score or the posts it rests on, and says how firmly it is held. A statement that would fit anyone is treated as a defect.
Never a diagnosis
No report says that you have or do not have a condition, and nothing sensitive is inferred about anyone: no health conditions, no sexual orientation, no political or religious views.

Questionnaire reports

This is a summary. The arithmetic and the thresholds are set out in full in the last chapter of each report, “Method and limits”, and the licence and translation notes for every source are on the report's own page.

LOVE

Romantic style report

How you tend to love, read from your Big Five scores

Built on
  • IPIP-NEO-120View source: IPIP-NEO-120The five Big Five domains and 29 facets. Four of the 120 items, which ask about voting and beliefs, are not asked, which leaves 116.
  • TLS-15View source: TLS-15Intimacy, passion and commitment: 15 items, asked only if you say you are in a relationship now.
Scoring

Each item is answered from 1 to 5, and items worded the opposite way are reversed before adding. A facet is the sum of its four items; a domain is the sum of all its items. The four-letter type code records, one letter each, whether four of the domain scores sit above or below the comparison group's mean. That rule is Human.ai's own, and the code may differ from a four-letter code you got elsewhere.

Comparison group

The 619,150 online respondents in the data Johnson has made public. They chose to take the inventory themselves, in English, and this is not a Korean norm. Percentiles are estimated from that group's mean and standard deviation. The love triangle is compared with the sample means printed in its source article; for answers given in Korean that comparison is approximate, because the Korean items are our own translation.

What it cannot tell you
  • It rests on your own answers, which can differ from how others see you and from what you actually do.
  • Each facet has only four items and is less steady than a domain, so it is shown as a range, not a point.
  • It gives no compatibility score, and no pairing of type codes has been shown to go well or badly together.
  • It describes tendencies of personality and makes no judgement about anyone's health.
Who writes the text

The explanation of each score is written in advance and selected by scoring rules. If you ask and agree, generative AI adds chapters that read the scores together, each one marked “AI-generated”.

The report's own page, with every source
BOND

Relationship security report

How settled and how open you feel in close relationships

Built on
Scoring

Each statement is answered from 1 to 5; statements worded the opposite way are flipped, and each scale is the average of the statements answered. A scale is scored only when at least 80% of its statements were answered. A position is put into words by whether the average is below 2.5 or above 3.5, which is a convention of Human.ai's and not a published cut-off.

Comparison group

None. No means or standard deviations are published for these three scales and Human.ai has collected no comparison data, so the report shows no percentile and no comparison with other people. It shows only where the average of your answers sits on the response scale.

What it cannot tell you
  • It does not assign an attachment style. The scales used were not built to measure attachment.
  • It cannot describe your relationship with one particular person: the statements mix partners, friends and people in general.
  • The scales are short, 7 to 9 statements each, so small differences in the first decimal place should not be given meaning.
  • It says nothing about the state of your health.
Who writes the text

The explanation of each score is written in advance and selected by scoring rules. If you ask and agree, generative AI adds chapters that read the scores together, each one marked “AI-generated”.

The report's own page, with every source
IPC

Interpersonal style report

Where you stand among other people

Built on
  • IPIP-IPCView source: IPIP-IPCThe two axes of the interpersonal circle, dominance and warmth: 4 statements for each of eight directions, 32 in all.
Scoring

Each direction's average is compared with the comparison group, and the published formula combines the eight results into two axis scores, dominance and warmth. The arrow's direction and length come from those two scores.

Comparison group

The sample in the source article: 501 adults in the United States, aged 22 to 90. The axis percentiles are estimates worked out from the article's table of correlations, and no Korean comparison data exist.

What it cannot tell you
  • It is a self-description, which can differ from how others experience you.
  • 32 statements give a quick sketch. A result near the border between two styles can change on a retake.
  • A single direction rests on four statements and is not reliable on its own, so the eight-direction chart is read as a shape and not as eight scores.
  • It does not measure how well you get on with any particular person.
Who writes the text

The explanation of each score is written in advance and selected by scoring rules. If you ask and agree, generative AI adds chapters that read the scores together, each one marked “AI-generated”.

The report's own page, with every source
MOODSelf-check

Mood check-in

How have the last two weeks been?

Built on
  • PHQ-9View source: PHQ-9How often, over the last two weeks, you had nine experiences to do with mood, interest, sleep, energy, appetite and concentration. The official Korean version is used unchanged.
  • GAD-7View source: GAD-7How often, over the same two weeks, you had seven experiences to do with tension and worry. The official Korean version is used unchanged.
  • IPIP CAT-PD-SF: PerfectionismView source: IPIP CAT-PD-SF: PerfectionismHow much you usually aim for perfection (6 statements): an everyday tendency, not the last two weeks.
  • IPIP CAT-PD-SF: Emotional DetachmentView source: IPIP CAT-PD-SF: Emotional DetachmentHow much you usually keep feelings in (7 statements).
Scoring

The PHQ-9 (0 to 27) and the GAD-7 (0 to 21) add up the answer chosen for each question (0 to 3). The range boundaries are the ones in the developers' manual; the range names, which only say how often, are Human.ai's own. Each tendency is the average of its answers (1 to 5). The four scores are never combined.

Comparison group

For the PHQ-9 total the report shows, for reference, the distribution found in a national survey of 10,759 Korean adults. There is no comparison group for the GAD-7 or the two tendencies, so the report never says those are higher or lower than other people's.

What it cannot tell you
  • It covers the last two weeks only, and cannot tell something long-standing from something set off by recent events.
  • It cannot say why. Short sleep, heavy stress or a physical illness can all produce the same answers.
  • The same person can land a point or two apart from one day to the next, so a one- or two-point difference means little.
  • No score and no interpretation comes from the drawing activity.
Who writes the text

Every sentence is written in advance and selected by scoring rules. No AI writes or reads any part of it, and nothing you answer is sent off the device.

The report's own page, with every source
FOCUSSelf-check

Focus style report

Your everyday focus and how today went

Built on
Scoring

The answers to the 12 statements (1 to 5) are added into a total of 12 to 60. The tasks record error counts and response times, and the stopping time is estimated the way a consensus guide written by researchers in the field recommends. A run that was interrupted, or shown on a display that redrew too slowly, is left unread under rules fixed in advance.

Comparison group

None. No reference sample could be verified for the statements, and none has been published for these tasks on touch devices. The study values quoted in the report are context only and are never turned into a score or a rank.

What it cannot tell you
  • Task results change with sleep, caffeine, noise and the device you use, and every device adds a different display and touch delay.
  • A single session is imprecise, even in a laboratory with keyboards.
  • A task of about four minutes cannot measure attention over a long stretch.
  • The statements are for looking at your own everyday tendencies. They give no position relative to other people.
Who writes the text

Every sentence is written in advance and selected by scoring rules. No AI writes or reads any part of it, and nothing you answer is sent off the device.

The report's own page, with every source

Instagram feed reading

The feed reading is not an instrument that measures personality. It is a reading of the picture a feed presents: counted, sorted into categories, and filtered.

What is read

  • Only posts from your own account that you hand over yourself: screenshots or photos of posts that you pick, or the archive from Instagram's “Export your information”. On the web, connecting a professional account yourself through Instagram Login may be offered as well. No account is looked up by its handle, and Instagram is not scraped.
  • The export archive is opened on your device. Messages, contacts, location data and other people's account names are not read in.
  • Posts from the last 36 months are read, newest first, up to 60. With fewer than 12, a first look reads the chosen images instead of counting tendencies; from 12 posts some tendencies are counted, and from 30 they are counted fully.
  • For a video, only the cover frame is read.
  • A post you leave out before the analysis is not read. Only its date counts, toward posting rhythm.

How a reading is made

Code does the counting and an AI reads the photos and captions. Each step says which of them does it.

  1. The material is prepared

    Device

    Each photo is scaled down to at most 768 pixels on its long edge and saved again, which drops the location, device and capture time that were attached to it.

  2. What can be counted is counted

    Code

    Weekday and hour of posting, the gaps between posts, the length and form of captions and the colours of the photos are counted by fixed arithmetic. These numbers are not psychological measures, and they are shown only as counts.

  3. Each post is sorted into categories

    AI

    For every post an AI sorts what the photo shows (how many people, the setting, the activity, how staged it is and so on) and what the caption is doing into fixed categories, and describes the photo in 25 words or fewer. It does not name anyone or guess who they are, and it does not read a mood into a face or a caption. This step describes; it does not interpret.

  4. The feed type is calculated

    Code

    The categories are counted into scores on six axes, and the type whose reference shape is nearest becomes the feed type. The share of 200 resamples of the posts that return the same type is shown with it as “signal strength”. The observations are written by code from these same numbers.

  5. Interpretations are proposed

    AI

    Five readings, each from a different angle, run separately and cannot see one another's output. They work from the categories, not from the photos, and every proposal has to name the posts that support it, the posts that cut against it and what would make it wrong.

  6. Rules filter and tier them

    Code

    A proposal that cites a post that does not exist, or a number that does not match the count, is removed. Proposals that say the same thing are merged, and two that contradict each other survive only when each clearly holds in its own context. Then the requirements below are applied.

  7. A critic goes through them

    AI and code

    A separate, adversarial pass judges each proposal: does the evidence support it, would it be true of most people, is it a stereotype, does it touch anything sensitive? What happens next is decided in code. A proposal flagged as sensitive is removed, never reworded.

  8. They are swapped against other feeds

    AI and code

    Each statement, with its numbers and post ids removed, is put beside summaries of the real feed and of three decoy feeds, and a judge says which feeds it fits. A statement that does not fit the real feed, or that also fits two or more decoys, is removed, because a reading that cannot be told apart from someone else's is not a reading.

  9. The wording is finished, then checked again

    AI and code

    Only what has passed is put into final wording. Code then checks the finished text: does every cited post exist, does every number in a sentence equal the recount, is any ruled-out wording present? A sentence that fails is rewritten once, and removed if it fails again.

The AI proposes, judges and rewords. What is shown is decided by rules, and an interpretation can only lose ground on the way through.

The three tiers

Observation

A sentence written by code from a count. No AI writes or rewrites it, and every number in it can be counted again. A report carries 3 to 5.

Counts about people and places are counts of the category an AI gave each post. A post can be categorised wrongly, so the posts behind each number are shown with it.

Pattern

It needs at least 5 supporting posts, and at least 15% of the posts read. It needs two or more kinds of evidence other than colour, audience response and posting rhythm, and the posts against it may not exceed a third of the posts for it.

It must have been proposed independently by two or more readings, or rest on an axis whose score is clearly at one end (70 or above, or 30 or below). It always carries what would make it wrong. A report has at most 4 (2 when only some tendencies can be counted).

Question

It needs at least 3 supporting posts and has to be worded as a question, not a conclusion. A question about something a feed rarely shows appears only when code worked that out from the counts. A report has at most 3.

What it never does

  • Colour, brightness, posting hour, gaps, the number of selfies and the number of likes are shown as counts only, and are never evidence for a psychological interpretation.
  • It says nothing about conscientiousness, agreeableness or emotional stability, the traits an AI reading text has been reported to misjudge in one direction.
  • It assigns no attachment style, gives no verdict on whether a self-presentation is genuine, and predicts nothing.
  • It does not name the people in your photos or guess who they are.
  • A post with content such as someone who appears to be a minor, a medical setting, a mention of health, a memorial, or drinking or smoking is evidence for no interpretation at all. Posts about religion or politics are left out of interpretations about values.
  • A post with language about harming oneself is not interpreted and no level of risk is assigned. The report carries contacts for help instead.

The accuracy ceiling

Many studies have tried to infer personality from digital traces such as posts, and the result is consistently low: correlations with personality questionnaires of roughly 0.2 to 0.4.

A correlation of 0 means unrelated and 1 means identical. “In 100” is how many people that correlation would place on the right side of the average, above or below. Picking with no information gets 50.

  • The authors of the pooled study conclude that predictions of this kind cannot be used to assess an individual.
  • These figures come from studies of text and activity on Facebook, mostly among English speakers. How well Human.ai's feed reading agrees with a questionnaire has not been measured.
  • So the feed reading is not an instrument that measures personality. It is a reading of the picture a feed presents, with its evidence. A statement about the feed can be checked by counting; a statement about you is a hypothesis. If an interpretation does not fit, it may simply be wrong.

About the feed type

The six axes and eleven feed types are this product's own classification. Their thresholds are provisional and have not been validated on real feeds. A type is a name for a feed, not a personality type.

Similarity says how close a feed is to a type's reference shape, from 0 to 100. It is not a chance of being right. Signal strength says how little the type depends on a handful of posts.

Sources and licences

  1. Marengo & Montag (2020)

    Marengo, D., & Montag, C. (2020). Digital phenotyping of Big Five personality via Facebook data mining: a meta-analysis. Digital Psychology, 1(1), 52-64.

    View source
    Licence
    Cited as the source of a published finding or formula. No text or data from the work is reproduced.

  2. Peters & Matz (2024)

    Peters, H., & Matz, S. C. (2024). Large language models can infer psychological dispositions of social media users. PNAS Nexus, 3(6), pgae231.

    View source
    Licence
    Cited as the source of a published finding or formula. No text or data from the work is reproduced.

  3. Marengo, Montag & Settanni (2025)

    Marengo, D., Montag, C., & Settanni, M. (2025). Inferring personality from social media activity using large language models: cross-model agreement, temporal stability, and convergent validity with self-reports. Journal of Personality, 94(4), 525-534.

    View source
    Licence
    Cited as the source of a published finding or formula. No text or data from the work is reproduced.

  4. Kim & Jo (2016)

    Kim, E.-K., & Jo, H.-H. (2016). Measuring burstiness for finite event sequences. Physical Review E, 94, 032311.

    View source
    Licence
    Cited as the source of a published finding or formula. No text or data from the work is reproduced.

Human.ai is not affiliated with, sponsored by, or endorsed by Instagram or Meta Platforms, Inc.

Text written by AI

Which reports can contain AI-written text and which never do, whose AI it is, and where it runs.

Where AI-written text appears

  • The personality reports (Romantic style report, Relationship security report, Interpersonal style report) are complete without AI: the explanation of each score is written in advance and selected by scoring rules. If you ask and agree, chapters that read the scores together are added. Those chapters are written by generative AI, and each one is marked “AI-generated”.
  • In the Instagram feed reading, the sentences of the pattern and question chapters are proposed and worded by AI. The observations and the feed type are produced by code from counts.
  • The self-checks (Mood check-in, Focus style report) contain no AI-written text. Every sentence is written in advance, and none of your answers, drawings, task records or scores is sent off the device.
  • AI interpretations are written by generative AI. They are not written by a person or individually reviewed by a professional.
  • AI-written sentences are checked in code before they are shown. In a personality report, each statement must name the scores it rests on and must be free of wording the report does not use; in a feed reading, an interpretation goes through every step described above.

The AI service, and where it runs

Recipient

Anthropic, PBC (Claude API)

Destination country

United States

The Human.ai server pins every request so that the AI inference runs in the United States.

What is transferred

AI interpretation of a personality report · Your scored results (scale scores with percentiles and bands), your type, response-quality indicators, the time you finished, the kind and version of the report, and the display language. Your answers to individual questions are not sent.

Instagram feed reading · Downscaled photos and captions of the posts you selected from your own account.

Human.ai has no accounts, so no sign-up details such as a name or an email address are sent.

When and how

Each time you ask, through the Human.ai server, over an encrypted connection (TLS).

How long the recipient keeps it

Deleted within 30 days of receipt or generation. Inputs and outputs flagged as a usage-policy violation may be kept for up to 2 years and their classification scores for up to 7 years, and anything the law requires to be kept is kept for as long as it requires.

It is not used to train AI models.

What Human.ai keeps

AI interpretation of a personality report · The Human.ai server only passes it on. It does not store or log a copy. The finished interpretation is saved only on this device.

Instagram feed reading · The Human.ai server handles the photos and captions in memory only and does not store them. The finished report is saved only on this device.

Withdrawing consent

You can withdraw at any time under “AI interpretation” in Settings. Nothing is sent after that. What was already sent is deleted on the schedule above.

Manage AI consent in SettingsPrivacy policy

If you need help

  • Human.ai is not a crisis service, and a report does not take the place of professional help.
  • If you are struggling right now or are worried about your safety, please do not go through it alone. Places you can contact straight away are listed under “If you need help” in Support and in Settings.