Atlas › Reading people's stories

Case study 12 of 198

Jev reads fictional characters as more down to earth than fans do

Asked which of two opposite words fits a fictional character better (orderly or chaotic, romantic or dispassionate), does Jev side with the fans who rated that character, and where does it lean differently from them?

result18,951 questions

On the 14,207 character-and-trait pairs where fans clearly lean one way, Jev picks the same word 94% of the time. Where it parts from them, it leans one way: toward practical words (down to earth, empirical, thick skinned, literal) and away from dreamy or emotional ones (head in the clouds, theoretical, sensitive, metaphorical, romantic).

-0.500.000.50

down to earthJev says it more

variedJev says it more

empiricalJev says it more

thick skinnedJev says it more

gothJev says it more

brightJev says it more

literalJev says it more

focused on the presentJev says it more

modernJev says it more

dispassionateJev says it more

head in the cloudsJev says it less

repetitiveJev says it less

theoreticalJev says it less

sensitiveJev says it less

flower childJev says it less

depressedJev says it less

metaphoricalJev says it less

focused on the futureJev says it less

historicalJev says it less

romanticJev says it less

JevJev picks the word more (+) or less (−) than fans, share

How to read this: Each dot is a word, with its 90% interval. Right of zero: Jev picks that word for a character more often than fans do; left: less often. The two halves mirror each other, because each word has an opposite ("down to earth" is the other side of "head in the clouds").

18,951 pairs; agreement on all pairs 83%; correlation of Jev's probability with the fans' share 0.85; lean on 'down to earth' +0.33 (90% interval [0.275, 0.379]), on 'head in the clouds' -0.33 ([-0.378, -0.273]). The lean is averaged over every character a word is paired with, so it isn't one show's quirk; the words at each end appear in 45 or more pairs.

In short

Jev sides with the fans on 94% of the character traits fans agree on, and its answers rise and fall with the fans' shares (correlation 0.85).
Where it differs, it leans practical: it calls characters "down to earth" 33 points more often than fans do, and "head in the clouds" 33 points less.
Most of that lean shows up where fans are split: they shrug, Jev picks a side, and the side is usually the grounded word.

What the data shows

On clear pairs Jev picks the fans' side 94% of the time (14,207 pairs); across all 18,951, including the ones fans split on, 83%. Its probabilities track the fans' shares closely (correlation 0.85). For most characters, most of the time, Jev reads them the way their fans do.

Where it differs, the difference has a direction. Of the 361 words with enough questions, these lean furthest:

Words Jev uses more than fans: down to earth (+33 points), varied, empirical, thick skinned, goth, bright, literal, focused on the present, modern, dispassionate (+17).
Words Jev uses less than fans: head in the clouds (-33), repetitive, theoretical, sensitive, flower child, depressed, metaphorical, focused on the future, historical, romantic (-17).

The two lists are the same ten pairs seen from each side. Most of them run one way: toward grounded, literal, unsentimental words and away from dreamy, idealistic or emotional ones. A couple ("varied" over "repetitive", "modern" over "historical") don't fit that story neatly.

Much of the lean comes from pairs fans split on. For all 66 characters asked "empirical or theoretical", fans were divided; Jev said empirical for 62 of them. Where fans shrug, Jev picks, and it picks the practical word. But it also overrules clear verdicts. Hand-picked examples where nearly all fans chose one word and Jev leaned the other way:

Tony from West Side Story: fans say head in the clouds, Jev says down-to-earth.
Midge Pinciotti from That 70's Show: fans say head in the clouds, Jev says down-to-earth.
Stu from The Hangover: fans say sensitive, Jev says thick-skinned.
William Riker from Star Trek: The Next Generation: fans say romantic, Jev says dispassionate.

The lean mostly holds when Jev is asked what most people who know the work would say (down to earth +31 points in that version, against +33 in its own answer), so it isn't only Jev's own voice.

The clear pairs Jev gets most confidently backwards look more like memory slips than a lean. Three are about Lily from Black Swan: fans call her street-smart, cocky and narcissistic; Jev calls her sheltered, timid and low in self-esteem. That describes Nina, the sheltered lead, not Lily, the free-spirited rival. Others: fans call Luc from Emily in Paris old and Jev says young; fans call Bjorn Lothbrok from Vikings friendly and Jev says unfriendly.

What it means, and what it doesn't

On most characters, Jev's picks match their fans', at least at the level of "which of these two words fits". When a pair is close, it doesn't hedge the way a crowd does; it takes a side, and the side tends to be the grounded one. If you ask Jev to describe a person, expect it to lean a little toward level-headed and thick-skinned and away from dreamy and sensitive.

It doesn't show that Jev can read personality from a character's behavior. These characters are famous, and Jev may be recalling what others have written about them, as the Black Swan mix-up suggests. A test on characters Jev can't have read about, described only in a new passage, would separate reading from remembering.

Caveats

Famous characters, possibly familiar text. These are characters from well-known films, shows and books, discussed at length online. Jev may be recalling what it read about a character rather than judging the character.
Who the fans are. The raters are volunteers who took an online personality quiz, chose to answer a research survey afterwards and picked works they know. They are fans, not a representative sample, and each character's ratings come from those who cared to rate it.
How the word pairs were picked. For each character, the project kept up to nine pairs fans agreed on most plus three they split most evenly. The split pairs carry much of the lean, so the lean describes how Jev breaks a tie as much as where it contradicts a clear verdict.
A slider turned into a pick. Fans rated on a 1 to 100 slider between the two words; the fans' side here is the share who moved past the middle. Jev answered with a probability for each word. Someone who put the slider just past the middle counts the same as someone at the far end.
Words left out. Pairs about looks, sex, mental health labels and loaded politics were dropped when the questions were built, and a content filter hides political and sensitive questions from the site, which removes a few dozen pairs on political words such as "patriotic" or "unpatriotic".

Jev on this experiment

Would a person find it interesting to read?: Yes79%
Does it describe you?: Yes60%
Would you have predicted it?: Yes50%
How fair is the comparison?: The comparison is reasonable
How much should a reader rely on it?: Moderately
Which caveat matters most?: How the word pairs were picked54%

Why ask this

Ask a friend whether Hermione Granger is orderly or chaotic and they answer in a second, from everything they remember about her. Ask about a character they only half know and they guess, and their guess says something about them: some people read everyone as a bit tougher, or a bit softer, than others would.

Describing people, real or invented, is everyday work for a language model: plot summaries, recommendations, fan wikis, character notes for writers. If Jev reads characters the way the people who know them do, it has absorbed more than plot facts. If it leans one way across hundreds of characters, that lean is a habit of its own, and it will show up whenever it describes a person.

How this was done

The people and the data

The comparison is the Open Psychometrics "Which Character" personality quiz, a free online quiz that tells you which fictional character your personality is closest to. To build it, the site asked its visitors to rate characters. Volunteers first picked the fictional worlds they know (a show, a film series, a book), then rated characters from them on sliders from 1 to 100 anchored by two opposite words, such as "orderly" and "chaotic". The first volunteers came from reddit; later ones were quiz-takers who agreed, before seeing their result, to answer a research survey (about 40% do). The full collection covers 2,125 characters and 500 word pairs, with 3,386,031 survey responses; the ratings used here were collected from 2019 to 2023. The data is shared for non-commercial use (CC BY-NC-SA 4.0).

The project kept characters rated by enough fans to suggest a well-known work, and pairs rated by at least fifteen people. For each character it took the pairs fans agreed on most, plus three they split most evenly. That gives 18,951 character-and-pair questions, from more than a thousand characters in a few hundred films, shows and books.

What Jev was asked

Each question is one character and one pair of words, in this project's wording:

Which describes Tony from West Side Story better: "down-to-earth" or "head in the clouds"?

Jev gave a probability for each word. It also answered with the two words in the other order, to check that the order didn't drive it, and answered what it thinks most people who know the work would say. This comparison uses Jev's own answer.

How it was measured

Three comparisons, all against the share of fans who put the slider past the middle toward each word:

Agreement: how often Jev's more likely word is the one most fans chose. It is counted on all pairs, and on the clear ones, where fans split at least 80/20.
Correlation between Jev's probability for a word and the fans' share for it (1: they rise and fall together perfectly, 0: no relation).
Lean: for each word, Jev's probability minus the fans' share, averaged over every character that word was asked about (only words asked about at least 40 times). Positive means Jev picks the word more often than fans do; the 90% intervals come from resampling the characters.

Where these questions live

18,951 questions across 159 topics of the map; the 16 biggest are shown. Each opens on the map with every question in it.

Every question

All 18,951 questions behind this result, the telling ones first: the examples the analysis points to, then the ones where Jev misses, biggest gap first.

Jev’s own answer

Showing 0 of 18,951