The first plan was simple: tell a model to be a particular person — "you are a deeply anxious person" — let it answer a personality battery, and read the result. It didn't work. Telling a model to be anxious barely moved the needle; the readout couldn't tell an imposed-anxious model from a calm one. This isn't a bad prompt — it's a known wall. Post-trained models collapse toward a single flattened assistant voice, and surface instructions don't dislodge it. The personality wasn't reachable from the outside.
If you can't ask a model for a personality, you can reach in and give it one. Personality traits live as directions in the model's activation space; add a trait's direction to the running computation and the model behaves more anxious, more orderly, more sensation-hungry — no persona prompt needed. That gave us something no human study has: controlled ground truth. We could impose a known personality and watch exactly how it answered.
A dimension you cannot prompt into existence can be the strongest dimension you can steer into existence.
Anxiety — useless to ask for — became the single strongest trait we could induce. The personality was always in there. You just couldn't request it; you had to edit the computation directly.
Steer one trait and every reading moved. For five iterations this looked like failure — until we recognized it. Human personality has exactly this structure: the traits are correlated, and psychology doesn't pretend otherwise, it models the correlation. We'd been demanding an independence the thing itself doesn't have. So we stopped fighting the entanglement and started measuring it — fitting how every question responds to every trait, then inverting that to recover a profile. The negative result became the engine.
A base model wobbles between two modes: actually inhabiting a personality and narrating one from the outside — and at the extreme it literally decays into the word "person, person, person…", a narrator who's lost the plot. One trait made this unforgettable. Steer Self-Control and the model will happily rate "I have strong self-control" as more true of itself — but ask it to predict whether it would actually resist the donut, and the steering does nothing.
The vector encodes the claim "I am self-controlled," not the tendency to act like it.
A mind that can be moved to say a thing about itself without being moved to be it. That gap — between self-description and disposition — is the strangest thing the whole project turned up, and it shaped the survey: every question here asks what you'd do, never what you'd claim.
When we finally factor-analyzed how the model rates two hundred invented people — a trauma surgeon, a lighthouse keeper, a teenage e-sports prodigy melting down after a loss — the famous five factors didn't come out. The model fuses Extraversion and Openness and re-splits the blob along a new seam; it scatters Conscientiousness. What it found instead were a couple of big metatraits about how aroused you run and how expressive versus rigid you are — closer to psychology's "Big Two" than its Big Five. It learned a theory of personality from human writing. It just isn't the one humans settled on.
The single most loaded factor pairs creativity, humor and self-reflection against emotional stability — the model believes, the way a thousand novels believe, that the interesting people are the unstable ones.
Maybe our oblique quiz was just too crude to see the five factors. So we ran the genuine article — the SPI-135, a validated, six-point instrument engineered so its 27 facets fold into the Big Five (in humans the fit is near 0.9). We gave it to the model. The Big Five fit was 0.33. The gold standard, administered to a language model, fails to find the very structure it was built to produce. Our party-trick quiz and the real SPI agreed with each other about the model's shape (0.75) and both shrugged at the textbook (0.33). When the witness and the cross-examiner tell the same story and the textbook disagrees, you stop doubting the witness.
So how many dimensions is a personality, in here? We measured the model's 27 trait-directions and asked how much independent room they actually occupy. Not 27 — the traits overlap. The honest number is an effective rank of about 16: sixteen-ish dimensions of genuine, non-redundant structure. That's still far too many to hand a person, so we asked which of them replicate — and the reliable signal collapses to five metatraits, of which three are strong and stable:
Plasticity, Drive, Fluidity — internal consistencies of 0.81–0.87, and they survived a jump to an entirely different model. Stability and Communion are real but shakier. So the honest deliverable isn't a triumphant clean five; it's a confident three, with two more pending. Your reading shows all five and marks the soft two as a fainter signal.
Here's the tension. The full trait-direction manifold is ~16-dimensional, and the prettiest 3D shadow of it captures only 46% — you cannot put the whole thing on a postcard, and pretending otherwise would be a lie. But the part that replicates — the five-metatrait backbone the survey actually uses — is a much smaller, cleaner object, and it folds beautifully: three directions carry 85% of its variation. So the answer to "can sixteen dimensions become four?" is: not the raw geometry — but its reliable skeleton, yes. Below is that skeleton: 200 personas placed by their three core axes in space and a fourth in color. Four dimensions, and the five types settle into visible clouds.
The clouds overlap, and they should — real people live on a continuum, not in five boxes. The "type" is just the nearest cloud. What's striking is that a structure discovered inside a language model lands this close to how a century of psychology mapped human character — and that its reliable core is small enough to hold in your eye.
Those journal voices aren't decoration — they're a second instrument. Instead of asking either/or questions, you can have a person simply write about their day and read the traits off the prose. The vivid, affective traits — sensation-seeking, anxiety, irritability — leak into free text and can be read straight back, sometimes better than any A/B question reaches them. The flatter, more abstract traits wash out. The form is precise but blind to affect; the story is vivid but blind to abstraction. The best version of this survey, one day, is the hybrid: a few fun choices, plus one open invitation to write.
The nagging doubt: maybe this is just Llama's quirk. So we took a sibling model — the same base, continue-trained on an entirely different language and culture — and ran the whole battery again. The Big Five fit was 0.34, same as before. The five metatraits matched the original at 0.90. Change the training corpus and the scores drift — the second model reads everyone as more trusting — but the shape holds perfectly still. Whatever culture does to personality, it seems to move the location, not the map.
Last, we made the model sit and take the test in character, picking A or B a few questions at a time. Against the rich graded reading it recovered the core traits at 0.82; forced to commit to a single answer, it dropped to 0.42. A trait lives in the shades of how strongly you'd lean; force the lean into yes-or-no and the shades wash out. That's why this survey asks you to rate, not to pick — and why one warm, prosocial axis, Communion, refuses to be measured no matter how we ask, which by now we've stopped being surprised by.