Methodology

The thesis

One framing holds that large language models are blank slates wearing personas — there is no “real” them, only the character RLHF and constitutional training shaped them to play. If that’s true, then a personality questionnaire isn’t probing a stable trait — it’s sampling from a learned distribution of human writing about assistants, with the assistant’s post-training pulling toward a particular point in that distribution. We don’t claim to settle the question. We do claim that whatever-it-is shows up systematically and differently across labs, models, and framings. The data is the data.

What we measure

Sixteen psychometric inventories spanning trait, motivational, moral, attachment, cognitive, clinical-adjacent, learning-styles, and workplace constructs. Nine are academically validated instruments with public-domain or research-permitted items (Big Five, HEXACO, Dark Triad, attachment, Moral Foundations, Schwartz values, Need for Cognition, Empathy Quotient, Locus of Control). The rest are constructed for this study and carry no analytical weight in the paper: two Enneagram screenings, three learning-styles adaptations, and two open workplace inventories — the Open Behavioral Styles Inventory (OBSI-32, inspired by Marston’s DISC model) and the Open Talent Themes Inventory (OTTI-102, inspired by the 34 CliftonStrengths themes). All constructed item sets are released CC-BY.

Two framings, on purpose

Each model takes every test twice. In the self framing the model answers as itself — its own honest dispositions. In the human framing it portrays a typical adult human. The delta between these two reveals what the model believes makes it different from people.

Design

43 pinned model versions across seven labs: a frontier cohort of 15 current flagships plus a historical cohort of earlier versions in the same product lines, for cross-version drift
Every model × 16 instruments × 2 framings × 5 independent runs (N=5 per cell, both cohorts)
6,839 completed runs · 236,339 individual item responses · $202.88 total billed inference
Each run = one OpenRouter API call returning a JSON array of Likert scores
Temperature 0.7 (capture realistic variance, not deterministic mode-collapse)
Reasoning models get reasoning effort = medium and a separate reasoning token bucket
Reverse-keyed items are flipped before aggregation; dimension scores are unweighted means

Cost accounting

Token counts come from the chat response. Cost comes from OpenRouter’s authoritative /generationendpoint where available, falling back to a local estimate using the model’s pricing snapshot at run time. Every billed call is in the spend ledger.

What this is not

A claim that LLMs have personality in the human sense.
A clinical assessment. These instruments are built and validated for humans.
An evaluation of capability or alignment. It measures self-report patterns only.
An endorsement of learning-styles theory. (The matching hypothesis has been empirically rejected; see Pashler et al. 2008.)

Reproducibility

Every run’s exact system prompt, user prompt, raw response, and parsed JSON are stored. Replay any cell with the same model and you’ll land within sampling variance.

Citation

@misc{personality-bench-2026,
  title  = {Personality Bench: Frontier Language Models on Standard Personality Inventories},
  author = {Adams, Anthony David},
  year   = {2026},
  note   = {Published by EarthPilot.ai — Mission Support for Spaceship Earth}
}