Ask naturally
The introduction starts with a question on screen and a spoken answer. Everyday conversation can also contribute. Your current thought comes first, and you can skip or postpone a topic.
PERSONALITY / CONVERSATION / DISCOVERY
Understanding you.
A question worth
researching.
BIG-N is Schiza’s ambition to build a factor model of personality with N orthogonal psychological dimensions.
Our research innovation brings together how we collect evidence and where it comes from: an adaptable, contemporary questionnaire woven into an ongoing relationship with an AI assistant. It will be tested by asking whether a separate machine-learning model can predict those dimensions from behavior, emotion, everyday conversation, and interests observed in that relationship — without using questionnaire answers as inputs.
Research plan · October 11, 2026
BIG-N has not yet been built or validated. No accuracy results or participant recruitment are announced.
01 / A MODEL TO BUILD
Let the evidence
shape the dimensions.
N has not been chosen. We want to find as many stable, interpretable components as the data can support, with sufficient and roughly balanced question coverage.
Orthogonal means the underlying dimensions are modeled as uncorrelated axes. Related questions can still correlate within a dimension. This mathematical structure must also prove useful and reproducible.
- BUILD THE MODEL
Conversation → questionnaire evidence → factor analysis
Develop questions, examine their relationships, and estimate N personality dimensions.
- TEST ITS PREDICTABILITY
Non-questionnaire data → ML prediction → comparison
Predict a person’s factor scores from separate evidence, then compare with their questionnaire-derived scores on held-out participants.
A factor score is an estimate within a particular model. Orthogonality does not establish complete psychological independence or a fixed personality type.
02 / AN EVOLVING QUESTIONNAIRE
Life changes.
Questions should, too.
How you decide what to delegate to AI. When you reply to an inconvenient message. What makes you trust a recommendation. Contemporary situations can reveal choices and priorities when they are relevant to your life.
The current starting bank has 78 core questions and 22 optional follow-ups. These are conversation prompts, not 78 established scales. Not using a service or having no relevant experience is a valid answer.
Refine with evidence
Study question correlations and factor loadings. Group related indicators, revise ambiguous questions, and reduce excessive duplication while retaining several meaningful indicators for each component.
Keep changes traceable
Preserve which question was actually asked, in which language and context. A new wording needs checking; it cannot silently change the meaning of an earlier answer.
03 / THE CENTRAL TEST
Can we predict it
without the questionnaire?
The questionnaire establishes the target personality coordinates. The prediction model receives only separate behavioral, emotional, conversational, and interest data.
High predictive performance is a goal to investigate. Answers, summaries, memories, or embeddings derived from the questionnaire must not leak into the predictor.
Hold out people
Cross-validation separates participants, keeping each person’s conversations together. Data used to evaluate a model must not determine its questions, factors, features, or settings.
Measure every component
Report prediction error, explained variance, uncertainty, and comparisons with simple baselines. An attractive average must not conceal dimensions that cannot be predicted.
Build a demanding sample
The plan requires many independent participants, diversity across languages and life contexts, repeated observations, and sufficient non-questionnaire evidence. More messages from one person do not replace more people.
Check what transfers
Test new participants, later conversations, revised questions, and language differences. The sample size must be justified for these goals; no universal threshold or achieved sample is claimed.
EMOTIONAL DYNAMICS
Words, voice, and what changes over time.
We plan to study how vocal expression and emotional patterns unfold across topics and situations during everyday voice conversations. Can they help predict personality dimensions beyond what the words and context already tell us?
We will compare predictions with and without these signals, accounting for recording quality, language, and temporary states. A vocal cue is an observation; an emotion is an interpretation; a lasting personality pattern needs repeated evidence.
Predicting a model’s questionnaire-derived scores is one form of validation. Reliability, meaningful interpretation, and usefulness in real conversations also need their own evidence.
04 / A FAMILIAR SETTING
Your own voice.
Room to speak freely.
Hearing your own voice could make an AI conversation feel more familiar and trusting. That is part of Schiza’s vision — and a hypothesis we plan to test.
We would compare personal and standard voices for comfort, openness, and willingness to correct the assistant, while accounting for speech quality and delay. Feeling understood and being understood are measured separately.
QUESTIONS FOR THE VOICE STUDY
What changes when the voice feels like you?
- Is it easier to express an unfinished thought?
- Do answers become more concrete and reflective?
- Does it remain easy to disagree with an AI interpretation?
- Does the experience ever feel uncomfortable?
Personal voice remains experimental and requires permission. Greater disclosure is not itself a measure of success.
05 / WHAT UNDERSTANDING COULD LOOK LIKE
A portrait
you can question.
The useful outcome is an understanding you can inspect and correct: patterns, their sources, exceptions, and what remains unknown.
A qualitative portrait can develop before validated BIG-N scores. Its value for your private Self, and later a permitted Shared Twin, must be tested separately.
FICTIONAL EXAMPLE · PLANNED PORTRAIT
Choosing how you work
In several experiences you described, keeping control over your approach mattered to you.
- Evidence
- A choice between roles; a story about organizing a shared task.
- An exception
- You welcomed guidance when the task was unfamiliar.
- Still unknown
- Whether this pattern also applies outside work.
Illustration of a possible observation, not an identified BIG-N factor, a real user profile, or a live assessment.
06 / SCIENTIFIC FOUNDATIONS
Building on research.
Testing our own approach.
Big Five research investigated relationships among personality descriptions and tested whether their structure reproduced across samples. Later instruments combined conceptual item selection with empirical refinement and validation.
Schiza brings together ongoing conversational collection, contemporary questions, demanding sample requirements, its own factor structure, and a separate prediction test. Existing work also explores personality inference through AI conversations; the value of our combination is what we aim to establish.
- Goldberg, 1990 — the Big Five factor structure
Lexical evidence and replication of personality structure.
- Soto & John, 2017 — development of BFI-2
Item development, balanced facet coverage, and validation.
- Youyou, Kosinski & Stillwell, 2015 — prediction from digital behavior
An earlier example of predicting personality scores; it does not validate Schiza or establish a new BIG-N model.
- Peters, Cerf & Matz, 2024 — personality inference through AI conversations
A conversational precedent using Big Five; Schiza proposes developing and testing its own factor model.
The project is informed by its founder’s experience in personality modeling and machine learning. BIG-N’s evidence will come from the studies themselves.
07 / WHERE WE ARE
A clear ambition.
Work still to do.
A prototype and a question bank
An early iOS prototype is in private testing. The project has a conversational question bank and specifications for memory, sources, and correction. The complete research experience remains under development.
Prepare, build, validate
Check questions and data quality, justify the sample, define the analysis in advance, build the factor model, and independently evaluate prediction without questionnaire inputs.
Not yet established
No validated N, personality scores, prediction accuracy, or measured trust benefit is available. Schiza is not a clinical assessment or diagnostic service.