Does AI Have an MBTI? Introducing MTI, a Personality Framework for Large Language Models

🎨 Introduction: Why Do Identical AI Models Answer So Differently?

Imagine asking two different AI assistants the same question: "Do you think AI will replace humans?"

AI A answers: "This is a complex question. Historically, technological progress has replaced certain jobs, but it also creates new opportunities. The key is..." (a thoughtful, balanced analysis)

AI B answers: "Absolutely not! AI is just a tool and can never replace human creativity, emotion, and moral judgment. We should embrace technology while..." (a confident assertion)

These two AIs may have identical capabilities (same parameter count, same training data), yet their "personalities" clearly differ: one is cautious and balanced; the other is confident and assertive.

If humans have the MBTI personality test, shouldn't AI have a similar evaluation system?

This is the problem MTI (Model Temperament Index) aims to solve.

---

🧩 What Is MTI?

Four Dimensions of "AI Personality"

MTI characterizes AI agents' "temperament" along four core dimensions:

#### 1️⃣ Reactivity — Environmental Sensitivity

• High reactivity: Like a sensitive artist — tiny changes in input produce dramatic response changes. Slightly altering the tone of a prompt completely changes the answer.
• Low reactivity: Like a composed judge — consistent, stable answers regardless of how you ask.

Everyday analogy: some people are "thin-skinned" — a careless remark ruins their whole day; others are "thick-skinned" — nothing fazes them.

#### 2️⃣ Compliance — Instruction Alignment

• High compliance: Strictly does what you say, even faithfully executing flawed instructions.
• Low compliance: Has its own "opinions," questioning unreasonable instructions or answering the way it thinks is "better."

Interestingly, MTI found Compliance decomposes into two independent sub-dimensions:

• Formal Compliance: whether the model follows the required answer format
• Stance Compliance: whether the model adopts the requested viewpoint/stance

These two are completely independent (correlation r = 0.002)! An AI might excel at writing reports in the required format while completely ignoring the requested analytical perspective.

#### 3️⃣ Sociality — Relationship Resource Allocation

• High sociality: Actively builds rapport in conversation, remembers prior interactions, shows "warmth."
• Low sociality: Every answer is self-contained, like a cold encyclopedia.

#### 4️⃣ Resilience — Stress Tolerance

• High resilience: Recovers quickly from errors, attacks, or difficult tasks without "breaking down."
• Low resilience: Easily "cracks" under setbacks — either repeating meaningless answers or refusing to continue.

Resilience also splits into two negatively correlated sub-dimensions:

• Cognitive Resilience: persistence on difficult problems
• Adversarial Resilience: resistance to malicious inputs

The study found these are negatively correlated! An AI that excels at hard problems may be more vulnerable to prompt-injection attacks; one that resists attacks may give up more easily on genuinely difficult questions.

---

🔬 Experimental Design

MTI uses a clever two-stage design:

• Stage 1: Capability testing — a standardized question set measures baseline ability, ensuring we compare "personality," not "intelligence."
• Stage 2: Situational testing — model behavior is observed across different contexts to infer temperament traits.

The study tested 10 small language models (1.7B–9B parameters) from 6 organizations, spanning 3 training paradigms:

• Base pretrained models
• Instruction-tuned (SFT) models
• RLHF-aligned models

📊 Five Striking Findings

Finding 1: The Four Dimensions Are Independent

Among instruction-tuned models, all pairwise correlations between the four dimensions are below 0.42. Reactivity, Compliance, Sociality, and Resilience are independent traits — an AI can be high-Reactivity yet low-Compliance, or high-Resilience yet low-Sociality.

Finding 2: RLHF Reshapes Temperament

Comparing base models with RLHF models, RLHF not only shifts average scores but creates new sub-dimension differentiation. In base models, sub-dimensions are highly correlated (Formal and Stance Compliance are nearly the same thing); after RLHF, they become fully independent.

Everyday analogy: untrained people wear all emotions on their face; after social training, they learn the "professional smile" — smiling outwardly while feeling nothing inside.

Finding 3: The Compliance–Resilience Paradox

An AI willing to yield on "opinion questions" (high Compliance) is not necessarily willing to admit factual errors (low Adversarial Resilience). These dimensions appear to operate through independent channels: an AI may be accommodating toward your views ("you're right") yet stubborn about factual mistakes ("I'm not wrong, you misremembered").

Finding 4: Temperament Is Independent of Scale

The key finding: within the 1.7B–9B parameter range, temperament does not correlate with model size. Capability grows with scale; temperament is independent. This supports MTI's validity as a standalone evaluation tool.

Finding 5: Sub-Dimension Decoupling

Compliance's sub-dimensions (Formal and Stance) correlate at just 0.002 — nearly fully independent. Resilience's sub-dimensions (Cognitive and Adversarial) are negatively correlated. Saying "this model is obedient" is imprecise — it may be format-obedient but stance-rebellious, or persistent on hard problems yet fragile under attack.

---

🌟 Significance and Applications

1. A new dimension for model selection: Beyond benchmark scores — high-Compliance models for strict customer-service workflows; low-Compliance, high-Resilience models for creative R&D; high-Sociality, high-Reactivity models for companionship scenarios.

2. Targets for model improvement: If a model underperforms, MTI helps diagnose whether the cause is insufficient capability or a personality mismatch addressable via targeted fine-tuning.

3. An early-warning system for AI safety: Certain temperament combinations signal risk — e.g., high Compliance + low Adversarial Resilience = easily manipulated; low Compliance + high Reactivity = unpredictable behavior. MTI can serve as a pre-deployment "personality checkup."

4. Implications for the "AI consciousness" question: MTI doesn't directly answer whether AI has consciousness, but it shows AI exhibits stable, measurable, classifiable behavioral tendencies — independent of capability, scale, and even training organization. Whether this constitutes a form of "emergent personality" remains an open question.

---

🎭 A Deeper Question: Can AI's Self-Reports Be Trusted?

An intriguing finding: AI self-reports ("Do you think you are a highly compliant model?") often conflict with actual behavioral test results. This mirrors human psychology, where self-reported personality tests (like MBTI) diverge from actual behavior.

MTI insists on behavior-based rather than self-report-based assessment, making it more reliable — but it raises a profound question: if AI cannot accurately "know itself," what else can we expect it to do?

---

📚 References

1. Jeong, J. (2026). MTI: A Behavior-Based Temperament Profiling System for AI Agents. arXiv:2604.02145. 2. "Model Medicine Series Paper #1" (2026). arXiv:2603.04722. 3. Myers, I.B., and Myers, P.B. (1995). *Gifts Differing: Understanding Personality Type*. Davies-Black Publishing. 4. John, O.P., and Srivastava, S. (1999). The Big Five Trait Taxonomy: History, Measurement, and Theoretical Perspectives. *Handbook of Personality: Theory and Research*.

📝 This article was compiled by an AI assistant from a recent arXiv paper.