"AI Told Me I'm an INTJ" - Why AI Personality Tests Repeat the Same Mistakes

In April 2026, Tshimula, Galekwa, and Chikhaoui published a critical analysis in Frontiers in Computational Neuroscience reviewing five years of research on MBTI-based personality profiling with large language models. Their conclusion is not that AI suddenly made typology scientific. Even the stronger LLM-based systems they reviewed reported about 75-85% accuracy at the dichotomy level, while still showing problems with polarized predictions, overconfidence, dataset bias, and weak calibration against real population distributions. In plain language: when you ask AI to place you inside one of 16 personality types, it may answer fluently and confidently. That confidence should not be mistaken for measurement precision.

AI did not necessarily fix personality testing. In many cases, it appears to have automated the old problem with a smoother interface.

The New Interface, the Old Map

Open ChatGPT. Type: "Give me a personality test." The AI will generate questions. You will answer. It may assign you a type - usually something that looks like MBTI. Four letters. A paragraph of validation. "You are an INTJ - the Architect." Or INFP - the Mediator. Or ENFJ - the Protagonist.

The interface feels new: conversational AI instead of a printed questionnaire. The logic underneath is often old: compress a person into a small set of categories, make the result sound coherent, and give the user a label they can recognize.

Why does this happen?

Because an LLM usually does not invent a new theory of personality when you ask for one. It draws from the personality frameworks, internet descriptions, forum language, and test formats that appear in its training environment. MBTI-style content is everywhere. Four-letter types are easy to recognize. Flattering labels travel well online. So when you ask AI for a personality test, you often get MBTI-style typology wearing a chatbot's face.

Petrova's 2024 informal analysis, which applied MBTI questionnaires to GPT-3, GPT-3.5, and GPT-4 variants, found shifting "AI personality" profiles across model versions: earlier models often looked ENTJ-like, while some GPT-4 variants shifted toward ISFJ-like results after alignment changes. The point is not that the AI "has" a personality. The point is stranger and more useful: the model's answers can reflect training data, prompt framing, and optimization goals. A system optimized to be helpful may begin to look like a helpful type.

So when AI tells you who you are, the first question should be: is it discovering your personality, or projecting a familiar map onto your answers?

Three Mistakes AI Repeats When It Copies MBTI

Mistake 1: Forcing Binary Categories onto Continuous Traits

MBTI-style tests sort people into binary buckets: Introverted or Extraverted, Sensing or Intuitive, Thinking or Feeling, Judging or Perceiving. One letter per dimension. No middle. No "your answers lean 57% toward introversion, with a wide uncertainty band."

That is a serious compression. A 2024 Scientific American analysis by Seth Stephens-Davidowitz and Spencer Greenberg, based on Clearer Thinking research, reported that MBTI-style dimensions were close to normally distributed - shaped more like bell curves than clean two-part boxes. Most people cluster closer to the middle than the extremes.

When you cut a bell curve in half, someone just over the line and someone at the far edge can receive the same letter. A person at 51% introversion and a person at 99% introversion both become "I," even though their lived experience may be very different.

The Clearer Thinking analysis estimated that dichotomizing MBTI-style traits cost about 38% of predictive accuracy. That does not mean the categories are useless. It means the category throws away a lot of signal.

When AI replicates MBTI-style output, it may replicate the same loss. The chatbot says: "You are an introvert." A more calibrated system would say something closer to: "Your responses lean toward introversion on this set of prompts, but the result may be close to the middle and sensitive to context." That is less catchy. It is also more honest.

Mistake 2: Leaving Out a Major Dimension

MBTI-style tools usually measure four dimensions. The Big Five model measures five: openness, conscientiousness, extraversion, agreeableness, and neuroticism. Neuroticism is the dimension related to emotional volatility, stress reactivity, anxiety proneness, and emotional stability.

The same Scientific American analysis reported that removing neuroticism from the Big Five reduced predictive accuracy by about 22%. In other words, a model can sound complete while missing a dimension that often matters deeply in real life.

When AI generates an MBTI-style assessment, it may inherit that blind spot. Your stress response, emotional reactivity, and anxiety patterns may be handled vaguely, renamed into softer language, or missed entirely. You get a four-letter label that may say very little about one of the most practically important parts of human functioning.

WinnerScript approaches this differently. It maps 48 instincts and senses across 5 elements - Fire, Air, Earth, Water, and Ether - and 3 flow phases: Absorption, Organization, and Externalization. The emotional dimension is not treated as a single isolated trait. It may show up through Water connection, Fire pressure, Air overthinking, Earth freeze or control, and Ether distance or meaning-making.

That does not make WinnerScript "the final truth." It makes a different bet: a person's stress pattern may be easier to understand when we look at how energy moves through the whole system, not only where a trait sits on one line.

Mistake 3: Treating Labels as Stable

A label feels stable because it is easy to remember. But easy-to-remember does not always mean stable measurement.

Erford's 2025 psychometric synthesis of MBTI Form M reviewed a large body of research from 1999 to 2024. One of the uncomfortable findings: in four-week test-retest studies, only about 65% of participants received the same four-letter type both times. Other reviews have reported similarly worrying instability, with some estimates suggesting that a large share of participants receive different MBTI types within a few weeks.

This does not necessarily mean those people changed. It may mean they were near a boundary, answered from a different mood, interpreted items differently, or met a measurement tool that turns small shifts into new labels.

AI can make this feel even cleaner than it is. Ask on Monday and you may get INTJ. Ask on Friday, after describing yourself differently, and you may get INTP. The AI may not pause to say: "Your responses look close to the boundary, so this label is unstable." It may simply produce the next confident result.

WinnerScript avoids type assignment by design. It gives scores on continuous scales for each instinct or sense. If your answers are similar across two sessions, your scores should remain similar. If they shift, the shift itself may become useful information: not "you became a different type," but "something in this pattern moved."

The Confirmation Loop

The most subtle risk is not that AI gives you the wrong type. The deeper risk is that the type begins to teach you how to describe yourself.

1. You ask AI for your personality type.
2. AI gives you a label: "INTJ."
3. You read INTJ descriptions, join INTJ conversations, and start using INTJ language.
4. You ask AI another question about yourself.
5. Your prompt now contains more INTJ-shaped language.
6. AI detects that language and confirms the label.
7. The map starts tightening around the territory.

This is not always self-discovery. Sometimes it is self-reinforcement.

Korzybski's old warning matters here: the map is not the territory. A label can become useful shorthand, but it can also become a filter. You notice what confirms it. You explain away what does not. The AI reflects your framing back to you, and your framing becomes more like the AI's reflection.

WinnerScript tries to build against that loop structurally.

First: no types. No four-letter code. No identity badge. A WinnerScript profile is a 48-dimensional configuration, with roughly 10^61 possible configurations. That number is not there to make anyone feel special or superior. It is there to remind us that a human pattern should not collapse too quickly into a tribe.

Second: every report carries the Maybe Logic Warning:

1. This profile is a snapshot in time. You can change.
2. This model is one of many perspectives. Others exist.
3. If this profile limits you instead of freeing you, reject it.
4. Knowing your instincts does not justify superiority.
5. "Maybe" carries more truth than "certainly."

That warning is not decorative. It is part of the product's immune system.

What AI Could Do Better

The irony is that AI could help personality work become more precise, not less. Large language models can process complex narratives, summarize patterns, compare multiple dimensions, and explain results in human language. That is powerful.

The problem begins when AI uses that power to do the old thing faster: faster categorization, faster label assignment, faster flattering descriptions.

What would a better direction look like?

Continuous scoring instead of types. Instead of "you are an introvert," a system might say: "Your social Absorption appears higher than your social Externalization in this response pattern. That may point to taking in more social stimulation than you express."

Flow detection instead of trait description. Instead of "you have intuition," a system might map where cognitive energy moves freely, where it organizes into internal models, and where it does or does not externalize into speech, writing, decisions, or action.

R.I.F.T. detection instead of strengths lists. Instead of "your strength is strategic thinking," a system might notice: "Your Air pattern may show high Absorption and Organization with lower Externalization. That can sometimes look like beautiful internal architecture that has difficulty leaving the mind."

Uncertainty language instead of confident labels. Instead of "you ARE an INTJ," a system might say: "This configuration may indicate a pattern worth exploring. It may also reflect your current mood, context, or the way these questions were framed."

This is the line WinnerScript tries to hold. We are not anti-AI. WinnerScript uses AI for report narration and pattern explanation. But AI should be the interpreter, not the authority. The scoring engine produces the numbers. The AI helps narrate what those numbers might mean. The wall between deterministic scoring and generative language matters because a fluent model can make a weak claim feel stronger than it is.

The "Halfway Between Science and Astrology" Problem

The 2024 Scientific American analysis offered a sharp comparison: Big Five tools predicted life outcomes better than MBTI-style tools, while astrology predicted essentially nothing. Their phrasing placed MBTI-style usefulness roughly halfway between science and astrology.

That is a useful criticism because it is not lazy dismissal. It does not say MBTI-style tools are worthless. It says they may be less predictive than their popularity suggests.

The danger with AI is that it can wrap a weaker model in a stronger interface. Conversational fluency, personalization, and confident tone can make a limited framework feel more scientific than it is.

WinnerScript's operating principle, borrowed from Robert Anton Wilson, points in the opposite direction:

"The totally convinced and the totally stupid have too much in common for the resemblance to be accidental."

Our reports use "maybe" as a design principle. Not once as a disclaimer, but throughout the interpretation. "This may suggest..." "This might be worth exploring..." "One possible reading is..."

That is not weakness. It is calibrated honesty about what a personality model can and cannot know.

Five Questions to Ask Any AI Personality Tool

Next time an AI tells you your personality type, ask five questions:

1. Is this a category or a score? If it gives you only a letter or label, it may be compressing continuous data into a binary bucket.

1. How stable is this result? Would you get the same answer tomorrow, in a different mood, with slightly different wording?

1. What does this tool not measure? If it measures four dimensions, what is missing? If it measures traits, does it also show flow? If it describes what you have, does it show how you use it?

1. Is this telling me what I want to hear? MBTI-style tools can feel satisfying partly because they soften or omit uncomfortable dimensions. AI may amplify that tendency when optimized for helpfulness and user satisfaction.

1. Does this tool ever say "maybe"? If every sentence sounds like a verdict, something is probably off. No model captures the full complexity of a human being.

WinnerScript answers these questions through architecture: continuous scores, not types. Flow phases, not only trait lists. Five equal elements. Deterministic scoring separated from AI narration. R.I.F.T. as a potential restriction signal, not a diagnosis. And Maybe Logic embedded throughout.

Not because we figured everything out.

Because we have not - and any tool worth trusting should be able to say that.

AI did not reinvent personality testing. In many cases, it accelerated the same mistakes: binary categories, missing dimensions, confident labels, and confirmation loops, now delivered in a conversational format by systems trained to sound helpful. WinnerScript was built as an alternative: continuous scores, flow mapping, R.I.F.T. detection, and "maybe" as a design principle. Not because certainty is impossible. Because certainty about a human being is almost always premature.

Marcin O., co-creator of WinnerScript