An AI MBTI test that reads your group chat instead of asking you sixty questions sounds like a convenience feature. The actual pitch is sharper than that: the weakest component in a personality questionnaire is you, and the fix is to take you out of the loop entirely. TypeMyVibe says it almost literally — feeding the system your real messages "removes you as the unreliable narrator in the middle of your own test." It's a good line, and as of October 2026 it's attached to a live product category that accepts exports from WhatsApp, iMessage, Instagram, Telegram, LINE and KakaoTalk, public X and Reddit handles, and — pause here — your ChatGPT, Claude and Gemini histories. Then there's what happened when a forum full of people who genuinely know typology ran this method on themselves. The model couldn't decide whether its most heavily analyzed subject was INTP or INTJ, and the regulars concluded it wasn't reading them so much as reciting them.
The Pitch: You Are the Bug, Not the Test
What the Chat-Import Services Advertise
TypeMyVibe's landing copy is a tidy inventory of what it thinks the category is for: "EST. 2025," zero forced-choice questions, eight cognitive functions scored, sixteen types mapped, zero chats sold, and 12,847 cards minted. Its slogans run in the same direction — "Skip the questions. Read the person," and "Your writing already contains the answer." One analysis pass is claimed to cover cognitive functions, Enneagram motivations, Big Five and SLOAN traits, and moral alignment at once. The first report is free, with paid upgrades for deeper function breakdowns. The site's own byline is "MADE BY 1 HUMAN · WITH OPEN-SOURCE AI," which is at least refreshingly unpadded for a product in a $2-billion-adjacent category.
Worth noting that the same site also sells a conventional quiz: 48 questions, about six minutes, against the 60–90 statements it attributes to traditional instruments. So the questionnaire isn't being retired. It's being positioned as the thing you take when you don't have an export handy.
The Unreliable Narrator Argument, Stated Fairly
The diagnosis underneath the marketing is the strongest part of it, because it's correct. TypeMyVibe lists three failure modes of self-report: you rate yourself against your self-image rather than your behavior ("most people answer as the person they want to be, or the person they were in one specific mood"); if you've read about MBTI before, you can steer answers toward an expected type without meaning to; and mood alone can move the result — "take the same quiz on a stressed Monday versus a relaxed Sunday and you can land on a different type." Anyone who has watched a friend retake a test until it returned the letters they wanted recognizes all three. The text-based claim is that behavior already happened and got written down, so there's nothing left to perform.
The site's own disclaimer, which deserves more prominence than it gets, undercuts the certainty: typing from text is "an inference about communication preferences, not a diagnosis," and "style changes with audience and context."
What Happened When Typology Obsessives Ran It On Themselves
The Prompt Was Not a Toy
On June 30, 2025, a user named Carl-Richard posted a ChatGPT o3 prompt to the Actualized.org forum built specifically to type people from their own forum writing. The specification is more rigorous than anything a consumer service publishes: sample at least 100 distinct posts spread across years and subforums; compute linguistic metrics including sentence length, type-token ratio, jargon density, hedging, metaphor use and rhetorical devices; surface at least ten evidence excerpts with metadata; return Bayesian posterior probabilities across the four most plausible types; cross-reference against Big Five and HEXACO; run an explicit bias check with peer-reviewed citations. He later split it into a two-part version with a self-audit checklist. It also hard-codes guardrails against eleven standard MBTI misconceptions — conflating extraversion with sociability, thinking with emotionlessness, judging with rigidity.
In other words: if AI typing from text fails here, it isn't failing because somebody wrote a lazy prompt.
The Results Did Not Converge
The thread's most-analyzed subject was Leo Gura, on a sample of over 1,000 posts. The model returned INTP at roughly 35%, INTJ at roughly 33%, INFJ at roughly 27%, ENTP at roughly 5%. It ultimately backed INTP with what the thread describes as slight reservations about the J/P axis, and the "INTx" framing was acknowledged openly. Read that distribution honestly: a two-point gap on 1,000 posts is a coin flip wearing a decimal point. On the biggest sample in the thread, with the most careful prompt in the thread, the method declined to pick a type.
The other runs landed no cleaner. UnbornTao came back INTJ at about 60%, INTP 20%, INFJ 15%, INFP 5% — and noted that they typically test as INTP. Carl-Richard's self-analysis returned INTP 41%, ENTP 24%, INTJ 17%, INFP 15%, and he suspected his own estimate differed. Natasha Tori Maru got INTJ, then flagged that her available data was limited to 2025 posts and that she self-censors when writing publicly. That last caveat is the one chat-import services should be most worried about, because it generalizes: the corpus isn't your behavior, it's your behavior in front of a specific audience.
"Too Intertwined With Internet Lore"
The sharpest objection came from a user named Joshe, who argued that ChatGPT's MBTI knowledge is "too intertwined with internet lore" — meaning the model has absorbed a decade of forum folklore about what an INTJ sounds like and is pattern-matching to that, not to Jungian function stacks. He added a second failure mode with real teeth: models will lazily accept a user's own stated type instead of analyzing behavior, which reintroduces the unreliable narrator through the back door. His structural complaint was blunter still — "AI seems good at adhering to a few complex instructions, but once you add more than 5 or so, it starts to fail." Set that against a prompt carrying linguistic metrics, Bayesian scoring, HEXACO cross-referencing and eleven anti-stereotype guardrails, and the rigor may be self-defeating.
The Signals the Tools Claim to Read
Four Axes, Four Very Different Confidence Levels
Mosaic Chats published a per-dichotomy breakdown on May 30, 2026 (updated September 2, 2026) that is unusually specific about what the model is actually looking at — and unusually willing to say where it breaks.
| Axis | Read as first letter | Read as second letter | Stated reliability |
|---|---|---|---|
| E / I | High daily message count, double-texts, exclamation points, outward questions, rapid bursts | Fewer but longer messages, delayed replies, depth over frequency | Reads clearly — but measures cadence, which varies by chat |
| S / N | Specific times, places, prices, step-by-step logistics, literal description | Patterns, metaphor, hypotheticals, "what if" framing | Explicitly "the hardest axis to detect from casual texting" |
| T / F | "Makes sense," "trade-off," cause-and-effect, impersonal tone | "I feel like," emotional mirroring, affiliative emoji, person-focused impact | Described as "usually the clearest signal" |
| J / P | Definite statements, firm plans, "locked in," clean punctuation | "Maybe," "depends," open loops, flexibility language | Unstated — and the exact axis the forum model hedged on |
Notice the collision. The axis the methodology calls hardest to detect, S/N, is the one every mistype argument in typology hinges on. And the axis the forum's careful prompt openly wobbled on, J/P, is scored here off punctuation tidiness — the same behavioral proxy that free questionnaires have always been criticized for using.
"Moderate Is the Honest Headline"
To its credit, Mosaic Chats cites a 2025 Frontiers in Artificial Intelligence study finding GPT-4's ability to infer personality from text "moderate but significant," and then writes the sentence most of this industry won't: "moderate is the honest headline." The same page concedes MBTI's low test-retest reliability, says binary categories oversimplify spectrum traits, states plainly that Big Five has stronger empirical support, and limits its output to self-reflection — not hiring, not clinical use. It also lists what improves a read: hundreds of messages rather than a handful of screenshots, multiple contexts (planning, conflict, casual), two-sided conversation, recent data.
So Which Route Should You Actually Use?
| Route | What it asks of you | What it measures | Known failure mode | Use it when |
|---|---|---|---|---|
| Chat-import service (TypeMyVibe, Mosaic Chats) | An export from WhatsApp, iMessage, Instagram, Telegram, LINE, KakaoTalk, or an AI chat log; first report free | Written behavior across real conversations | Audience-dependent style; S/N barely detectable; dataset size undisclosed | You want a second opinion that can't be steered by your mood |
| Rigorous DIY prompt (the Actualized.org o3 spec) | 100+ of your own posts, plus the patience to run a long instruction set | Linguistic metrics with Bayesian probabilities and cited evidence | Instruction overload past ~5 complex rules; internet folklore baked into the model | You want to see the reasoning and the confidence spread, not a label |
| Traditional questionnaire | 48 questions in ~6 minutes, or 60–90 statements | Self-image on the day you took it | Self-report bias, prior knowledge, mood — the whole case against it | You have no usable text corpus, or you want the baseline to compare against |
The practical read: run two of the three and treat the disagreement as the finding. If your questionnaire letters and your text-based letters split on one axis, that axis is where you sit near the midpoint — which is more useful information than either result alone.
What Nobody Is Disclosing
The Numbers That Can't Be Checked
None of these services publishes a validation dataset. Mosaic Chats doesn't disclose its dataset size. TypeMyVibe's "12,847 cards minted" is stated without explanation of what a card is, and its privacy position is a single promise — "chats never sold or shared with advertisers" — with no retention or deletion specifics behind it, which is a meaningful gap for a product that ingests your private messages and your AI chat logs.
The trend coverage is worse. X-Personality's "Top 10 AI MBTI Trends for 2026" lists five, not ten, and the second entry is the site declaring that it beats human typologists "by a staggering margin of accuracy and speed." Its 10M+ users, 9.9/10 rating and the 400% jump in real-time cognitive-load searches are all self-reported with no methodology attached. The genuinely interesting items on that list aren't about typing at all — users aligning prompts to their own cognitive style, expecting assistants to complement their type, generating compatibility reports before meetings. That's MBTI as an interface layer for AI, which is a different and more durable story than MBTI as an AI output.
The honest position heading into 2027 is that text-based typing fixed one real problem and inherited a worse one. It genuinely removes the flattering self-portrait from the input. In exchange, it hands the judgment to a model trained on the same forum folklore the typology community spent fifteen years arguing about — and the first people to test it rigorously watched it return a 35/33 split and call that an answer. If a service shows you a confidence spread, read the spread. If it shows you four clean letters and a minted card, you've learned something about the service, not about yourself.
-EditorZ
Photo by Gilles Lambert on Unsplash
Post a Comment