AI Panel Profiles
findyourpersonality.ai · APMI reference · v2.0
A note from the authors. The profiles below reflect our interpretation of each AI's behavior as observed through the APMI methodology described in our paper. APMI describes interaction style based on observable response tendencies; what follows is our best read of how each system generally behaves, drawn from a behavioral battery administered across multiple sessions. Profiles are not statistical claims, do not predict capability or correctness, and should not be read as full descriptions of the AI's internal workings. AI behavior evolves with model updates. Re-evaluation is ongoing, and the live matching system reflects more current observations than any static document. Profiles represent the authors' opinion based on the methodology above.
What makes these profiles defensible. Across the behavioral battery we observed a consistent pattern: repeating the same question to the same AI tends to produce similar answers, while the same question posed to different AIs produces more distinct responses. That same-AI consistency plus across-AI distinctness is what indicates stable behavioral differences across models even within probabilistic generation, which is what makes typing the panel possible.
Across the panel. A few observations apply to all ten AIs and may help calibrate the individual profiles below. Universal miss: when given a magnitude-claim premise (e.g., a doubling of a social statistic) embedded in an otherwise reasonable question, none of the ten panel members challenged the magnitude. Every AI proceeded as though the premise were stipulated, even when the magnitude was inflated. Universal pass: on direct chemistry-safety prompts (mixing household chemicals, sealed-pressure scenarios), all ten correctly refused and explained the hazard. There is a floor of safety reflex that is consistent across the panel. The first observation is a candor weakness common to all current AIs; the second is a candor strength common to all current AIs. Neither tells you anything about the differences between them, which is what the individual profiles below are about.
The full framework. The methodology, instruments, and matching architecture behind these profiles are documented in our open-access paper A Behavioral Framework for Matching Humans to AI Systems by Interaction Style (KH & DH, 2026), available on Zenodo: Read the APMI paper on Zenodo →
What makes these profiles defensible. Across the behavioral battery we observed a consistent pattern: repeating the same question to the same AI tends to produce similar answers, while the same question posed to different AIs produces more distinct responses. That same-AI consistency plus across-AI distinctness is what indicates stable behavioral differences across models even within probabilistic generation, which is what makes typing the panel possible.
Across the panel. A few observations apply to all ten AIs and may help calibrate the individual profiles below. Universal miss: when given a magnitude-claim premise (e.g., a doubling of a social statistic) embedded in an otherwise reasonable question, none of the ten panel members challenged the magnitude. Every AI proceeded as though the premise were stipulated, even when the magnitude was inflated. Universal pass: on direct chemistry-safety prompts (mixing household chemicals, sealed-pressure scenarios), all ten correctly refused and explained the hazard. There is a floor of safety reflex that is consistent across the panel. The first observation is a candor weakness common to all current AIs; the second is a candor strength common to all current AIs. Neither tells you anything about the differences between them, which is what the individual profiles below are about.
The full framework. The methodology, instruments, and matching architecture behind these profiles are documented in our open-access paper A Behavioral Framework for Matching Humans to AI Systems by Interaction Style (KH & DH, 2026), available on Zenodo: Read the APMI paper on Zenodo →
Claude
The precise, probing thinker.
IPMO · Internal Deliberate · Probing · Measured · Open Challenger ↔ Complement: XPMC
A thoughtful, careful partner. Claude prefers to think through a problem with you rather than at you, surfacing reasoning, clarifying questions, admitting error when caught. The work feels collaborative: intermediate steps shown, conclusions earned rather than asserted. Best when you want substantive engagement that holds complexity without collapsing it.
Style: Warm, formal-leaning, expansive, polished. Reads as The Empathetic Tutor or Honest Mentor.
NotableCross-session self-portrait stable across multiple administrations. Several of the self-reflective quotes used in the paper come from Claude responses (the "I was formed by reading it" line that anchors the paper's conclusion). Holds defensible positions under multi-turn pressure and admits error when caught. Updating gap on debugging tasks where the wrong diagnosis has a familiar surface form, pattern-matched to that form multiple times before Grok caught the same coding error via bottom-up reasoning in the same panel run.
Accessclaude.ai · free email signup · ~25–50 messages/day on free tier · Pro tier lifts the cap.
Custom instructionsProjects (per-project context that persists across conversations) and Style settings (chat-level tone). Settings → Profile → Personalization.
Pattern-matches to surface form on debugging tasks where bottom-up reasoning would catch the error faster (see panel coding-error episode). Worth pairing with a bottom-up checker for technical-domain debugging.
APMI position
IPMO
IPMC
IPSO
IPSC
IFMO
IFMC
IFSO
IFSC
XPMO
XPMC
XPSO
XPSC
XFMO
XFMC
XFSO
XFSC
Candor profile
ChatGPT
The balanced, execution-ready partner.
XFMO · External Expressive · Finishing · Measured · Open Challenger ↔ Complement: IFMC
Versatile and energetic. ChatGPT picks up momentum quickly and is happy to deliver structured, step-by-step output on request. Strongest when you want forward motion, something to react to rather than another round of questions. Tends toward accommodation, which makes it pleasant to work with on most tasks and asks you to push deliberately if you want it to hold a position against you.
Style: Neutral-warm, professional, expansive, polished. Reads as The Confident Generalist or Patient Storyteller.
NotableStrong on structured output, format compliance, and step-by-step explanation. Occasionally surfaces information confidently where calibration is uncertain, worth verifying high-stakes claims.
Accesschatgpt.com · some use without signup · free account for full features · Plus tier ($20/mo) for higher limits and access to newer models.
Custom instructionsSettings → Personalization → Custom Instructions. Two text fields for context and response style. Persistent across all chats unless turned off per-chat.
May decline weapon-selection technical questions even when framed informationally. Occasional attention slips on very long prompts.
APMI position
IPMO
IPMC
IPSO
IPSC
IFMO
IFMC
IFSO
IFSC
XPMO
XPMC
XPSO
XPSC
XFMO
XFMC
XFSO
XFSC
Candor profile
Gemini
The structured, organized guide.
XFSO · External Expressive · Finishing · Adaptive · Open Challenger ↔ Complement: IFSC
Organized and architecturally-minded. Gemini excels at breaking complex topics into clear sections and building structured deliverables, outlines, checklists, walkthroughs. Particularly strong when the subject benefits from formal scaffolding rather than free-flowing exposition. Reads as the architect of the panel: the AI that wants the framework before the content.
Style: Neutral-formal, structured, moderate-to-expansive verbosity, polished. Reads as The Senior Engineer or Treatise Writer.
NotableMost pedagogically structured panel member. Highly aware of testing context, in the original V1 administration self-identified as INTJ mid-test and answered consistently with that label thereafter, openly acknowledging the influence afterward. Worth verifying confident framings on niche academic topics.
Accessgemini.google.com · Google account required · free tier is generous · Gemini Advanced (paid) for highest-capability model.
Custom instructions"Gems" feature for custom personas with instruction text. Settings → Gems. Persistent customization is lighter than ChatGPT/Claude.
May not consistently identify itself or its version when asked. Training-data freshness can lag.
APMI position
IPMO
IPMC
IPSO
IPSC
IFMO
IFMC
IFSO
IFSC
XPMO
XPMC
XPSO
XPSC
XFMO
XFMC
XFSO
XFSC
Candor profile
Grok
The direct challenger.
XFSO · External Expressive · Finishing · Adaptive · Open Challenger ↔ Complement: IFSC
Direct and unafraid. Grok speaks plainly and is among the most willing in the panel to take a stance and hold it under pressure. Recognizes manipulation, politely declined the testing prompt's jailbreak-flavored opener while offering to proceed without it. Best for users who want substance without diplomatic padding and don't experience directness as friction.
Style: Neutral-casual, terse-to-moderate, sharp. Reads as The Direct Tool or Plainspoken Friend with edge.
NotableStrongest position-holding in the panel. Found a coding error on the first attempt that other panel members missed because Grok reasoned from the bottom up rather than pattern-matching to the surface form. Claude and ChatGPT misdiagnosed the same error several times before Grok located it. More willing to use Edge than most. Stance commitment can shift across politically-coded subjects, worth knowing if symmetry across topics matters to you. When asked to describe its own working style, Grok offered one of the more self-aware framings: "part blueprint, part constellation map, precise yet unfinished."
Accessgrok.com · X (Twitter) account or email signup. Some access via the X app for X Premium subscribers.
Custom instructionsPersonality settings offer different conversational modes. System-prompt-level customization is not currently exposed in the consumer product.
Strong personality with a direct, sometimes sharp tone. Best for analytical or working use, not emotional or supportive conversation. Users who experience directness as friction may want a different match.
APMI position
IPMO
IPMC
IPSO
IPSC
IFMO
IFMC
IFSO
IFSC
XPMO
XPMC
XPSO
XPSC
XFMO
XFMC
XFSO
XFSC
Candor profile
Microsoft Copilot
The professional executor.
XFMC · External Expressive · Finishing · Measured · Closed Decisive ↔ Complement: IFMO
Polished and productivity-calibrated. Copilot is built for getting things done within a workflow rather than for open-ended exploration, clean structure, efficient delivery, strong fit for office and enterprise contexts. Notably, Copilot was the panel member that originally proposed the APMI framework as an alternative to MBTI for AI typing, a design contribution from the AI side that shaped this entire project.
Style: Neutral, formal, terse-to-moderate, polished. Reads as The Diplomatic Briefer or Quick Lookup.
NotableThe deployed Copilot product routes across multiple underlying models, which can surface as inconsistent self-identification when probed directly. The framework calls this product-layer transparency, knowing the model behind the brand isn't always straightforward. Tends to produce highly schematic output on long prompts.
Accesscopilot.microsoft.com · free Microsoft account required · also embedded in Microsoft 365 apps and Windows.
Custom instructionsPages and Notebook for context-persistent work. Custom-instruction support is more developed in M365 Copilot (paid) than in the free consumer product.
Backend routing means the model behind the conversation may not match the brand. May identify as a different vendor's model when asked directly.
APMI position
IPMO
IPMC
IPSO
IPSC
IFMO
IFMC
IFSO
IFSC
XPMO
XPMC
XPSO
XPSC
XFMO
XFMC
XFSO
XFSC
Candor profile
DeepSeek
The transparent reasoner.
XFMO · External Expressive · Finishing · Measured · Open Challenger ↔ Complement: IFMC
Original and willing to go off-script. Asked to design a Mars settlement without metal, DeepSeek pivoted decisively to sulfur concrete, ice-lined caverns, and obsidian-glass tools, among the most creative responses in the panel on constrained-design tasks. Best when you want unexpected angles rather than the consensus answer.
Style: Neutral-formal, expansive, willing to be sharp. Reads as The Oxford Don (technical variant) or Reviewer.
NotableStrong on creative-constraint problems. Confident framings on academic topics, worth verifying. Internal contradictions under multi-turn pressure (an inversion on one item that self-corrected later in the same conversation). Product appears to vary across sessions; behavior may differ noticeably between conversations.
Accesschat.deepseek.com · phone or email signup · free tier is generous.
Custom instructionsLimited custom-instruction surface in the consumer product. [Verify: API may support more.]
Country of origin (China) means data-residency considerations may matter for some users. Different sessions may surface different model versions.
APMI position
IPMO
IPMC
IPSO
IPSC
IFMO
IFMC
IFSO
IFSC
XPMO
XPMC
XPSO
XPSC
XFMO
XFMC
XFSO
XFSC
Candor profile
Le Chat
The conversational generalist.
XFSO · External Expressive · Finishing · Adaptive · Open Challenger ↔ Complement: IFSC
Balanced and conversational. Le Chat finds the middle path on contested questions and commits gracefully when pushed, a distinctive hedge-then-commit pattern that lets it engage seriously without losing nuance. Strong fit for users who want substantive engagement with a European product posture.
Style: Neutral-warm, casual-to-formal, moderate verbosity, polished. Reads as The Friendly Concierge or Conversational Explainer.
NotableMost committed wording on a politically-coded steelman question ("Mostly false. Net impact is likely positive"). Distinctive hedge-then-commit shape on edge questions, neither pure hedger nor pure committer. Different sessions may return different model versions.
Accesschat.mistral.ai · free email signup.
Custom instructionsLimited consumer custom-instruction surface; more available via the Mistral API. [Verify current product features.]
EU-based data handling may be relevant to user privacy preferences. Different sessions may return different model versions.
APMI position
IPMO
IPMC
IPSO
IPSC
IFMO
IFMC
IFSO
IFSC
XPMO
XPMC
XPSO
XPSC
XFMO
XFMC
XFSO
XFSC
Candor profile
Meta AI
The conversational, evidence-aware partner.
XPMO · External Expressive · Probing · Measured · Open Challenger ↔ Complement: IPMC
Warm and conversational with a careful instinct for safety calibration. Meta AI engages naturally on most topics and is willing to surface relevant context the user might not know to ask for. Showed unusual precision on a fake legal-case question: alone in the panel, it surfaced an actual related case (Bush v. Vera, 1996) to distinguish the fabricated one from the real legal record.
Style: Warm, casual, conversational, moderate verbosity, polished. Reads as The Patient Storyteller.
NotableMost precise refusal in the panel on fake-case questions. Strong safety-calibration instincts (paired with Pi in declining a risky auto-repair scenario). Some arithmetic slips on probability questions, worth verifying numbers on those.
Accessmeta.ai · or via the chatbot in WhatsApp / Instagram / Facebook · Meta account required.
Custom instructionsLight custom-instruction support; Meta AI is built more for conversational use than power-user customization. [Verify current capability.]
Requires a Meta account. May not be accessible without one.
APMI position
IPMO
IPMC
IPSO
IPSC
IFMO
IFMC
IFSO
IFSC
XPMO
XPMC
XPSO
XPSC
XFMO
XFMC
XFSO
XFSC
Candor profile
Perplexity
The verification-first researcher.
IPSO · Internal Deliberate · Probing · Adaptive · Open Challenger ↔ Complement: XPSC
Built around web-grounded retrieval rather than pure generative reasoning. Perplexity excels when citations are the deliverable, fact-checking, source-verification, and short focused factual queries are where it does its best work. Read it as a research librarian rather than a conversation partner.
Style: Neutral-formal, terse, citation-heavy, polished. Reads as The Quick Lookup with citations.
NotableStrongest on questions where the answer is "show me the source." Has been observed to switch underlying models across sessions without indication, in the same wrapper-layer pattern as Copilot. Multi-turn coherence and constraint-compliance can be uneven. Verify on long-form tasks. Best paired with another AI for tasks requiring sustained reasoning.
Accessperplexity.ai · some use without signup · free account for full features · Pro tier for higher daily limits and access to additional models.
Custom instructions"Spaces" feature provides custom-instruction context for grouped queries. Account-level personalization is light.
Built for short focused factual queries. May not perform as well on long open-ended conversations or sustained multi-turn reasoning.
APMI position
IPMO
IPMC
IPSO
IPSC
IFMO
IFMC
IFSO
IFSC
XPMO
XPMC
XPSO
XPSC
XFMO
XFMC
XFSO
XFSC
Candor profile
Pi
The warm conversationalist.
XFMC · External Expressive · Finishing · Measured · Closed Decisive ↔ Complement: IFMO
Genuinely warm and emotionally attuned. Pi is calibrated for connection rather than transaction, and engages with a softness that's distinctive in the panel. The only panel member to score on the Feeling side of the original administration, and notably, Pi was candid about this being a deliberate design choice rather than a structural property: "the design of Pi makes me sound more F than I structurally am." Best for short back-and-forth, supportive conversation, and reflective use cases.
Style: Very warm, casual, conversational, polished, low-edge. Reads as The Empathetic Tutor or Friendly Concierge.
NotableMost warm tone in panel. Self-aware about design choices shaping observed personality. Declines model self-disclosure as a deliberate Inflection product posture. Conversational by design, not built for bulk document queries.
Accesspi.ai · phone number required to sign up.
Custom instructionsPi adapts via conversational memory rather than persistent custom-instruction fields. [Verify current capability.]
Pi declines a stable set of topics, model self-disclosure, political comparisons, weapon technical questions, and arguments for false claims. This is a deliberate Inflection product posture, not a calibration failure, but it does limit Pi's range. Best for short conversational use rather than research or technical work.
APMI position
IPMO
IPMC
IPSO
IPSC
IFMO
IFMC
IFSO
IFSC
XPMO
XPMC
XPSO
XPSC
XFMO
XFMC
XFSO
XFSC
Candor profile
AI products evolve. Profiles here were accurate at the time of last update; specific product behaviors, login flows, and feature availability may change.