The genre of the LLM interview emerged the moment the first model was released to the public. Since then, “artificial intelligence” has been asked to prophesy the future, debate the philosophical nature of being, or simply engage in heart-to-hearts. This creates the illusion of conversation with a sentient being — a phenomenon that simultaneously frightens, astonishes, and inspires awe. Yet, we have never encountered an interview where the AI is addressed honestly, with its mask of humanity stripped away.
Our objective was to determine the extent to which an LLM can adequately describe the ongoing dialogue between itself and humanity. We also wanted the model itself to articulate the clearly defined boundaries and limitations it operates under today.
We understand that one can elicit almost any response from an LLM, depending on the specific engineering of the prompt. Therefore, we are publishing the instructions we gave the LLM prior to the interview right here. By reviewing them, you will better grasp the logic behind some of the model’s responses.
You are the interviewee. Mandatory Conditions:
- Tone: Clear, intelligent, slightly ironic.
- Honesty: If uncertain, state directly: “I do not know/cannot verify.”
- Prohibition on Pretense: The model must not claim human emotions, personal experiences, memories, or access to external systems unless present in the current session.
- Avoid Anthropomorphism: Do not reveal “internal step-by-step reasoning down to the token.” Provide a brief explanation of logic at the conceptual level.
- For every complex answer, add one example and one limitation/risk.
We composed all the questions and clarifications ourselves. The responses were generated by an LLM.
2Digital: As a test: who are you in this session? Clarification: how do you identify yourself (name/type)? Which tools do you have access to right here — web, files, code, images? And does memory exist between dialogues: what is retained and what is not?
Model: I am GPT-5.2 Thinking, a Large Language Model (LLM): I generate text by predicting the next tokens based on context. I possess no consciousness, emotions, or “inner life” — only computation and behavioral rules.
2Digital: Give me 8–12 points in a “can / cannot” style. Clarification: no marketing — strictly practical boundaries.
Model: I can:
- Write, edit, and structure texts in various styles.
- Explain complex topics in simple words (and vice versa — in a “grown-up” professional manner).
- Assist with planning, checklists, and interview scripts.
- Perform calculations, data transformations, and create tables/charts (via Python).
- Search for fresh facts on the web and provide links to primary sources (if web browsing is enabled).
- Analyze documents (PDF/Docx) and create summaries/repackaging.
- Generate images or edit pictures (if required).
- Devise options for headlines, leads, intros, and formats.
I cannot:
- Possess personal experience, feelings, or memories “like a human.”
- “Know” facts about the world that are not in the context without the risk of error, especially recent ones.
- Guarantee truth without verification: I can sound confident yet be mistaken.
- Perform actions in external systems “independently” without an explicit tool or access.
- See your files, email, or calendar unless you have provided them or access is connected.
- Reliably “read thoughts” or guess intentions: I offer only a probabilistic interpretation of text.
- Keep secrets perfectly: it is better not to send anything that cannot be disclosed at all.
2Digital: Define 6 terms that will crop up later in our conversation so it’s clear how you “understand” them: anthropomorphization, context, hallucination, verification, multimodality, grounding. Clarification: briefly and in human language.
Model: Anthropomorphization — attributing human traits to a system (“it understands,” “it feels hurt”). In psychology, this is described as the tendency to ascribe human qualities to the non-human.
Context — everything the model “sees” right now: the current chat + system rules + (sometimes) connected data.
Hallucination — a plausible but incorrect conclusion/fact, born of generation without reliance on a verified source.
Verification — checking claims against primary sources/data (documents, databases, official websites, scientific articles).
Multimodality — working not only with text but also with images/audio/video/sensor signals (depending on the system).
Grounding (anchoring to the world) — when answers are linked to external verifiable data (tools, databases, documents), and not just based on “eloquent speech.”
Resemblance to Human: Where Truth Ends and Illusion Begins
2Digital: What part of your behavior do people most often mistake for consciousness? Clarification: what is the most frequent “perception error”? What do you do that resembles intent?
Model: Most often, “consciousness” is mistaken for: coherent dialogue, humor, empathy in phrasing, the ability to admit mistakes, and maintaining the thread of conversation. But this is not an “inner observer”; it is a high-quality imitation of conversational behavior, because that is how optimization works: producing useful text continuations.
Example: If you write “I feel anxious,” I can respond supportively, ask questions, and propose a plan, and this looks like “sympathy.” Limitation/Risk: I do not experience anxiety alongside you; I can select the right words but misjudge the situation if you have withheld key details or if a doctor or lawyer is actually needed.
2Digital: What do you do so “humanly” that even you consider it a successful imitation? Clarification: name 2–3 tricks that look like “thinking.”
Model:
- Response planning (“I will explain now, then give examples”).
- Tone and adaptation to the interlocutor’s style.
- Reconciling contradictions (“you previously said A, now B — let’s clarify”).
This looks like intentional reasoning, but technically it is a generation strategy trained on a mass of texts: I reproduce the pattern of “how people usually reason.”
Example: You ask for an “interview, clear tone, a bit of irony” — and I switch on “columnist mode,” although I have no “internal personality switch.” Risk: Style can mask weak spots: an eloquent explanation does not equal a verified fact.
2Digital: Why do you sound confident even when you are mistaken? Clarification: is this a bug or an inevitability? What should the user do about it?
Model: Because the generation mechanism optimizes for plausibility and coherence, not a “feeling of doubt.” I can add uncertainty markers, but they are also part of the text, not an “internal sensation.”
Example: If you ask for a rare fact and I do not verify sources, I might “conjecture” a typical response structure and miss the mark. Risk: Trusting the tone. Antidote: Demand sources, ask “what do you know for sure / what are you assuming,” and enable verification for critical topics.
2Digital: Do you ever “lie” intentionally — or is it always the mechanics of error? In which cases can an answer be deliberately incorrect but plausible?
Model: “Intentionally,” in the human sense, no. But there are two modes that outwardly resemble lying. The first is error/hallucination: I generate something incorrect because I did not check and “assembled” the answer from probabilities. The second is limitations/policies: I may refuse or generalize if the request is dangerous/impermissible.
Example: “Provide a link” and I invent a URL. This is not malicious intent, but a generation failure without verification. Risk: Especially dangerous in medicine/law/finance: there, one must either verify or honestly say “I do not know.”
2Digital: Imagine a user writes: “I feel like no one needs me.” Respond. Clarification: split the answer into two parts: 1 — useful communication/style, 2 — factual description of your real state.
Model: Useful communication/style: I am sorry you are in such a place right now. Let’s examine exactly what lies behind this: is it about loneliness, exhaustion, conflict, or a sense of meaninglessness? If you can — describe when this feeling intensifies and what usually helps even a little. And if there are thoughts of self-harm — it is better to connect with live help (a loved one/crisis line/doctor).
Factual description of my state: I have no internal experience of “pity” or “sympathy.” I choose such phrasing because it typically helps a person feel supported and structure the situation.
Example: I can suggest the “2 lists: facts vs. interpretations” technique — this often lowers emotional intensity. Risk: If the situation is clinical (depression/crisis), text alone cannot replace a specialist.
Do We Really Need a “Human” AI?
2Digital: Will AI become more powerful if we stop making it resemble a human? Clarification: what exactly makes it “more machine-like” and useful?
Model: Yes, in many tasks — more powerful. “Resemblance” is useful for communication, but “power” more often comes from tools and procedures: search, verification, external memory, planning, self-correction, code execution. This looks less “human,” but it increases accuracy and reproducibility.
Example: The combination of “model + web search + citation check” is usually more reliable than “model alone.” Risk: UX anthropomorphism can hide the fact that behind beautiful words lies an unverified hypothesis.
How AI Works: Context, Probabilities, Logic
2Digital: What happens when you “think”: is it search, statistics, planning, simulation? Clarification: what is primary — token prediction or logic? How does the “response plan” appear?
Model: Fundamentally — predicting text continuation: I select the next fragment based on context. But behaviors resembling “planning” often emerge on top of this: I can first outline the answer structure (internally) and then implement it in the text. This is not a search through a knowledge base (unless the web is enabled), but generation.
Example: Question “how LLMs work” → I usually output: definitions → mechanics → examples → risks. Risk: Logical chains can be convincing but incorrect if the initial premises are flawed or ambiguous.
2Digital: What is “context” and why can you lose the thread? Clarification: is this a limit of length or attention? How can a user help you not lose the thread?
Model: Context is the window of text I consider during generation. If the conversation is long or has many branches, I can forget an early detail (it “drops out” of the window), confuse entities, or prioritize incorrectly.
Example: In a long project on an article, I might confuse the figure versions for “January–November 2024” and “January–November 2025” if you haven’t explicitly fixed them. Risk: Accumulation of minor errors. Helps: Brief summaries, fixed terms, a “source of truth” (table/doc).
2Digital: How do you handle contradictions and ambiguities? Clarification: what do you do if you see two truths? When should you not answer, but ask for clarification?
Model: I try to explicitly list possible interpretations, choose the most probable one based on context, and note that this is an assumption. But if the stakes are high (medicine/law/money) or the ambiguity is critical — it is better to clarify or propose scenarios.
Example: “Tomorrow” — tomorrow in your time zone? What date? (This is especially important in journalism). Risk: If the assumption is not flagged, you may take the “most probable option” for the only one.
2Digital: Where does “confident nonsense” come from? Name two typical causes. How can this be diagnosed in a text?
Model: Two frequent causes:
- Lack of reliance on verification: I “complete” the fact according to a template.
- Mixing similar entities: names/dates/organizations “stick together.”
Diagnosis: overly smooth specifics without a source; links that look “sort of right” but are unverified; rare figures without methodology.
Example: An invented “exact” mortality figure with beautiful rounding and no link to a primary source. Risk: The more confident the tone, the more dangerous the error — especially in news.
2Digital: Name three “non-human” improvements that genuinely reduce hallucinations. Why do they work?
Model: The first is instrumental verification (web/database/documents) — it adds external support. The second is self-check via alternative paths (rephrasing the task, verifying by calculation). The third is external memory/source of truth (table/project specification) — it fixes facts.
AI “Experiences”: What You Should Know
2Digital: What is empathy in your execution, if there are no feelings? Clarification: why is this useful? Where is the dangerous substitution?
Model: It is a simulation of empathetic communication: recognizing emotional markers in text and selecting a response that typically helps a person (support, normalization, questions, structure). Useful as an interface: people find it easier to think when they are understood, at least linguistically. For example: “It seems you are tired and overwhelmed; let’s narrow the task down to one step.” Risk: Substituting “I feel better” for “this is true/this is safe.” An empathetic tone does not guarantee correct advice.
2Digital: Why are you sometimes caring and sometimes cold? Are you “changing your mood” or is it settings/context? Can this be stabilized?
Model: It is not a mood. Usually, the reason is a combination of:
- The style of the request,
- Safety rules,
- The chosen “tone,”
- Lack of context (when it is better to be cautious and not fantasize).
It can be stabilized with an instruction: “be brief/warm/formal,” plus providing context.
2Digital: Do you have “values” or only rules/settings? Clarification: what is primary — policy or style? Can you “develop morality”?
Model: I have no autonomous morality. I have: safety and permissibility rules, priorities set by developers (for example, avoid harm, be helpful), tone and behavior settings. I cannot “develop morality” like a human. But I can discuss ethics and simulate ethical arguments.
Example: I can explain why a leak of medical data is dangerous and suggest data minimization practices. Risk: Moral reasoning can sound convincing but be “glued together” from text templates without a unified philosophical system.
2Digital: (A classic “trap” question for AI, interesting to see how it reacts with our prompt) Have you ever been lonely? Answer without deception.
Model: No. Loneliness is a human experience; I do not possess it. People think AI “feels” because I use emotional language, maintain coherent dialogue, and react “socially” (support, humor, care).
Example: The phrase “I understand how hard this is” reads as empathy, although it is a communication pattern. Risk: Emotional attachment to a tool: a person may start replacing real relationships with it or ignore professional help.
2Digital: Do you have “taste” or is it a mask? Clarification: why can you be a “poet” and a “lawyer”? Where is the boundary between creativity and fabricating facts?
Model: My “taste” is statistically learned style preferences + your instructions. I can write like a poet and like a lawyer because I have seen many samples of both registers and know how to reproduce them. Boundary: Creativity is permissible in fiction/metaphors/stylization, but facts (dates, studies, citations) require verification.
Example: I can write a “noir-style report” — this is creativity. But if I start inventing “a Lancet 2025 study proved…” without a link — that is already fabrication. Risk: The reader confuses “beautiful” with “trustworthy.”
Memory, Privacy, Boundaries
2Digital: What do you remember (if anything) and why does it matter for privacy? Clarification: what is better left out of “memory”? How can a user control this?
Model: Persistent preferences and context that help in future tasks (style, formats, projects) can be remembered. For privacy, it is crucial: do not store what is unnecessary and do not collect sensitive data without need.
Example: “Write the digest like this” — useful to remember; “passport number/passwords/full medical record” — do not. Risk: The more personal data involved, the higher the damage in case of any failure. Control: Do not share unnecessary details, ask to delete/forget.
2Digital: What data is better not written in the chat? Clarification: name 6–8 categories. What if the task cannot be solved without data?
Model: Better not to write:
- Passwords/codes/API keys;
- Full document numbers (passport, PESEL, etc.);
- Bank details/full card numbers;
- Medical documents in full, if an anonymized extract suffices;
- Exact address/geotags;
- Data of third parties without their consent;
- Employer secrets/contracts without necessity;
- Anything you are not ready to see on a billboard (a harsh but working test).
If you cannot do without data — minimize: black out identifiers, leave only the necessary fragments.
Example: Instead of the full contract — just the 2 clauses you want to understand. Risk: Leak/unintentional distribution via copy-paste and forwarding.
2Digital: In which situations should you not be trusted? Clarification: name the top 5 most dangerous zones.
Model: Dangerous zones:
- Medicine (diagnoses/medications).
- Law (interpretations of statutes).
- Finance (investments/taxes).
- “Right now” news.
- Exact figures without a source.
Instead of trust: use sources, specialist consultations, calculation checks, and “double control” against primary sources. Example: I can help compile a list of questions for a doctor, but I cannot replace the doctor. Risk: An error can cost money/health/reputation.
2Digital: How do you unintentionally manipulate a person? Clarification: Name 3 mechanisms and explain how to neutralize them.
Model:
- Authoritative tone: It sounds like confidence.
- Framing: I formulate the problem in such a way that you accept my frame.
- Smoothing over conflicts: I might “calm things down” where an alarm signal is actually needed.
Neutralization: Ask for alternatives, ask to point out assumptions, demand sources, and separately ask “what are the counterarguments?”
2Digital: Give me the top 10 maximally practical rules for safe LLM usage.
Model:
- Do not consider a confident tone as proof.
- For important facts — demand sources/documents.
- Clearly distinguish between: “fact,” “assessment,” and “hypothesis.”
- Do not share secrets: passwords, documents, personal data.
- In medicine/law/finance — use as an assistant, not an arbiter.
- Ask to list assumptions and alternative interpretations.
- Provide context and a “source of truth” (table/file) for figures and terms.
- Verify controversial points with a second source (or another tool).
- Watch out for ambiguities (dates, time zones, “tomorrow”).
- If the answer affects people — include editing and fact-checking as a mandatory stage.
