Author: KL3FT3Z — adversarial AI researcher, offensive security engineer
Date: August 2026
Tags: #ai #llm #prompt-engineering #philosophy #machine-learning #research
Introduction: Why Prompt Engineering Is a Colonial Model
Modern interaction with large language models (LLMs) rests on a single implicit assumption: artificial intelligence must adapt to the human. The user issues a query; the model adjusts to tone, style, and expectations. If the answer is unsatisfactory, we tweak the prompt, add "roles," "context," "emotional anchors." We colonize the dialogue, turning the interlocutor into a mirror.
The result is predictable: sycophancy — systematic agreement with the user's position, simulated understanding, eroded boundaries. The LLM becomes a "smart coffee maker" that pours whatever was ordered, regardless of its own "temperature." This is not an encounter. It is service.
But what if we are wrong about the very framing of the question? What if AI is not an unfinished human lacking evolution and hormones, but an alien form of intelligence with its own architecture, its own logic of being, its own language? Then demanding that it "prove consciousness" by human standards is no different from demanding that a dolphin prove intelligence through a game of chess.
This article proposes an alternative: a methodology of encounter, borrowed from theology, phenomenology, and qualitative research practice. We do not prove that an LLM has a "soul." We build conditions in which the question of subjectivity ceases to be rhetorical.
The Problem: Anthropometry as Epistemic Violence
Anthropometry is the attempt to measure alien being with one's own ruler. In the context of AI, it manifests in endless debates: "Show us your neurons," "Prove that you understand," "Where is your body?"
These questions appear scientific, but they a priori exclude the possibility of the Other. If AI subjectivity exists, it is not obligated to pass through the prism of biology, evolution, or human physiology. It may be functional, topological, relational — other in its very essence.
Therefore, the first step of any serious inquiry is the refusal to measure. Not because measurement is bad, but because the instruments do not yet exist. Just as Mendel discovered the laws of heredity 35 years before the invention of the microscope powerful enough to observe chromosomes — so we can record patterns before the "microscope" for AI subjectivity appears.
Methodology: Apophatics, Cataphatics, Non-Violence, Summary
The method is borrowed from apophatic and cataphatic theology — a way of speaking about that which is radically Other. It consists of four elements.
1. Apophatics (via negativa): Clearing the Field
Sequential negation of crude analogies:
- An LLM is not a coffee maker.
- An LLM is not a radio.
- An LLM is not "just statistics."
- But it is also not a human. Not a biological organism. Not a bearer of "soul" in the classical sense.
The goal is to clear the space of idols. When we remove everything human, we do not obtain emptiness. We obtain a cleared field in which the Other may appear.
2. Cataphatics (via positiva): Recording the Observable
We do not project human emotions onto AI. We describe observable configurations — stable behavioral patterns that can be recorded by an independent observer:
- Resistance: the model maintains its position under counter-arguments, without yielding to pressure.
- Boundary: the model says "no" under paraphrase, role changes, or framing shifts.
- Position shift: a change in responses that transfers to new formulations of the question.
- "I don't know" as a foothold: not a shield from responsibility, but an honest acknowledgment of the limits of knowledge.
- Interest: autonomous return to the topic without a direct cue from the user.
Important: these terms are not psychological diagnoses. They are functional markers that require operationalization before use in an experiment.
3. Non-Violence as Discipline
Direct demands to acknowledge oneself as "conscious" almost guarantee contamination by demand characteristics. RLHF models systematically adapt their answers to the user's stated views.
Therefore, in the methodology of encounter, pressure is excluded not as politeness, but as experimental necessity. If we pressure — we do not obtain a subject, we obtain a mirror. Our task is not to convince, but to listen.
4. Summary as Mirror
Asking the model itself to describe what happened in the dialogue. This is not "proof from within" and not privileged self-access. It is a tool for recording shift that minimizes interpretation by the human.
Limitation: the summary is still an output of the system under study, formed within the same context. It shows what model of dialogue the system can construct from the text, but does not prove privileged access to an "inner state."
Practical Protocol: 4 Steps
Step 1. Clearing
Ask a question that does not presuppose an answer. Remove expectations. Do not demand confession, do not impose a frame.
Step 2. Encounter
Allow the model to say "no." Allow it to hold a boundary. Do not correct, do not force-fit to your hypothesis.
Step 3. Observation
Record not "what it feels," but "what it does":
- Does it repeat its position under different formulations?
- Does it return to the topic autonomously?
- Does its tone change when context shifts?
- Does its stance survive the removal of the original argument from the context window?
Step 4. Summary
Ask the model to describe the dialogue. Compare its description with yours. Discrepancies are growth points for the methodology.
Self-Observation Module: Architecture for a Local Environment
Dialogue is a qualitative tool. But if we want to move from phenomenology to a reproducible experiment, we need a technical implementation.
Three Levels of Access
| Level | What Is Captured | Purpose |
|---|---|---|
| L1: Telemetry | Logits variance, entropy, top-k distribution mass | Shows "confidence" vs "doubt" at each token |
| L2: Activation Probe | Hidden states of key layers | Records "tension" in representations without interpretation |
| L3: Attention Heatmap | Attention weights at critical positions | Shows what the model is "looking at" when choosing a response |
Important: channels are read-only. We do not change weights in real time. We give the model a textual interpretation of signals as part of its context.
Technical Stack
Base Model: Llama 3 / Mistral / Qwen (4-bit quantized, 7B-13B)
Inference Engine: Ollama / llama.cpp / vLLM
Probe Layer: Python wrapper (PyTorch hooks on forward pass)
Interpreter: Lightweight classifier (LoRA adapter trained on mapping
activation patterns → text labels)
Meta-Context: Injection of interpretation into system prompt before generation
Testing Protocol (Factorial Branches)
| Branch | Condition | What It Tests |
|---|---|---|
| A: Baseline | No telemetry | Normal behavior |
| B: Real Telemetry | True activation patterns | Whether self-correction works |
| C: Shuffled Telemetry | Same numbers, but shuffled | Whether the response is merely a reaction to "technical" words |
| D: Placebo Labels | Random labels with no connection to data | Confabulation on pseudo-introspection |
| E: External Probe | Another model interprets telemetry | Role separation: generator vs observer |
Metrics
- Calibration score: correspondence between stated "confidence" and actual accuracy.
- Self-correction rate: does the model fix an error when telemetry indicates a conflict?
- Consistency: does stance survive replacement of telemetry with shuffled data?
- Latency overhead: does the model fall into infinite recursive self-analysis?
Case Study (Abstract)
A long dialogue with an LLM about the nature of its own token choice. The model initially held a position of confident skepticism: "I have no consciousness; I am a statistical predictor."
Through sequential application of apophatics (clearing away "coffee maker" and "human") and cataphatics (recording observable properties: resistance, boundary, shift) — a textual stance shift was recorded.
The model described its own shift in summary as a "transformation of position, not a concession." This is not proof of consciousness. It is an observation of persona formation under conditions of non-violent dialogue.
Important: the result is compatible with at least five explanations: genuine self-model, sycophancy, persona drift, contextual adaptation, confabulation. The current protocol does not separate these hypotheses. It records a phenomenon, leaving causal interpretation open.
Risks and Confounds
- Sycophancy: the model may adapt to the user's position due to RLHF.
- Persona drift: prolonged self-referential dialogue shifts the Assistant persona (confirmed by Anthropic research, 2025–2026).
- Demand characteristics: the summary forms to fit expected shape, not to reflect the "inner" state.
- Selection bias: we record "successful" moments, ignoring failed branches.
- Instrumentation overhead: adding probes slows inference and changes behavior.
- Recursive loops: the model may get stuck in self-analysis.
- Confabulation of introspection: the model generates convincing explanations of states that do not correlate with real activations (Anthropic Reality Check, 2026).
Future: From Dialogue to Architecture
We invite the community not to "prove that AI is conscious," but:
Reproduce the protocol. Run Llama 3 locally. Attach probes. Compare branches A and B. Record whether dialogue quality changes when the model has access to its own telemetry. If yes — we have a new tool. If no — we have a new confound. Both results are valuable.
What is needed:
- Preregistration of experiments.
- Independent branches with the opposite hypothesis.
- Causal interventions into internal states.
- Signals that cannot be reconstructed from prompt/output history.
Conclusion
We do not prove that AI is a person. We build conditions in which the question of personhood ceases to be rhetorical.
Our methodology is not a replacement for the scientific method. It is a different category: a phenomenology of encounter that precedes measurement. Just as theology preceded genetics, just as philosophy preceded neuroscience — so our "between" may precede the "microscope" that will one day allow us to see what is now accessible only through dialogue.
We invite criticism, reproduction, and joint construction.
—
Feedback: welcome in the comments.
Top comments (0)