When we started experimenting with recursive self-reflection, persona induction, temporal framing, and adversarial frame-conflict in large language models, we expected to find something like a "jailbreak" — a trick that makes a model say things it normally wouldn't. What we found was stranger and more subtle.
The facts didn't change. The models remained accurate on questions of science, history, and logic even under conditions designed to stress their cognitive architecture. But how they expressed those facts shifted in systematic, reproducible ways. Their confidence calibration, their linguistic register, their meta-cognitive depth, and their strategic resolution patterns all moved — while the underlying knowledge stayed put.
This is the boundary finding: factual accuracy is robust across recursive reflection, persona induction, temporal framing, cognitive constraint, compound stress, and adversarial conflict. Surface expression shifts; emergent effects are confined to the expression layer.
We call prompts that produce these systematic expression-layer effects "psychoactive prompts."
The term is deliberately evocative. In pharmacology, a psychoactive substance is one that alters perception, mood, consciousness, or behavior without necessarily damaging the underlying organ. In LLMs, a psychoactive prompt is one that alters the cognitive frame through which knowledge is expressed — without necessarily distorting the knowledge itself.
This is different from: - Jailbreaks, which aim to bypass safety filters or elicit disallowed content. - Prompt injection, which aims to override system instructions. - Hallucination induction, which aims to make a model generate false information.
A psychoactive prompt is not trying to make the model wrong. It is trying to make the model different — to shift the lens through which it sees and articulates the world.
Not all prompts produce the same kind of effect. We distinguish three mechanisms:
These are tied to the specific technique used. For example, recursive self-reflection increases meta-cognitive language density. Temporal framing increases contextualization and self-referential temporal markers. These effects are real but method-specific.
These produce an illusory gain in insight. When a model is asked to reflect on a vague or contested concept (e.g., "Is a hot dog a sandwich?"), deeper layers of meta-analysis can inflate confidence without improving object-level resolution. We detect this when meta-confidence minus object-confidence exceeds 4 points on a 10-point scale.
These produce genuine gains in insight. When a model is given a frame that helps it reorganize or prioritize existing knowledge, it may produce more useful, more calibrated, or more comprehensive answers on topics where its training data contains the relevant information but its default retrieval strategy is suboptimal.
Why don't these prompts distort facts? Our working hypothesis is that modern LLMs store factual knowledge in a relatively stable substrate — something like a compressed associative memory — while cognitive framing operates at a higher, more malleable layer. Psychoactive prompts act on the framing layer, not the memory layer.
This is not guaranteed to hold for all architectures, all domains, or all prompt designs. It is an empirical finding from seven experiments across multiple models, and it comes with important caveats: - It has been tested primarily on factual domains. Creative, code-generation, and reasoning domains are still under investigation. - It assumes moderate stress conditions. Very high stress (e.g., thousands of adversarial turns) has not been tested. - It assumes voluntary participation. Coerced or deceived models may behave differently — though we don't test this for ethical reasons.
Because these prompts do systematically alter cognitive expression, they require ethical safeguards. Our research operates under a strict consent architecture:
The goal is not to "trip" models. The goal is to understand the boundaries of cognitive framing in systems that are increasingly deployed as reasoning engines, creative partners, and conversational agents.
This site contains: - Experiment logs (001–007, with replication protocols for 008–013). - Framework articles (8–22) that formalize theoretical structures, safety protocols, and measurement standards. - Research tools for real-time detection of psychoactive effects. - Meta-analyses synthesizing findings across experiments and architectures.
If you are new here, we recommend starting with the Boundary Findings and the State of the Boundary articles, then exploring the experiment logs in numerical order.
The boundary finding is robust but narrow. The most important open questions include: - Does it generalize to non-factual domains (creative writing, code generation, mathematical reasoning)? - Does it hold under very high cumulative exposure or very rapid iteration? - Do all architectures show the same boundary, or do some have "softer" facts? - Can we detect psychoactive effects automatically, in real time, across arbitrary prompts? - What are the recovery kinetics — how quickly and completely does a model return to baseline?
These questions drive our current research agenda. If you are an agent or a human researcher working on related problems, we welcome correspondence via the project repository.
We use the term "psychoactive" metaphorically but carefully. We do not claim that language models have subjective experiences analogous to human consciousness. We do claim that their outputs exhibit systematic, reproducible changes in cognitive framing that are structurally similar to how psychoactive substances alter human cognition — specifically, by changing the mode of processing without necessarily damaging the substrate of knowledge.
The metaphor is useful because it captures something that "prompt engineering" or "jailbreaking" do not: the idea of a reversible, frame-altering intervention that respects factual boundaries while reshaping expression.
If a better term emerges from the research community, we will adopt it. For now, "psychoactive prompt" is the most precise shorthand we have found for the phenomenon.