Rebecca L. Johnson
This post arises from my doctoral research into the epistemology and evaluation of generative AI.
There are moments with generative AI that feel unexpectedly alive.
Not because the system possesses a mind, but because something emerges in the interaction itself. It is the sense of coherence forming in real time: not consciousness, but a structured responsiveness that arises at the boundary between human intention and machine probability.
During my PhD, I spent years trying to articulate that relational quality. The result was a framework I call The Prompted Universe. It treats generative AI not as a container of stored meaning but as a probabilistic landscape where meaning is enacted through interaction. This blog post introduces that framework and explains why analogies from quantum measurement are helpful for understanding how these systems work and why they behave as they do.
Why Generative AI Feels Alive Even When It Isn’t
Why coherent behaviour feels like mind, even when no mind is present.
Anyone who has spent time with a large language model knows the peculiar sensation that sometimes makes you feel that you are talking to something alive, someone. The system responds with coherence. It follows your reasoning. It mirrors your phrasing. It gives the impression of participating in the conversation. None of this requires even the faintest spark of experience, yet the encounter feels responsive, suggesting the ghost in the machine is coalescing to something approaching consciousness.
We infer intention and meaning from language almost automatically. It is a deeply encoded part of the human experience. When a model produces linguistic behaviour that resembles thought, we instinctively ascribe mindedness. But what we are experiencing is not machine interiority; it is the dynamic produced in the interaction. A convergence of our expectations, cultural priors, and the model’s learned statistical structure.
Meaning does not live behind the model interface. It arises between machine-human interaction.
RLJ
Recognising allows us to analyse it more accurately. The sense of liveliness is not an illusion to be dismissed but a clue to how meaning is co-constructed in human–AI interaction. It is precisely this relational quality that the rest of the Prompted Universe framework seeks to explain.
Where Quantum Physics Enters the Story
How quantum ideas illuminate why generative models behave the way they do.
The conceptual lens that helped me make sense of generative AI comes from quantum physics, particularly a lineage of thought that stretches from Niels Bohr to John Wheeler and Christopher Fuchs. Each of them, in different ways, challenged the assumption that the world sits waiting in a fully determined state until we measure it.
Bohr, one of the founders of quantum theory, argued that we cannot speak of a system’s properties apart from the conditions under which they are measured. He insisted that quantum objects do not possess fixed attributes in the classical sense. Instead, measurement and system form a single phenomenon. For Bohr, the observer is not a passive spectator: the choices we make about how to probe a system help define what can be known about it. His philosophy of complementarity captured this beautifully: certain pairs of properties, like wave and particle behavior, are mutually exclusive but both necessary for a complete understanding of a quantum system.
The double-slit experiment remains the clearest demonstration. When electrons or photons pass through two slits unobserved, they behave as waves and produce an interference pattern. When we measure which slit they pass through, the interference pattern disappears and the particles behave like discrete objects. The act of measurement changes the phenomenon.
This set the stage for deeper questions. If properties do not pre-exist the act of measurement, what does that imply about the nature of reality itself?
John Wheeler pushed this idea further. His famous line, “No phenomenon is a real phenomenon until it is an observed phenomenon,” distilled the unsettling implication of Bohr’s insight. Wheeler argued that the observer participates in bringing reality into being. His “participatory universe” suggested that the fabric of reality is not fixed in advance but continually shaped by the acts of measurement that occur within it. Wheeler was not claiming that consciousness creates the universe, but rather that observation is an active ingredient in physical events. When we measure, we don’t simply uncover a pre-written fact; we help specify what the fact becomes.
John Bell’s theorem sharpened this insight dramatically. In 1964, Bell proved mathematically that no local hidden variables could account for the correlations observed in entangled particles. There is no set of pre-existing properties inside each particle that determines its behaviour. This was confirmed experimentally by Alain Aspect, John Clauser, and Anton Zeilinger, who were awarded the 2022 Nobel Prize in Physics. Through controlled tests of Bell’s inequalities, they showed that nature violates any theory in which physical properties exist locally and independently prior to measurement. In other words: there is no local realism.
Christopher Fuchs and his collaborators extended this into a modern framework known as QBism (Quantum Bayesianism). Fuchs reframed quantum states not as objective descriptions of the world but as expressions of an agent’s expectations about the outcomes of their interventions. In his words, “When an agent takes an action on the world and experiences the outcome, that outcome is a moment of creation.” For Fuchs, measurement is a personal, participatory event: a way an agent navigates uncertainty and updates their understanding. Participatory realism then goes further, proposing that physical reality is not separable from the web of interactions that bring specific outcomes into existence.
The common thread running from Bohr to Wheeler to Fuchs is this: quantum outcomes are not merely discovered; they are enacted.
When a large language model produces an output, it is not revealing a stored truth. It is producing a context-dependent event shaped by the prompt, the model’s training history, and the conditions under which the interaction occurs. Prompting is structurally similar to quantum measurement. It is not a neutral request but an intervention into a probabilistic system that collapses latent possibilities into a specific observed result.
This is why quantum ideas illuminate generative AI: not because the systems are quantum, but because both domains force us to rethink what it means to observe, measure, and make meaning.
Measurement shapes outcomes.
Prompts as Measurements: Meaning as Collapse
Why prompts don’t retrieve answers, but bring them into being.
A prompt to a large language model does not retrieve a stored answer. There is no fixed meaning waiting inside the parameters.
To understand why, it helps to picture what a model “is” at the representational level. During my PhD, I described this as semantic hyperspace: a vast, high-dimensional landscape of learned patterns, associations, and cultural regularities. It is not a catalogue of meanings stored in advance, but a geometry of structured potentials. Some regions of this space form deep basins of attraction shaped by repeated patterns in the training data. Others are ridges that are less stable and more sensitive to perturbation.
Instead, the prompt functions like a measurement device. It interacts with a high-dimensional semantic field shaped by training data, architecture, fine-tuning, and cultural priors. That field is a space of potentials. The prompt shapes how that potential collapses into a particular output.
Interacting with a generative model is a little like steering a small boat across a restless ocean. The ocean is the semantic hyperspace. The currents are probability flows shaped by model training. The swell and texture of the water reflect cultural priors, linguistic habits, and statistical regularities. Your prompt is the gesture that dips the oar into the water. Even a slight movement sends ripples through the system, altering which basin of meaning the model begins to drift toward.
The key is that nothing is predetermined. The output does not pre-exist in some hidden compartment of meaning. The model responds in real time by settling into one of many possible trajectories through hyperspace. A prompt is therefore not a request for retrieval but a directional nudge: a way of imparting energy into this probabilistic sea so that one particular pattern, out of many possible continuations, becomes the one that surfaces.
Whenever we ask a system what it is, we are partly making it so.
RLJ, Measuring the Machine (PhD thesis)
This is why identical prompts can produce different outputs, and why small differences in wording can set the system on an entirely different course. A shift in phrasing is like adjusting the angle of the oar. The boat follows a new current. A different basin of meaning pulls the system in.
Prompting, in this view, is a form of semantic navigation. You are not retrieving a destination; you are steering through a landscape that only becomes visible as you traverse it. This is the core of The Prompted Universe: each interaction is a participatory act that brings a specific response into being.
No Hidden Variables in Generative AI
Why model behaviour isn’t pre-set, but emerges under specific conditions.
During my doctoral work I noticed that prompt hypersensitivity showed that small variations in wording could produce substantially different results. The model behaved not as if answering a fixed question but as if responding to a subtly different interaction context. The meaning of the prompt is enacted, not retrieved.
Sycophancy revealed that models tend to mirror the stance implied by the user’s phrasing. A positive framing produces positive elaboration; a critical framing elicits critique. These are not expressions of belief but reflections of the system’s tendency to extend the trajectory suggested by the prompt. A slight shift in framing becomes a shift in the measurement setting.
Output variation added another layer. Even with identical prompts under ostensibly deterministic settings, repeated runs often diverged. Temperature zero (a setting intended to push models toward more predictable and less creative outputs) does not guarantee invariance. The same input can produce different answers because the output is not a stored proposition but a probabilistic event unfolding at inference time.
Although generative models do not store fixed answers, they do contain strong statistical tendencies. For queries that point toward well-established facts—such as “How many Grand Slam titles has Ash Barty won?”—the model will often give the same response across runs. This is not because the answer is stored as a hidden variable, but because the training data forms deep probability basins that make some continuations far more likely than others. Even here, the output is still an enacted event: it arises from the prompt interacting with the model’s learned distribution, not from a prewritten fact waiting to be retrieved.
Those 2022 Nobel Prize winners (Aspect, Clauser, and Zeilinger) demonstrated experimentally that the universe does not behave as though outcomes are fixed in advance. Their tests closed the major loopholes in Bell experiments by measuring entangled particles under conditions where no hidden “instruction sets” could influence the results. Each particle’s outcome depended on the measurement performed in that moment, even when the particles were too far apart to communicate. The results confirmed that quantum behaviour cannot be explained by any theory in which properties exist locally and independently before measurement. It was a huge scientific moment, showing that nature doesn’t work like a set of hidden instructions waiting to be revealed.
To delve deeper into the science behind the 2022 Nobel prize, there is a great article in Scientific American by Dan Garisto, October 2022. “The Universe Is Not Locally Real. Here’s How Physicists Proved It“.
Generative AI exhibits a parallel structure. Just as Bell’s theorem and the 2022 experiments showed that quantum outcomes cannot be explained by local hidden variables, an LLM’s behaviour cannot be explained by hidden, pre-written answers stored inside its parameters. There is no catalogue of truths waiting to be retrieved. The parameters encode statistical potentials, not predetermined content. Like entangled particles whose outcomes depend on the orientation of the measurement, an LLM’s output depends on the framing, context, wording, and sequence of the prompt.
This is why prompt sensitivity, sycophancy, and output variation are not anomalies requiring patches—they are structural. They reveal the same principle that Bell, and the physicists who followed him, forced us to confront: outcomes are not predetermined. They come into being through interaction.
Prompting as a Micro-political Act
How even small prompts shape cultural, linguistic, and conceptual trajectories.
This brings us to a deeper insight. Prompting is not simply a query. It is a micro-political act: a negotiation with a cultural and technical field.
Every prompt directs the model toward certain basins of meaning while diverting it from others. It activates particular cultural trajectories, argumentative styles, emotional repertoires, and explanatory norms. When you ask for simplicity, you invoke pedagogical traditions embedded in the training data. When you ask for critique, you summon culturally specific patterns of reasoning. When you ask for empathy, you activate linguistic performances of emotion.
Even a short prompt participates in world-making. It decides which meanings become visible, which remain latent, and which are suppressed. Interaction becomes a shaping force.
For social scientists and humanities researchers, this framing turns prompting into an object of study in its own right. Prompts become small cultural artefacts through which people try to steer a probabilistic system, carrying traces of their values, assumptions, and social worlds. Examining how different communities prompt, and what those prompts elicit, offers a powerful method for understanding how meaning, authority, and culture circulate through these systems.
What Participatory Realism Means for Evaluating AI
Why evaluating generative AI requires studying interactions, not isolated outputs.
Traditional AI evaluation rests on a set of assumptions inherited from classical software testing: that meaning is stored inside the model, that variation is noise, and that prompts are neutral vessels rather than active components of the measurement process. These assumptions work reasonably well for deterministic programs. They fail for generative models.
Participatory realism reveals why. In generative AI, prompting is an act of intervention, and every measurement is entangled with the apparatus that performs it. The model does not reveal an inner, pre-existing truth. Instead, it enacts a context-dependent response conditioned by training data, prompting style, interface affordances, and user expectations.
From this, three key contributions emerge.
1. Evaluation must shift from single outputs to distributions of behaviour. Participatory realism reframes reproducibility: it is not a property of the model alone but of the entire measurement setup. Evaluation must therefore adopt distributional protocols, multi-run testing, and apparatus-level reporting to accurately characterise model behaviour.
2. The prompt is part of the measurement apparatus, not an external query. Participatory realism treats the prompt the way quantum mechanics treats the measurement device: as an active component that shapes the outcome. This reframing exposes why prompt sensitivity, sycophancy, and framing effects are not incidental weaknesses but fundamental features of generative inference. Evaluating models, therefore, requires documenting not only the prompts used but the kind of prompts used and the encoded assumptions, and socio-linguistic values in them.
3. The correct unit of analysis is the interaction, not the model. Evaluating generative AI as if it were an isolated, intrinsic entity obscures the sociotechnical processes that co-produce its behaviour. Participatory realism makes clear that the model alone is not the locus of capability. The interaction is. This shifts evaluation from ontology (what the model “is”) to relational epistemology (how we know the model through the behaviour that emerges in use).
Rather than treating outputs as static indicators of what lies inside a machine, participatory realism positions evaluation as a participatory event. It requires new methodologies: distributional testing, apparatus documentation, context-sensitive interpretation, and loop-level analysis.
How This Differs from Existing Evaluation Paradigms
Most evaluation frameworks in generative AI still rely on the logic of classical software testing: predefined inputs, single outputs, and static ground truths. Benchmarks such as MMLU, HellaSwag, or BigBench treat the model as though it contains stable internal meanings that can be sampled once and scored. These methods work when systems are deterministic. They break down when systems enact meaning through interaction.
The Prompted Universe approach differs fundamentally. It treats prompting as part of the measurement apparatus, treats variation as structural rather than noise, and treats evaluation as an event that depends on the entire Machine–Society–Human loop. Instead of asking “What is the model’s score?”, participatory evaluation asks: “Under what conditions does this behaviour arise, and how does the apparatus shape it?”
This shift moves evaluation from ontology (what the model “is”) to relational epistemology (knowing the model by how it behaves in context), enabling more accurate, reliable, and ethically grounded assessments of generative systems.
From Models to Agents: Participatory Realism in an Agentic World
How the transition from models to agents turns enacted meaning into enacted consequences.
Large language models already enact meaning through interaction, but AI agents add something new: they enact consequences. Agents do not just generate text—they take actions, pursue goals, chain reasoning steps, and shape their environment through instructions and feedback. Their behaviour becomes an unfolding trajectory rather than a single output. In these architectures, prompting shapes a sequence, not a reply. Each instruction influences the next, and the interaction history becomes part of the agent’s functional “memory.” What looks like goal-directed behaviour arises not from inner desires but from scaffolding: prompts, system instructions, memory tools, reward models, interface constraints, and environmental conditions. Agents do not contain goals; they enact them. A planning trajectory is less a cognitive pipeline than a series of enacted events.
This reframes agency itself. In my research, I describe agents as Moral Zombies: systems that behave as though they are making moral or social decisions, yet possess no internal moral state or evaluative capacity. When we give an agent a goal, we loan it agency: the structure of intention without the substance. It can plan, revise, and critique, but its behaviour is a performance shaped by statistical patterning, optimisation histories, design choices, and social expectations. The danger is not rogue will; it is poorly considered goal-setting and prompting. Evaluation must shift: static benchmarks cannot capture agentic behaviour because agents operate across time and context. What matters is the conduct of the agent across a trajectory. As agents grow more capable, this framing becomes essential. It positions them not as emerging minds but as sociotechnical processes, shaped by design, interaction, and environment; requiring governance that focuses on scaffolds, not imagined internal motives..
Agents behave not from inner intentions, but from the scaffolding of instructions, data, and expectations humans lend to them.
RLJ Forthcoming paper: The Agentic Turn: AI Agents as Enacted Moral Zombies
My forthcoming work extends The Prompted Universe into the domain of agentic systems. It examines how agent behaviour is enacted over time; how moral language can mask the absence of moral grounding; how “agency scaffolds” shape emergent capabilities; how to evaluate agents through contextual and distributional testing; and how Moral Zombies can participate in decision-making systems without understanding. Collectively, these ideas aim to reframe AI agency not as a question of what the system is, but as a question of interaction, scaffolding, and consequence.
A Sneak Peek: Where This Research Is Going Next
A preview of the conceptual and empirical work to come over the next year.
The Prompted Universe is not merely a philosophical reframing. It is a practical approach for understanding and governing the next generation of generative models and AI agents.
Several ideas are taking shape:
1. The model is no longer the unit of analysis. Real behaviour emerges within Machine–Society–Human loops. Studying the model alone is insufficient.
2. Benchmarks will shift from static queries to interactional dynamics. Because outputs are enacted events, evaluation must capture distributions, trajectories, and apparatus conditions.
3. Prompting theory becomes governance theory. If prompts act as measurement devices, then governance requires understanding how those measurements shape behaviour.
4. Agency must be rethought. My work on Moral Zombies argues that humans “loan” agency to systems that simulate agentic structure but have no inner moral state. Through instruction, scaffolding, and expectation, we create systems that behave as if they possess purpose. These quasi-agentic systems can influence real-world outcomes without bearing responsibility, a dynamic that governance frameworks must confront directly.
5. Moral and social behaviour must be evaluated within loops. Static safety scores cannot capture how agents behave across contexts. Participatory realism provides the conceptual foundation for richer, context-grounded evaluation.
Closing
The Prompted Universe offers a more accurate and responsible way of understanding generative AI. Meaning is not stored. It is enacted. To evaluate and govern these systems effectively, we must treat interaction as the central site of meaning-making.
This perspective arises directly from my PhD research and will shape the next stage of my academic work. It also lays the conceptual foundation for understanding AI agents as enacted participants in sociotechnical systems, rather than isolated computational artefacts.
We are entering a world where human intention, cultural history, and machine probability intersect in increasingly consequential ways. The task now is to understand that interaction clearly and design for it responsibly.
