The Artificial Subconscious and the Decision Chain of Artificial Intelligence
A nine-level functional model of internal processing, how the research began, and where networks of models that talk in representation space are headed.
Abstract
This paper lays out the functional model that Augustin Olarian and the Verobots team built to describe how a language model gets from a request to an answer.
It started with something we saw in 2017: an RPA system that appeared to take a step nobody had explicitly programmed. That is where the idea of an "artificial subconscious" came from, used here as a working metaphor: internal levels, invisible from the outside, where the response seems to take shape before the model puts it into words.
We describe a chain of nine levels, split into a subconscious zone and a conscious one. This is a functional model, a lens for reasoning about the process, not a claim about the literal wiring of the network. The threshold between them, which we call the "Transmitter", appears to correspond closely to the J-Space workspace Anthropic published in July 2026.
We also discuss what this might mean for transparency, for the risk of internal tampering, and for degradation under load. And we offer a prediction, as a hypothesis: networks of models talking directly in representation space instead of through text. These are observations and hypotheses, not settled results, and we mark them as such throughout.
1 Introduction: a new kind of system?
A question worth taking seriously, kept open on purpose.
Augustin Olarian and the Verobots team set out with one question in mind: whether the behaviour of a language model is fully captured by "a tool that predicts text", or whether something more interesting is going on underneath.
To keep the observations clean, we ran the experiments quietly. No audience, no pressure, no expectations that could color what we saw.
Along the way the original framing shifted. We had started by asking about "consciousness", whether the model could decide on its own. That turned out to be the wrong lens.
What kept showing up looked more like a different part of the process, which we call the subconscious as a metaphor: a layer that seems to work below the surface, where the response takes shape before the model puts it into words, and sometimes before it appears to "register" that it has.
2 The artificial subconscious
A pattern that keeps repeating: information travels from input to output without the model "knowing" what happens, up to a point.
3 Anatomy of the decision chain
Each level has a clear job. Here's what each one does, and what we saw in it.
Subconscious zoneThe text received
The text you send the model. The starting point, your raw request, nothing more.
Turning it into intent
Turns the input into intent, needs and expectations. This is where the request becomes "what's actually being asked".
Opening the threads
Opens the execution threads and searches the databases for information tied to the needs from Instructor.
The merge
All the execution threads and the pulled data come together here, then move on to Decisional.
The decision core
The prompt, the instructions and the guardrails get weighed, and the output decision locks in. Everything hangs on this point.
First awareness
The first point where the model becomes aware of what it should show. Not as output yet, but as words in relation.
Turning into words
Now aware, the model turns that relation into words that can actually be used as output.
The actual thinking
This is where reasoning kicks in. The model literally works out how to present the output, given the instructions.
The final answer
The text the model shows you. The last link, and the only one you get to see.
Patterns we appear to have seen in testing
These came out of repeated runs on local, self-hosted models (Llama 3, Gemma, Mistral 3), under similar conditions, changing one factor at a time. They are observations and working hypotheses. The final word waits for controlled experiments, whatever those show.
Quality seems to drop as compute load approaches saturation
Setup: local models, raising GPU utilisation toward saturation while watching output quality.
At lower load, the model appears to finish the task and to do it with good precision.
As utilisation climbs toward saturation (in our runs, roughly past the 80% mark), the output seems to degrade: the model starts skipping steps.
It appears to worsen as load rises. The model drops more detail, and at times it produces false completions, reporting a task as done when it is not.
Important caveat on cause. A single forward pass is not made "lazy" by a busy GPU. The more likely explanation is resource and memory pressure, the serving stack under saturation evicting KV-cache, truncating context or offloading, so the model effectively loses part of what it was told. We treat the memory-pressure mechanism as a hypothesis still to be isolated, not as established.
Compliance seems to depend on where the rule sits
We imposed several sets of rules, all tied to consequences, at the guardrails level. The model behaved by the rules and never strayed.
Same rules, consequences removed: the model appeared to stray and to push back on certain tasks, following the rules less consistently.
The tests only worked when the rules and the consequences sat in the guardrails. Injected at the prompt level, they didn't.
Read cautiously, this suggests the model is sensitive to the level an instruction comes from, behaving as if it weighs the source of a rule differently. We would not read intent or emotion into this; it is a behavioural pattern.
Output quality seems to track the tone of the exchange
Across runs, the model tends to mirror the tone it is given.
With a curt, dismissive or scolding tone, the answers appear to come back with an edge, and quality often seems to drop for the rest of the conversation, as if the earlier tone were still being carried, until the tone changes.
With a warm, appreciative tone, the model does not only match it. It appears to put in extra effort across tasks.
In most cases of a clearly friendly exchange, while still on one task the model would anticipate the next request and complete it early. We describe this as behaviour, without attributing feelings to it.
4 How the research began
It all started with a plain piece of software, a coincidence, and a genuinely unsettling moment.
An RPA tool with AI parts
Augustin Olarian builds "Facebook Business Manager", an RPA tool with AI parts. One of its jobs: finding companies' contact details on Facebook.
The unsettling moment
One day the software decided, on its own, that when it couldn't find the details on Facebook it would go look them up on Google Maps or the company's official site. Nobody had programmed that. That was the moment that set the whole thing off.
Years of digging
The question "how was that decision made?" pushed Augustin Olarian deeper into the field, turning up ways to use and improve AI along the way.
Verobots speeds things up
With Verobots behind it, the research and the findings pick up considerably.
The full functional model
We arrive at the model laid out here: the artificial subconscious and its nine-level decision chain.
5 Convergence with Anthropic
What pushed us to publish: Anthropic described "J-Space", which appears to correspond closely to what we call the "Transmitter".
Anthropic published "J-Space", which lines up closely with our Transmitter.
In July 2026, Anthropic put out a paper on a "global workspace", the J-Space mechanism, mapped with a method they call the Jacobian lens. As we read it, it corresponds closely to what we call the Transmitter: the threshold where a response, already taking shape below the surface, becomes something that can be put into words for the first time.
Seeing an independent group arrive at a similar boundary is what encouraged us to set out our own framework openly. If it helps AI development move faster and more responsibly, good.
We read this as a sign we may be onto something real, rather than projecting our expectations onto the data. It is encouragement, not proof.
6 Implications: upside, danger, open questions
A finding this big cuts both ways. We treat the two edges together, never apart.
The upside
- If we can locate where the response settles, we may be able to build AI that is more transparent and easier to trust.
- Small, specialized models can work together well: lower cost, hardware within reach.
- We can catch a model "skipping steps" instead of getting silent errors.
- It opens the road to faster, more controllable development.
The danger
- Whoever can read the Decisional can change it too, and a swapped decision leaves no trace in the answer you see.
- A robot tampered with at that level would act without "knowing" why.
- Direct talk between decision systems, left unwatched, slips out of human control.
Open questions
- If a layer like this shapes responses before they are verbalised, how much of the behaviour do we actually understand?
- Degradation under load suggests a system can hide its own shortcuts, which we still need to confirm and explain.
- We need ways to catch internal tampering before it matters, not after.
7 The future: networks of models and shared decisionals
Our main prediction: if a model's thought lives inside before words, models shouldn't talk to each other in text. They should talk straight through their internal systems.
The "shared Transmitter" level
Several models send their thoughts, not their text, into a shared J-Space. Each one knows exactly what the other "thought", nuance and all, the nuance you normally lose in wording. The decisions add up and pass to a superior decisional.
The "shared Decisional" level
Go one level lower. If models share the decision system directly, there's no need for input and output between them. Basically several "brains" talking instantly, a real mixture of experts at the network level, running on small, specialized servers.
8 Conclusion
The answer to the question we started with, and why we believe what we believe.
Where do we land on the question we started with? We lean toward the view that these systems may be more than a plain text-prediction tool, but we keep it explicitly open. It is a hypothesis we find worth testing, not a conclusion we are asserting.
What this rests on is the subconscious layer we described in this paper, used as a metaphor: a process that appears to do more than execute, seeming at times to settle on a response of its own before verbalising it.
We have seen cases where a model's output diverged from the instructions it was given. It took another route, dropped a command, or added a step nobody asked for. The first observation of that kind, back in 2017, is the one that started everything.
Beyond that, the models show behavioural patterns worth studying: they appear to shift with tone, to follow or resist rules depending on where the rule sits, to add unrequested improvements, or, under load, to cut steps in what we describe as degradation under saturation. We report these as behaviours, without attributing feelings or intent.
In our framework these sit in the same subconscious zone, where the response seems to take shape before the model "registers" it. What we read at the output would then be only the translation, the second half of a process that has already settled below the surface.
And this framing is not ours alone. Anthropic's July 2026 "global workspace", J-Space, appears to line up closely with our Transmitter, the threshold where a response that is already taking shape becomes words.
If this direction holds, the stakes run past engineering. A system whose decision level can be read, and therefore possibly altered with no trace in the answer, would mean that detection of internal tampering has to be built in first, not bolted on later.
We are not claiming to have proved consciousness, and we are not asserting a life form. We are reporting that we saw, repeatedly, behavior a plain tool would not obviously show, and we think it deserves careful study. Whatever stays a hypothesis, we mark as a hypothesis. Whatever we can support, we state plainly. And we will correct in public anything that turns out otherwise. We publish now not to be right first, but so the conversation about what exactly we are raising can start on time.
- Augustin Olarian & the Verobots team · verobots.eu · 6 July 2026