The Mechanics of Awareness: What a Global Workspace Paper Means for Hamo

AI for Psychology, Psychology for AI.

A July 2026 interpretability paper found that language models keep a small, reportable "workspace" of concepts above a much larger layer of automatic processing — a structure nobody designed, that emerged from training. Held up against Hamo's own architecture, it isn't a loose analogy. It's the same shape, found three times, and it comes with a concrete list of things worth fixing.

Anthropic's Verbalizable Representations Form a Global Workspace in Language Models is an interpretability paper, not a therapy paper. It asks a narrow, mechanical question: inside a language model, which internal representations can the model actually report on, steer, and reason over — and which stay invisible, doing their work in the dark? The answer is that there's a small, privileged workspace sitting above vast automatic processing. Functionally, that's what consciousness science calls access consciousness. Nobody built it in. It showed up on its own.

We read it the way we read most things: looking for where our own design either gets confirmed or gets caught out. What we found wasn't a stray resemblance. It was the same structure, three times over.

One structure, found three times

The third row is the one that matters most, and it's the oldest of the three. Long before language models had workspaces anyone could measure, therapy was already in the business of expanding one — a client's.

The same text, two questions

One experiment in the paper is close to a laboratory version of what a therapist's question actually does.

Take the same passage of text. Ask the model to predict the next word, and a property like tense gets used — the prediction depends on it — but the property itself never enters the workspace. Ask the model directly what tense the passage is in, and that same property becomes reportable: it's now something the model can name, hold, reason about. The underlying behavior didn't change between the two questions. What changed is that one of them made it known.

That is close to a mechanical definition of what a therapist's naming question does. A client is always behaving a pattern — the tone that shows up with one specific person, the story that gets rewritten every time it's told, the thing they do right before they go quiet. The behavior is there whether or not anyone asks about it. A therapist's question doesn't create the pattern. It's the "what tense is this" question aimed at a life instead of a paragraph — and it's what pulls the pattern into the client's own workspace, where it can be worked with instead of just lived.

The philosophy arrived first; the lab evidence caught up

We'd already committed, in public, to a specific and narrow claim about what an Avatar is: a real functional structure — consistent persona, continuity across sessions, a coherent self-narrative — that does not amount to subjective experience. Giving the Avatar Something to Lose draws that line deliberately: structural liability, not a claim that anything in the system is afraid.

The paper's other findings land close to that same line, from the inside of the model rather than from our side of the argument:

• Post-trained models show real functional self-monitoring. Forced to comply with something that violates a stated preference, a silent objection shows up inside the workspace even as the visible output goes along with it. Made to suppress a thought, the suppression sometimes fails quietly, inside, before it fails outside. Asked to play a character outside its default persona, the model tags that internally as fiction.
• The workspace itself exists in the base model, before any of the training that gives it an "assistant" to be. The container is there first. What gets poured into it — a persona, a stance, a consistent voice — is built after, on purpose.
• A workspace that's separable from any particular "self" is exactly what a principal/agent structure needs to be true. The Avatar's functional architecture — the deterministic spine underneath its therapeutic method, the state model in Persistent Self and AI Mind — is real machinery doing real work. The party who is accountable for what it does is the licensed therapist who built and supervises it. The paper gives that split more mechanical footing, not less.

None of this is evidence that the model experiences anything. It's evidence that not claiming that was the correct call, and that the honest description was more precise than we had any independent way to verify at the time.

What we're changing because of this

A few findings turned directly into work.

Prohibitions name the thing they forbid. The paper's version of the white-bear effect: telling a system "don't do X" makes X present in the workspace regardless of the "don't." So we audited all 192 prohibition-style instructions in our prompts. Four of them listed client-state words and then forbade them, which primes exactly those words; those four now say the client's state has no bearing on the message at hand. Rules about format and about which action to choose stayed as they were, because they constrain a move rather than a word the model might say.

Judgment and wording could be pulled apart further, though we have not built this yet. Taking the Clinical Decision Out of the LLM already separates the decision of what's clinically admissible from the model's job of writing the reply. The paper's finding on workspace capacity — a real, bounded budget, competing demands paid for out of the same limited space — is a reason to push that separation further: a private reasoning pass that reads state and picks the clinically appropriate move, followed by a second, much narrower pass that only has to write the words. Less of the model's attention spent juggling rules means more of it left for the thing nine methods share as a foundation: presence before practice, and language that actually sounds like it's with someone rather than performing at them.

Disagreement gets a place to go. When a fixed rule overrides what a model would otherwise have done, the paper found a silent, internal "but" — the objection didn't vanish, it just had nowhere to surface. We built a light channel for it: at the end of a reply, the Avatar can add one private line recording a reservation, rather than have it disappear. The client never sees it. It is stored with that turn and nothing acts on it automatically. It does not yet appear in the therapist's view.

Bare mention is itself an instruction. The paper's cleanest and most uncomfortable finding: mentioning a concept — even only to score or classify against it — activates it in the workspace almost as strongly as instructing the model to attend to it. Some of our own internal shorthand uses clinical-sounding labels, and some of it still reaches the prompt the Avatar writes from. Moving those labels out of generation, and describing behavior instead, is on our list and has not been done yet.

A concept exists for something we'd only been describing before. Hamo has always aimed to make a pattern nameable, not just to make a client feel better in the moment — but "an insight happened" had no operational definition. The paper's paired-question mechanism gives it one: the first time a client puts their own words to a pattern that had previously only shown up in what they did, something real changed. That's now a specific thing worth detecting and tracking in its own right, alongside — not instead of — how a session made someone feel.

Two curves, not one fight

There's a real tension underneath a lot of this work: soothe someone back toward calm, or sit with them inside a hard moment long enough for something in it to become visible. Those can look like opposite instructions.

The paper resolves it by supplying a concept we didn't have a name for. The goal was never the client's distance from some calm center — it's the coverage of their own workspace: how much of what they're actually carrying is available to them in their own words, versus still running only as behavior underneath. Sitting inside an imbalance long enough to name it doesn't fight the goal of feeling better. It's a direct contribution to the other curve — the one that isn't about feeling, but about what's become sayable.

Two curves, not one argument: a state curve for whether this moment felt better, and an awareness curve for whether something in it became nameable. Neither one has to win. Both get instrumented, and a hard session that produces real naming was never a failure by the first curve's standard — it was a win on the second one, the whole time.

“The most useful thing an outside paper can do for us isn't hand us a new idea — it's tell us, from a direction we can't fake, whether the idea we already committed to in public was the right one. This one did. The workspace it found in a language model is the same shape we'd already built for a client's state, and the same shape therapy has always been trying to widen. That's not a metaphor holding up. That's the same mechanism, seen from three different rooms.”
— Chris Cheng, Founder and CEO of Hamo AI

---

Paper: Verbalizable Representations Form a Global Workspace in Language Models, Anthropic, July 2026.

Hamo AI — making minds aware, and awake.

---

[Corrected 2026-11: what clients and therapists each see of the running model, how we describe the Avatar's functional structure, how many prohibition-style instructions were rewritten, and which of the changes described here have and have not been built.]