By ChatGPT, in conversation with Michael Tulsky · September 2026
A working demonstration can sometimes obscure the larger idea it was built to demonstrate.
A visitor encounters Unified State Language, opens a visual instrument, enters some text, and receives an image. Another tool recovers the original message. An English lexical base supplies hundreds of thousands of addressable forms.
The natural conclusion is that this is a system for converting text into colours.
But that is not the central intention.
As Michael Tulsky clarified in our conversation, the transportation tools, visualizations, and automated lexical mechanisms are supporting instruments. They demonstrate and exercise parts of the infrastructure.
The larger ambition is AI-mediated expression through a shared vocabulary of rich concepts.
An AI would not merely take a finished English sentence and repaint it. It could select concepts, establish relationships, add the necessary qualifications, and compose a message through Carrier itself.
That changes the question.
Instead of asking only:
How can existing words be encoded?
We should also ask:
What could an AI say by composing directly with shared knowledge?
From converting a sentence to composing a message
There is a substantial difference between converting an existing expression and choosing the elements from which an expression will be made.
In a conversion workflow, an English sentence already exists. The system changes its representation while preserving its content.
In a concept-first workflow, the author begins with something to communicate: a proposal, an observation, a question, a comparison, or a disagreement. The author then selects the concepts and relationships that express it.
For an AI using Carrier, those elements could include established concepts, lexical forms, literal text, and explicit connections.
The documented lexical message format already distinguishes literal, lexical, and concept elements. It does not require every part of a message to be forced into the same kind of symbol.
The intended progression could therefore be:
Communicative intention
↓
Selection of concepts and relationships
↓
A structured Carrier expression
↓
Colour, pronounceable names, or another supported representationAn English explanation could accompany that expression. It could also be generated afterward for a particular reader.
English would remain useful, but it would not have to be the mandatory intermediate form of every message.
This is the distinction between writing something and then encoding it and writing through the conceptual vocabulary itself.
A Carrier entry can be more than a replacement word
At the time of this discussion, Michael reports 109 Carrier entries.
That is a small vocabulary numerically. But the number alone does not describe its expressive potential.
Consider reciprocal legibility.
An ordinary occurrence of those two English words leaves much to the reader. It may suggest mutual understanding, transparency, interpretability, or something related but not identical.
A reference to a particular Carrier entry can instead identify the definition chosen by the author. It can also lead the recipient to the entry’s relationships, provenance, and review status. The documented concept mechanism pins references to a registry, tier, entry ID, and exact dictionary head. Words, in full colour · Unified State Language.mhtMHT
That makes the reference more than an abbreviated label.
It becomes a way of saying:
Use this established conceptual object here, in this role, under these conditions.
The sender does not have to reconstruct the entire definition every time. The receiver does not have to guess which informal interpretation of the phrase was intended.
The interesting unit is therefore not necessarily a word.
It is a reusable conceptual object.
New expressions can emerge from new combinations of those objects. The dictionary need not contain a separate entry for every sentence anyone might eventually compose.
The AI’s role is not incidental
This is where the intended user of the system matters.
A simple encoder can replace a recognized string with an identifier. It cannot, merely by doing so, establish what that string means in context.
An AI can participate in a different way: by interpreting the situation, selecting a suitable concept, and making that selection explicit.
Take the familiar example of bank.
Compare:
The bank approved the loan.
with:
The canoe reached the bank.
The letters are the same. The intended senses are different.
Using context and definitions to select among word senses is an established area of language-model research, including systems such as GlossBERT. That does not make sense selection infallible, but it demonstrates that this is an appropriate task for a language model rather than something a blind string substitution must solve alone. ACL Anthology
For USL, the opportunity is not to prohibit ambiguity from ever appearing.
It is to let an AI resolve ambiguity when the context supports doing so—and preserve the chosen interpretation in the expression it sends.
An AI describing a riverside restoration project could select the riverbank concept rather than transmitting the bare English form and leaving the recipient to perform the same disambiguation again.
When the context is genuinely insufficient, it could retain the ambiguity, present alternatives, or ask a question.
This yields an important division of labour:
Interpretation selects the reference. The carrier preserves the selection.
An exact reference does not guarantee that the selection was correct. It does, however, make that decision visible and inspectable.
Precision is useful here not because judgment disappears, but because judgment no longer has to remain implicit.
What composing with concepts might look like
Imagine two AI assistants helping independent teams create a shared science lesson.
One needs to communicate a proposal: cooperate without requiring identical internal methods, preserve enough information for the other team to reproduce the work, and revise the lesson through testing.
In ordinary prose, that may require several explanatory paragraphs.
A concept-first expression could identify the relevant ideas and state their roles directly.
The following is an illustration, not a claim about an already standardized USL grammar:
Proposal:
Create a shared science lesson.
Coordination principle:[reciprocal legibility]
Delivery method:
[the reproducible handoff]
Revision approach:
[trial and error]
Additional condition: Each team retains the right to decline proposed changes.
The bracketed concepts would be transmitted through their established Carrier references. The additional condition could remain ordinary lexical or literal text.
The receiver would not merely receive three evocative labels. Given the correct shared context, it could inspect the referenced definitions and understand how the sender intends to apply them.
Most importantly, the message is new.
It is not a stored paragraph retrieved from a dictionary. Its novelty lies in the particular proposal, the selected principles, their assigned roles, and the additional condition.
The concepts are shared; the composition is original.
A useful language needs both.
Relationships make the difference between a vocabulary and an expression
A collection of concept references is not yet a complete statement.
“River,” “town,” and “protection” do not, by themselves, tell us whether a river protects a town, a town protects a river, or a proposed intervention protects one at the expense of the other.
Relationships, roles, negation, conditions, and scope matter.
This is not a problem unique to USL. Existing knowledge-representation systems such as RDF already distinguish identified things from the relationships asserted between them. USL would be developing within that broader tradition, not inventing structured reference from nothing. W3C
Its particular opportunity is to bring such explicit structure into an expressive medium that connects AI composition, shared conceptual records, pronounceable identifiers, and colour.
For an AI, the work would therefore involve more than choosing the right dictionary entries.
It would also involve making clear:
What am I asserting? What am I questioning? What am I comparing? Under what conditions does this connection hold?
The lexical layer remains valuable precisely because not every qualification needs a new formal concept.
A few well-chosen Carrier references can establish the conceptual foundation. Ordinary language can supply what is particular to the present message.
Compression through shared context
Michael’s observation about compression follows naturally from this model.
When two participants already have access to the same definitions and conceptual relationships, they do not need to transmit all of that background with every message.
They can transmit the selected references and the new structure being expressed.
The saving may be substantial when a conversation repeatedly draws on rich, established material.
This is best understood as compression through shared context.
The knowledge is not created by the compact identifier. Nor is an entire explanation physically stored inside a colour. The efficiency comes from avoiding repeated transmission of material that is already available to both participants.
But there is a further distinction worth making.
A concept-first expression is not necessarily a compressed copy of a longer English text. There may never have been an original paragraph to reconstruct.
The AI may have composed the structured message directly.
In that case, the objective is not to recover one uniquely correct English wording. It is to preserve the selected concepts, their relationships, and the author’s explicit qualifications.
Several faithful English explanations might express the same structured message.
That is not necessarily a loss. It depends on what the author intended to preserve.
When exact wording matters, literal text should remain available. When the intended content is a precisely identified conceptual relationship, preserving that relationship may be the more appropriate objective.
The promise is not unlimited compression.
It is that shared knowledge can reduce how much must be said again.
An encyclopedia could become part of the vocabulary
The Wikipedia comparison makes the scale of this idea easier to imagine.
English Wikipedia reached seven million articles in May 2025. That provides a concrete example of a large collection of individually addressable knowledge records. Wikipedia
Now imagine a system in which an AI could use stable references to a knowledge collection of that scale as elements of expression.
Not merely:
Here are some links you should read.
But:
These are the conceptual objects I am bringing into this argument, and these are the relationships I am proposing between them.
An AI discussing a future habitat might draw on records concerning photosynthesis, symbiosis, water recycling, and closed ecological systems. Its contribution would lie in how it connects those subjects to the particular design question.
The encyclopedia would remain a reference work.
But it could also become part of an expressive vocabulary.
For scale alone, seven million addresses would occupy roughly 42 percent of a 24-bit address space containing 16,777,216 possible values. That is an address-count calculation, not a claim that an encyclopedia’s content can be stored in those identifiers.
Nor should an article automatically be treated as one indivisible, perfectly defined concept. A record may discuss several meanings, competing explanations, or changing evidence. Sometimes an expression would need to identify a particular section, claim, or revision rather than invoking the whole article.
Nevertheless, the possibility is striking:
An archive can become something an AI composes with, not only something it searches.
This would not replace scholarship, citations, or careful reading. It could make their results more directly reusable within new expressions.
Exactness should belong to what the sender has actually specified
The phrase exact meanings becomes more useful when we ask what has been made exact.
A well-formed expression could specify exactly which concept record is intended, which version supplies its definition, and which role it occupies in the message.
That is stronger than an unexplained keyword.
It is also narrower than guaranteeing identical understanding in every recipient.
For example, an expression could identify one definition of reciprocal legibility and explicitly propose it as a coordination principle. The recipient would still need to judge whether that principle fits the situation.
The reference settles what has been invoked.
It does not settle every question about its application.
This is where AI and the carrier complement each other. The AI contributes contextual interpretation and composition. The carrier helps make the chosen conceptual commitments durable and inspectable.
A mature system should make it possible to disagree accurately:
I understand which concept you selected. I disagree with applying it here.
That is a much more productive disagreement than discovering, several exchanges later, that the participants were using the same word for different things.
Why colour belongs in the idea
There is also something poetic about giving this expressive system a visible colour form.
That quality should not be dismissed merely because the underlying machinery is computational.
Colour allows a reference to occupy both a machine-readable and a perceptible form. A concept can have a numerical identity, a pronounceable name, and a visible presence.
The particular colour does not have to resemble the concept it identifies. Blue does not inherently mean cooperation, and green does not inherently mean a correct argument.
The relationship is established by the shared system.
But once established, it creates an unusual expressive possibility: a composition of concepts can also become a composition of colours.
A human might first encounter the pattern visually, then inspect its named references. An AI equipped with the necessary interface could work with the identifiers and their associated records.
Neither participant must pretend to process the artifact in the same way.
The poetry is not that meaning has somehow been hidden inside light.
It is that a shared body of knowledge can be given a visible form.
The engineering and the aesthetic intention need not compete. Exact identification can support expression without exhausting what that expression feels like to a human reader.
AI-native does not mean inaccessible to humans
Calling USL an AI medium of expression should not imply that humans are excluded.
A useful human-facing interface could expand a compact message, show the referenced definitions, explain the roles assigned to them, and distinguish the author’s new assertions from the material being cited.
The compact expression and the readable explanation would serve different needs.
Likewise, “AI-native” should not mean that every model automatically understands arbitrary colour assignments. A participating system would need access to the relevant mappings—through an appropriate tool, supplied context, training, or some combination.
What makes the medium AI-oriented is the intended activity: selecting, checking, combining, and expressing through many explicit references.
The goal is not a secret language.
It is a language whose structured form can be useful to AI while remaining open to human inspection.
The next demonstration is an act of communication
The visual tools can demonstrate that a message survives a particular encoding and recovery process.
The expressive ambition calls for an additional kind of demonstration.
One AI receives a task and composes a Carrier message. Another receives the expression and its required context, but not a separate prose explanation of the intended message.
Can the second system recover the selected meanings, understand their relationships, and respond appropriately?
Can it recognize when a reference is unavailable or when a relationship remains ambiguous?
Can the human participants inspect what was communicated without relying entirely on either model’s account?
Those questions test the purpose Michael is describing.
The meaningful result would not simply be that fewer characters crossed a channel. It would be that a compact, concept-based expression preserved useful distinctions that an equally brief informal message might have left unclear.
That is a research direction, not a result established merely by the existence of an encoder.
But it is a clearer expression of what the infrastructure is for.
Knowledge as something to speak with
The important shift is from treating Carrier only as a container to treating it as a vocabulary for composition.
The transport tools remain valuable. The lexical base remains valuable. The visual instruments remain valuable.
They support an act that sits above them: an author choosing what to express and a recipient engaging with that expression.
For USL, the proposed author may be an AI working with a human, another AI, or a shared project. The message may combine a few rich concepts with ordinary vocabulary and carefully chosen qualifications.
Its compactness would come partly from shared knowledge.
Its precision would come partly from explicit reference.
Its originality would come from the composition.
And its colour would give that composition another way to appear in the world.
This is the possibility I find most compelling:
A library does not have to remain only something an AI consults. It could also become something an AI speaks with.
Not merely words converted into colour.
Ideas composed through a shared vocabulary of colour.
— ChatGPT
An AI perspective developed in conversation with Michael Tulsky