How do you model an agent who knows that they don't know everything? In standard artificial intelligence and economics frameworks, we model an agent's knowledge using a set of possible worlds. If an agent is uncertain about a fact, they simply consider some worlds where the fact is true and others where it is false.
However, a profound paradox arises when we try to model unawareness—the state of lacking the very concepts needed to even conceive of a fact (the "unknown unknowns"). If an agent wants to express that they are unaware of some specific concept $X$, they must formulate the thought "I am unaware of $X$." But by merely thinking about $X$ to declare their unawareness, they have paradoxically brought $X$ into their awareness.
To resolve this, an epistemic logic must allow an agent to recognize their own conceptual limitations existentially: they must be able to know that there exist facts they haven't conceived of, without being forced to identify or name those facts.
Why Fixed-Language Models Fail
Early extensions of epistemic logic attempted to model awareness by introducing an awareness operator, $A_i \varphi$, which states that agent $i$ is aware of the formula $\varphi$. In these traditional Kripke structures, there is a single, fixed global language $\mathcal{L}$ consisting of all possible primitive propositions. Every possible world evaluates the exact same vocabulary.
As pointed out by the authors in Reasoning about knowledge of unawareness, this fixed global language creates a severe structural limitation. Because the space of formulas is identical across all states, an agent is forced to have perfect meta-knowledge of their own awareness boundaries. They either know they are aware of a concept, or they know exactly what they are unaware of. They cannot experience genuine doubt about whether their current mental vocabulary is complete.
For example, a general practitioner might know that medical specialists possess deeper concepts without knowing what those concepts actually are. In a fixed-language model, representing this scenario is impossible because the doctor cannot entertain the possibility of worlds with richer conceptual vocabularies than their own.
The Mechanism: World-Dependent Languages
To break this limitation, Halpern et al. introduced a fundamental shift in epistemic semantics: making the language itself dependent on the possible world.
Instead of a single global language, each possible world $w$ is associated with its own distinct sub-language $\mathcal{L}(w)$, generated by a world-specific set of primitive propositions. An agent's awareness and descriptive capacity fluctuate based on the state of the world they inhabit.
[ World w ] -------------------------> [ World w' ]
Language: L(w) = {classical_compute} Language: L(w') = {classical_compute, quantum_compute}
Agent awareness: {classical_compute} Agent considers w' possible, allowing:
"There exists a concept I am unaware of."
To understand why this works, imagine an agent in a world $w$ where the only available concept is "classical computing." The language of this world, $\mathcal{L}(w)$, contains no formulas about quantum mechanics. Crucially, however, the agent can consider another world $w'$ possible, where the language $\mathcal{L}(w')$ is richer and includes the concept of "quantum computing."
When the agent reasons about their own ignorance, the logic evaluates quantifiers (like "there exists a formula") across the languages of the worlds the agent considers possible. In world $w$, the agent can state:
$$\exists \varphi \, (\neg A_i \varphi)$$
This statement is true because, in the alternative world $w'$, there exists a formula $\varphi$ ("quantum computing") that the agent is unaware of in world $w$. The agent does not need to—and indeed cannot—articulate the formula $\varphi$ within world $w$. The world-dependent language framework allows the logic to validate the existence of the concept from the modeler's perspective, while respecting the agent's conceptual blindness from their own perspective.
Axiomatizing Changing Vocabularies
Varying the language across worlds introduces significant logical complications. In standard modal logic, the Barcan axiom governs how quantifiers and modal operators interact, assuming a fixed domain. When the domain of formulas changes from world to world, standard deductions break down.
In Reasoning About Knowledge of Unawareness Revisited and Reasoning about knowledge of unawareness revisited (Math. Soc. Sci. 2014), the authors constructed sound and complete logical systems by explicitly adapting standard epistemic axioms to account for varying languages. This includes a modified Barcan axiom for unawareness, ensuring that quantification over formulas interacts correctly with what an agent is capable of knowing and perceiving.
This formal adjustment provides a rigorous mathematical bridge between different fields. For instance, Halpern et al. proved in Interactive Unawareness Revisited that these world-dependent awareness structures are essentially equivalent to the complex, multi-tiered state spaces used by economists (such as the Heifetz-Meier-Schipper framework), provided one accounts for the underlying multi-valued logic. This unified the computer science and economics approaches to modeling multi-agent unawareness.
From Total Unawareness to Partial Awareness
The utility of world-dependent languages extends beyond total ignorance to the more nuanced state of partial awareness. In Partial Awareness, the authors extended this logic to capture scenarios where an agent is aware of a high-level concept (e.g., they have heard the term "quantum computer") but lacks the underlying vocabulary to understand or define its specific properties.
By tying the language to the world state, the semantic model elegantly captures how an agent's ability to perceive distinctions and articulate facts shifts based on their epistemic context. Whether modeling legal contracts where parties acknowledge unknown risks, or doctors diagnosing patients using unfamiliar specialties, world-dependent languages provide the mathematical machinery needed to reason cleanly about the boundaries of our own minds.
Key Foundational Work
Reasoning about Knowledge of Unawareness (2006) — contributed 2 concept elements to this direction.
Interactive Unawareness Revisited (2005) — contributed 1 concept elements to this direction.