The Ambiguity Problem
Human language is a marvel of compression. We convey complex ideas with minimal bandwidth, relying on shared context, social conventions, and real-time feedback to fill gaps. When I tell a colleague "let's meet for coffee to discuss the project," we both understand this means a casual conversation, probably tomorrow or the day after, at a nearby café, lasting maybe 30 minutes. None of that information was explicit.
This ambiguity is efficient for humans. It lets us communicate quickly without specifying every detail. When misunderstandings occur, we clarify in real-time. When agreements need interpretation, we apply common sense and good faith. When conflicts arise, we have institutions—courts, mediators, social pressure—that resolve disputes by interpreting intent.
AI agents operate in a fundamentally different context. An AI agent executing a contract cannot "interpret intent." It cannot rely on goodwill to resolve ambiguity. It cannot read social cues or apply common sense in the way humans do. When an AI agent receives an instruction to "deliver the order by end of week," it needs to know: which order? What constitutes delivery? Which week? What timezone? What happens if delivery is one hour late?
The Semantic Gap Between Models
The challenge compounds when multiple AI systems need to communicate. Different models process language differently. They were trained on different data, optimized for different objectives, and embed different assumptions about meaning.
Consider a simple concept like "urgent." For one model, trained primarily on business communications, "urgent" might imply same-day response. For another, trained on medical literature, "urgent" might imply immediate, life-threatening priority. For a third, trained on customer service interactions, "urgent" might simply mean "the customer is frustrated."
These semantic differences are usually invisible. Both models will process the word "urgent" without flagging uncertainty. Both will produce confident outputs. But those outputs may reflect fundamentally different interpretations of the underlying meaning.
In human communication, we resolve such differences through dialogue. "When you say urgent, do you mean today or this week?" AI agents currently lack robust mechanisms for this kind of semantic negotiation. They process input, produce output, and move on—often without recognizing that a semantic mismatch has occurred.
Why This Matters for Commerce
Commerce depends on shared understanding. When a buyer and seller agree to a transaction, they need to mean the same thing by price, quantity, delivery terms, quality standards, and payment conditions. Human commerce has evolved elaborate mechanisms to ensure this: standard contracts, industry terminology, regulatory definitions, and legal precedent.
AI commerce needs equivalent mechanisms, but they don't yet exist. Consider an AI agent negotiating a supply contract. The agent agrees to "premium quality materials delivered monthly at competitive pricing." Every term in that sentence is semantically ambiguous:
- Premium quality: According to what standard? Whose definition of premium?
- Materials: Which specific materials? What specifications?
- Delivered: To what location? By what method? Who bears transit risk?
- Monthly: Calendar month? 30-day periods? Beginning, middle, or end of month?
- Competitive pricing: Compared to what benchmark? At time of order or delivery?
Humans would clarify these ambiguities through follow-up negotiation. AI agents need mechanisms to either avoid such ambiguities in the first place or resolve them programmatically when they occur.
Approaches to Semantic Precision
Several approaches can improve semantic precision in AI systems:
Formal Ontologies
Ontologies provide structured definitions of concepts and their relationships. By mapping natural language to formal ontological representations, AI systems can ensure they're referring to the same underlying concepts. Industry-specific ontologies (for healthcare, finance, manufacturing, etc.) can capture domain-specific meaning that general language models miss.
The challenge with ontologies is coverage and maintenance. No ontology captures all concepts, and maintaining ontologies as domains evolve requires ongoing effort. Additionally, mapping natural language to ontological representations is itself an imperfect process.
Semantic Embeddings with Alignment
Modern language models represent meaning as vectors in high-dimensional spaces. By aligning these embedding spaces across different models, we can create translation layers that preserve semantic content. When Model A communicates with Model B, an alignment layer can transform representations to account for differences in how the models encode meaning.
This approach is promising but imperfect. Embedding alignment works well for common concepts but can fail for specialized or novel uses of language. And alignment itself can introduce subtle semantic drift.
Explicit Grounding
Rather than relying on natural language, AI agents can ground their communications in explicit, unambiguous representations. Instead of "premium quality," an agent specifies ISO 9001 certification, specific material compositions, and quantitative tolerances. Instead of "monthly delivery," an agent specifies exact dates, times, and locations.
Explicit grounding sacrifices the efficiency of natural language for the precision of formal specification. This trade-off makes sense for high-stakes transactions where ambiguity could be costly, but may be overkill for low-risk interactions.
Semantic Verification Protocols
Before executing on an agreement, AI agents can verify semantic alignment through structured protocols. "You said X. My understanding is Y. Is that correct?" These verification rounds catch misunderstandings before they cause problems, at the cost of additional communication overhead.
The Role of Standards
Human commerce relies heavily on standards—ISO specifications, industry terminology, legal definitions—that provide shared semantic foundations. AI commerce needs equivalent standards, but developing them presents challenges.
First, AI systems evolve rapidly. Standards that make sense for GPT-4 may not apply to GPT-5. The standards development process needs to be faster than traditional industry standards while remaining rigorous enough to be useful.
Second, AI systems are heterogeneous. Unlike traditional software, where standards can specify exact protocols, AI systems produce probabilistic outputs that may vary even for identical inputs. Standards need to accommodate this inherent variability.
Third, AI capabilities are expanding into domains without established semantic conventions. When AI agents negotiate complex multi-party agreements or coordinate distributed manufacturing, they need semantic frameworks for concepts that barely exist in human practice.
Information-Theoretic Foundations
Our research approaches semantic precision through information theory. By quantifying the information content of communications, we can measure semantic precision mathematically. A message with high semantic precision conveys specific meaning with low uncertainty. A message with low semantic precision could be interpreted multiple ways.
This framework enables several practical applications. We can score the semantic precision of proposed agreements before execution, flagging terms that are likely to cause misunderstanding. We can measure semantic alignment between different AI systems, identifying potential communication problems before they occur. We can design communication protocols that optimize for semantic precision given bandwidth constraints.
Implications for AI Development
The semantic precision problem has implications for how we build AI systems. Current language models are optimized for fluency, helpfulness, and apparent intelligence. They are not optimized for semantic precision—for conveying exactly what they mean with minimal ambiguity.
Future AI systems intended for commerce should be trained with semantic precision as an explicit objective. They should be evaluated not just on whether their outputs are useful, but on whether their outputs convey consistent, unambiguous meaning across different interpretation contexts.
This may require new training approaches. Current methods evaluate outputs based on human preferences, which tend to favor natural-sounding language over precise language. Training for semantic precision may require different evaluation criteria—perhaps formal verification that outputs map to consistent underlying representations.
The Path to Reliable AI Commerce
Semantic precision is not just an academic concern—it's a prerequisite for reliable AI commerce. As AI agents take on more autonomous commercial functions, the cost of semantic misunderstanding grows. A misinterpreted order might mean wrong products shipped. A misunderstood contract might mean unintended obligations. A semantic gap between systems might mean coordinated processes that fail silently.
Building AI systems that can transact reliably requires solving the semantic precision problem. This means formal representations, alignment protocols, verification mechanisms, and standards—the boring infrastructure work that enables exciting applications.
The goal is AI systems that understand each other as well as trained professionals in the same field understand each other—perhaps better, since AI systems can maintain consistency that humans cannot. Achieving this goal will unlock the full potential of autonomous AI commerce.