Semantic drift happens when the meaning of words, and the relationships between words, gradually diverge from what they were intended to mean – or when their meaning was never defined in the first place. This misalignment, or so-called drift, can often go unnoticed, preventing machine-readable knowledge from moving smoothly and consistently between AI systems.
A well monitored and maintained ontology can help to prevent semantic drift, by establishing a controlled vocabulary and rulebook for the domain. It can constrain what data the AI is permitted to use, and how it can be used. This tends to keep the domain’s terms aligned with the source of truth, and interoperable with other systems in line with ontology standards.
What are the symptoms of semantic drift?
Semantic drift reveals itself through inconsistency in language, data or AI behaviour. Common symptoms include terms that gradually change their meaning; long chats that lose focus; categories that blur; and updates that break earlier definitions.
So for example, an intelligent system might answer a question correctly using a defined term, but then answer the same questions differently, using the term in a new way. It may seem confused about what the user means. Longer conversations might go off-track as the meanings of words no longer fit the evolving context – although this can also be caused by LLMs ‘forgetting’ for example.
Drift is not just a system-wide phenomenon. A 2024 paper published by Meta studied semantic drift in text generation, and the trade-off between information quantity and factual accuracy. It presents a semantic drift score that measures the degree of separation between correct and incorrect facts in generated texts.
What causes semantic drift?
Aside from model updates and new training data, a major trigger for semantic drift is translating terms and concepts into another language, especially where there are multiple meanings for the same word (polysemy). Introducing new technology to an organisation is another key trigger.
Let’s say an organisation’s financial department defines a customer as someone who pays for a service. In another department that’s just launched a free app, new users are also called customers. Because one kind of customer pays and the other doesn’t, these two definitions are now in conflict, causing semantic drift.
In this example the concept ‘customer’ has not been clearly defined by the organisation, or maybe it’s been described inconsistently. Without an application ontology to guide it, the organisation’s AI might fill in the gaps with guesses from general-language patterns, resulting in hallucination.
What are the costs of semantic drift?
Semantic drift creates confusion and mistakes within an organisation, which can be critical in domains such as medical research or advanced manufacturing that depend on data-sharing, semantic precision and collaboration. Risks grow when knowledge is fed into automated workflows without appropriate validation and flagging processes.
The likelihood of semantic drift going undetected is highest in large, data-heavy organisations with siloed knowledge, or where knowledge is shared between multinational or multidisciplinary teams. Semantic drift is a significant hazard in domains where specialised terminology needs to comply with strict legal or ethical rules.
How can an ontology fix semantic drift?
Ontologies are designed or approved by humans, acting as semantic anchors to prevent drift. Drift can be detected by monitoring concept similarity. Validation of meaning can be performed by ontology-driven test suites. Memory can be stabilised using ontology-typed storage, and updates can be gated using ontology layer contracts. Ontology-grounded prompts can then be applied to realign systems.
None of this simple however.
How do I build an ontology?
Emerging ontology-as-a-service platforms potentially offer non-technical people the tools to develop, implement and maintain ontologies by themselves, although complex domains might require support from a knowledge engineer. To begin the process of building consensus around an ontology, it can help to ask subject matter experts three key questions.