If you spend your days building ontologies and knowledge graphs, your job essentially boils down to relentless standardisation (1984 is calling).
You sit with domain experts, architects, and engineers, debating whether a term is a class, an attribute, a relationship… or none of those. You build clean data models so that fifty different databases can finally talk to each other without choking on synonyms.
Under the hood, entire industries are quietly converging on the exact same logic.
We are constructing the shared reasoning boilerplate that allows machines to make sense of our world, and we know that without rigid semantic structure, enterprise AI and GraphRAG pipelines simply collapse into hallucination.
And yet, the moment that same machine turns around and generates natural language back at us, we recoil?
Cringe!
We have all developed an uncanny sixth sense for “AI prose”:
“In today’s rapidly evolving landscape, it is essential to leverage holistic paradigms to unlock synergistic value…”
It is a fascinating psychological tension, because we demand total standardisation for the machine’s knowledge layer, but we have an almost allergic reaction when the machine attempts to standardise our human voice and throws it back at us.
Why are we so comfortable automating the underlying architecture of thought, yet so fiercely protective of how that thought is expressed?
Natural Language Is Human Telemetry
In computer engineering, language is often treated like an API payload: an encoder packages information, sends it across a wire and a decoder unpacks it. Under that model, the ideal message is frictionless, predictable, and standardised.
Machines go: 😎👌
And humans go: 😭👎
And that’s because communication has never been just about transmitting facts.
When a human writes to another human, there is an unwritten contract of cognitive effort and intent (Saussure’s ideas show this quite elegantly, here’s an intro to his work):
- The friction of associative choice (The Paradigmatic Axis): In Saussurean linguistics, speaking is an active navigation of the associative axis deliberately selecting one concept while consciously rejecting dozens of virtual alternatives (in absentia). Choosing a specific metaphor, taking a stance, or framing an idea requires cognitive commitment. In constrast, an LLM predicts the statistically probable token along a linear chain (in praesentia).
- The human imprint of Parole (The living sign): Saussure drew a strict line between la langue (the abstract, shared system of rules) and la parole (the living, individual act of speech). We, humans, don’t connect over the abstract dictionary; we connect through parole: the rhythm, the tone, and the psychological bond between an author’s real mental concept (signifié) and their chosen expression (signifiant).
When an LLM generates prose, it produces the statistical average of language. It eliminates the cost, the effort, and the friction, and because the effort is zero, the social signal collapses to zero. No human observer behind something that is said makes us feel cheated.
My Two Cents On Knowledge Engineering: Structure the Knowledge, Free the Voice
As practitioners working at the intersection of ontologies, NLP, and LLMs, this paradox gives us a clear operational boundary:
- Build rigid, deterministic knowledge graphs, unambiguous data models, and strict validation layers. Use structured semantics to keep AI accurate, grounded, and auditable.
- Don’t use AI to synthesise your perspective or flatten the descriptions.
Standardisation is the death sentence in human conversation.
We are building a world where machines share our logic, but let’s make sure we keep the messy, opinionated friction of human language for our reality.
Clément Dhollande 