AI provenance: keeping systems honest through ontology alignment
Every AI system has a moment of origin, a birth certificate it cannot read. A model trained on a sprawling corpus of text, images, and code inherits the assumptions, biases, and outright errors of everything it consumed, yet it can never point to the specific source of a single belief. This is the problem of provenance, and it is not a philosophical nuisance. For a large language model advising a logistics manager on shipping routes, or a diagnostic tool suggesting a treatment plan, the inability to trace a conclusion back to a defensible foundation is a liability that no amount of clever prompting can fix. The system may be consistent, always producing the same answer for the same input, but consistency is not the same as coherence; a broken clock is consistent twice a day, but it never tells the right time.
The gap between what an AI says and why it says it becomes existential in fields where the cost of a wrong answer is measured in more than user frustration. Defence logistics, for instance, operates on a web of interconnected data standards where a single misaligned term can cascade into a failed supply chain or a misdirected unit. The UK and NATO have spent years building upper ontologies — formal, machine-readable frameworks that define the fundamental categories of their data — yet these frameworks have historically existed in isolation. The Information Exchange Standard (IES), the Higher Quality Data Model (HQDM) that underpins the National Digital Twin, and Basic Formal Ontology (BFO) each describe the world in their own vocabulary, and no public alignment between IES and HQDM existed until a recent effort by researchers at the University of Cambridge who decided to stop assuming and start checking. [1]
The concrete gain here is not a new algorithm or a faster chip; it is the unglamorous work of reconciliation. The research team took a hand-curated crosswalk of seventeen correspondences between IES and HQDM and translated it into a formal logic language, then let a reasoner — a piece of software that performs logical deduction — loose on the complete merged ontologies. [1] What emerged were three results that matter far beyond the defence niche. First, the alignment proved that the two frameworks could be unified without contradiction, a non-trivial outcome when two systems have been developed independently for years. Second, the process exposed subtle mismatches in how each ontology defines concepts like “person” or “asset,” mismatches that would have remained invisible to a human reader flipping through documentation. Third, and most importantly, the reasoning step turned a static mapping into a living check: any future change to either ontology can now be tested against the alignment to see if it breaks the whole.
This is where AI stops being a magic black box and becomes something more useful: a sceptical colleague. The reasoner does not defer to the authority of the original designers; it asks, relentlessly, whether the logical consequences of the merged system hold. A term defined in IES as a subclass of “event” might, under HQDM’s rules, end up as an instance of “process,” and the reasoner will flag that inconsistency even if no human ever thought to look there. The result is a form of AI-assisted honesty, a tool that forces the underlying data to answer for itself. This matters because the alternative is the silent drift of meaning, where two systems use the same word for different things and nobody notices until a decision made on one side of the data pipeline collides with a decision made on the other.
The structural asymmetry in this story is worth sitting with. Those who build the ontologies, the data architects and standards bodies, are not the ones who pay the price when the alignment fails. The cost falls on the operators in the field, the logistics officers, the analysts, the people who trust the system because they have no way to audit it themselves. They are the ones who receive a recommendation from an AI-assisted tool and have to act on it without knowing whether the underlying categories are compatible. The researchers who did this work are not selling a product; they are providing a public good, a way for the people downstream to hold the system accountable. That is a rare position in an industry where the incentives usually point toward speed and market share, not toward verifiability.
What makes this effort genuinely uplifting is not the technical achievement itself but the posture it represents. Somewhere in the chain of human decisions, someone said: let us not assume that because two documents use the same word, they mean the same thing. Let us check. Let us formalise the check so that it can be repeated, automated, and trusted. That is the quiet version of AI lifting us, not by replacing human judgement but by extending it, giving it a tool that can see around corners the human eye cannot reach. The reasoner does not tell you what to think; it tells you what follows from what you already decided, and that is a gift of clarity in a field drowning in ambiguity.
The voice of those affected, heard last, is the one that matters most. Consider a defence analyst in a joint operations centre, staring at a dashboard fed by multiple data sources. The dashboard looks seamless, but the seamlessness is a lie if the ontologies underneath disagree. With the alignment in place, the analyst can trust that a “vehicle” in one system is the same category as a “vehicle” in another, not just a similar string of characters. That trust is the difference between a recommendation that is acted upon and a recommendation that is quietly ignored because nobody can explain why it makes sense. The researchers have not made the analyst’s job easier; they have made it possible to do it at all, and that is the real measure of progress.
This is the kind of AI advancement that does not make headlines, because it does not promise to write poetry or diagnose cancer. It is infrastructure, the plumbing of meaning that makes everything else work. The reasoner that checks the alignment is not intelligent in any human sense; it is relentless, which is better for this task. The people who built this did not ask what AI could do for them; they asked what it could do for the people who would never read the paper, never see the code, never know the names of the ontologies. Those people will simply experience a system that, for once, does not surprise them with a contradiction. That is the lift, and it is real. It is the quiet work that keeps the machinery of trust running.