I have a habit I’m slightly embarrassed by. Every so often, halfway through an argument, I stop and ask myself why I actually believe what I’m saying. More often than I’d like, the honest answer is “because everyone around me does”, or “because I read it somewhere once and it stuck”. It’s an unsettling question, but a useful one.
Psychologists call this habit metacognition: thinking about your own thinking. John Flavell coined the term in the 1970s, and Nelson and Narens later gave it a neat structure with two levels. At the object level, you do the thinking: remembering, reasoning, judging. At the meta level, you keep a running model of that thinking. Information travels up (“do I really understand this?”) and decisions travel back down (“I’d better read it again”).
The longer I work with data, the more this looks like Master Data Management to me. Good MDM isn’t about having more data. It’s about having an honest, shared picture of what your data means, who looks after it and how far you can trust it. That is metacognition, only scaled up to an organisation.
Two levels, different setting
In MDM, the object level is the data itself: customers, products, suppliers, accounts. The meta level is everything that explains it: definitions, ownership, rules and lineage. None of it adds a single customer record, yet without it nobody can say what a “customer” actually is. Anyone who has watched sales and finance argue over that question will know exactly what I mean.
How sure are we?
A careful thinker keeps asking how confident they ought to be. The data world has its own version in quality monitoring: profiling, validity checks, scorecards that show where things are going wrong.
There’s a subtler parallel too, which is calibration. If you say you’re 80 per cent sure, you should be right about four times in five. MDM does something similar when it matches records. Two entries might describe the same person, so the system estimates how likely that is, and survivorship rules then decide whose address wins. The much-loved “golden record” is really a best judgement, not a fact carved in stone.
Doing something about it
Noticing a problem is only half the job. A student who realises they haven’t understood a paragraph reads it again. In data, that means stewardship: cleansing, merging, chasing up exceptions and fixing errors at the point of entry. Watching without acting is just worrying, and acting without watching is guesswork. The organisations that do this well keep the two joined up.
What really carries the weight
Nobody can question every belief with equal energy, and no organisation can govern every attribute with equal rigour. That’s why the idea of Critical Data Elements is so helpful. These are the fields where a mistake hurts most, such as a tax identifier, an account number or a legal entity name.
It’s the same question a thoughtful person asks of their opinions: which few of these are actually holding up my decision? Answer that honestly, and you know where to spend your effort.
Where did this come from?
Opinions often owe more to upbringing and habit than to evidence. Numbers can be just as slippery. A figure in a report may have passed through five systems, two transformations and somebody’s stray spreadsheet before it reached you. Data lineage and impact analysis let you trace it back and see what depends on it. Without them, an organisation is a bit like a person with firm views and no memory of how they arrived at them.
One step further: ontologies
If MDM is metacognition for an organisation, an ontology is that metacognition written down in a language machines can reason over. It’s an explicit model of what exists in a domain, how those things relate and what they mean. In practice it’s the meta level made formal: the classes, properties and rules (the T-box, if you speak description logic) sitting above the actual instances (the A-box). It says what a Customer is, rather than who your customers are.
The other metacognitive ideas carry over surprisingly well. Checking instance data against the ontology, with something like SHACL shapes or a consistency check, is monitoring: the system is telling you “this doesn’t fit what I believe about the world”. A reasoner that classifies new entities, spots contradictions or triggers a steward is the control side. Provenance answers where an assertion came from, and your critical data elements are the classes and properties you choose to guard most tightly.
Calibration is the awkward one. Plain OWL is binary: something is consistent or it isn’t. “How sure are we?” needs extra help, such as confidence annotations, provenance weights or probabilistic extensions.
An ontology is a belief too
This is the part I find most interesting. An ontology feels like neutral truth, but it’s really someone’s view of the domain, shaped by their context, just like an opinion. So the same questions apply. Who decided a “customer” works this way, and why? What would make us change the model? When two teams hold different ontologies, where exactly do they disagree? Ontology alignment is, when you strip it down, two people comparing worldviews.
There are risks, too. Formal models are rigid where human beliefs are fuzzy. It’s easy to spend months perfecting the meta level while the data itself stays untouched. And a reasoner is only as wise as the axioms you give it, so a confident but wrong ontology is worse than a vague one.
Where the comparison wobbles
I don’t want to push the analogy too far. Human metacognition is private, intuitive. I like to provoke it, changing it and often unconscious. MDM can’t rely on any of that, because lots of people have to share the same understanding. The meta level has to be written down and agreed: a data catalogue, a business glossary, named owners, clear decision rights.
There’s a cultural point as well. Good metacognition needs a little humility, and so does good data work. If finding an error gets you blamed, people will stop reporting them. If it’s treated as useful information, you’ll hear about problems much sooner.
A few things worth trying
- Spend time on definitions and ownership, not just on collecting more data.
- Show people how confident the system is, rather than hiding it behind a single answer.
- Pick your critical data elements and look after those first.
- Make sure every check has someone who can act on what it finds.
- Let people challenge the data, and the assumptions behind it.
End of rambling.
Metacognition doesn’t make anyone think more, its about thinking better and achieving at a personal perspective on our own thinking. MDM does the same thing for data. The real prize isn’t a bigger pile of information. It’s knowing what it means, where it came from, how far you can trust it and what happens if it’s wrong. For people and organisations alike, a good decision tends to begin with taking a step back.
Leave a Reply