What Are Priors in AI-Mediated Discovery?
One of the most important ideas in AI-mediated discovery is also one of the least understood.
Priors.
Most people encounter AI systems as if they begin from scratch.
A question goes in.
An answer comes out.
It feels immediate.
Objective.
Neutral.
But that isn’t how intelligent systems work.
They don’t approach every question as though they’ve never seen the world before.
They begin with assumptions.
They begin with priors.
What Is a Prior?
A prior is the understanding that exists before new evidence arrives.
It is the system’s starting point.
The expectations it already carries.
The patterns it has learned.
The assumptions through which fresh information is interpreted.
In statistics, priors are the beliefs you hold before seeing new data.
In AI-mediated discovery, they are the model’s accumulated understanding of:
→ who entities are
→ what they do
→ when they are relevant
→ how confidently they can be recommended
They shape the answer before retrieval even begins.
The Discovery Illusion
Traditional search conditioned us to believe discovery starts with a blank page.
You type a query.
The system searches.
The best result wins.
AI-mediated discovery increasingly works differently.
The system already understands something about the world.
Retrieval doesn’t always create understanding.
Often, it refines it.
Fresh evidence arrives.
The prior interprets it.
Resolution emerges.
Which means discovery isn’t just a contest for visibility.
It’s increasingly a contest to shape expectation.
How Priors Form
Priors don’t appear overnight.
They emerge through repetition.
A pathway works.
It gets reused.
Successful outcomes accumulate.
Confidence grows.
The pattern stabilises.
Recommendation
↓
Experience
↓
Reviews
↓
Corpus
↓
Weights
↓
Expectations
↓
Recommendation
The internet becomes memory.
The models inherit it later.
Eventually, certain entities stop being possibilities.
They become expectations.
Why Priors Matter
Two organisations can be equally visible.
Equally retrievable.
Equally cited.
Yet one consistently emerges as the answer.
Why?
Because one already occupies a sharper region of the model’s understanding.
It has a stronger prior.
Priors reduce uncertainty.
And systems optimised for resolution tend to favour lower uncertainty.
The recommendation feels easier.
Safer.
More predictable.
Less costly to reconsider.
The Three Prior Games
Not every organisation faces the same challenge.
Some are building priors.
Some are defending them.
Others are trying to change them.
Build
You don’t yet have a prior.
Your challenge is becoming distinctive enough to matter.
Defend
You have the right prior.
Your challenge is preserving coherence and preventing drift.
Overwrite
You have the wrong prior.
The system understands you.
It simply understands an older version of you.
Your challenge is changing what the world remembers.
Each game requires a different strategy.
The mistake is assuming they’re the same.
How To Diagnose Priors
You can test this today.
Ask multiple AI systems:
What is this company?
What is it best at?
Who is it for?
Then repeat the exercise without retrieval.
Three outcomes emerge.
Accurate and confident
Your prior aligns with reality.
Uncertain
You may not yet have a strong prior.
Confident and wrong
You may have an overwrite problem.
The system understands you.
But it understands the wrong thing.
Confidence can disguise drift.
Beyond Defaults
Defaults explain repeated selection.
Priors explain why repeated selection becomes expectation.
Defaults are behavioural.
Priors are cognitive.
A default says:
“We usually choose this.”
A prior says:
“This is what we already expect to be true.”
One is repetition.
The other is understanding.
Conclusion
The search era asked:
How do we get found?
The default era asked:
How do we get selected?
The prior era asks:
What assumptions already exist before the question is even asked?
Because increasingly, AI-mediated discovery doesn’t begin with retrieval.
It begins with expectation.
The internet becomes memory.
The models inherit it later.
And eventually, the organisations that thrive won’t simply be the easiest to find.
They’ll be the ones future systems already understand.