There’s a quiet problem at the heart of the AI boom: everyone says they want “trustworthy AI,” but almost no one agrees on what that actually means.
In our recent paper—The epistemologies of trust: conflicting worldviews in the “Trustworthy AI” discourse—we argue that this confusion isn’t a technical gap. It’s a clash of worldviews. You can read the full paper here: https://rdcu.be/feRMf
“Trust,” we show, isn’t a single measurable property of AI systems. It’s a boundary object: a concept that different communities interpret in fundamentally different ways, while still believing they’re talking about the same thing.
That insight has serious consequences. It means that much of today’s AI governance debate isn’t just unresolved—it’s often misaligned from the start.
The Illusion of a Shared Goal
Across policy documents, corporate statements, and academic research, “trustworthy AI” has become the north star. But as we document, the term lacks a consistent definition despite its ubiquity.
This leads to a paradox:
Users report distrust of AI systems—yet rely on them implicitly
Others trust systems based on narrow competence (e.g., coding), and mistakenly infer broader moral reliability
Behavioral research calls this a mismatch between trust and reliance.
These contradictions aren’t irrational. They reflect fundamentally different ways of understanding what “trust” even is.
Four Epistemologies of Trust
Our paper identifies four distinct worldviews that shape how people define and evaluate trust in AI. These aren’t just academic categories—they map onto real disagreements in policy, design, and public discourse.
1. The Technocratic Paradigm: Trust as Metrology
Here, trust is something you measure.
If a system performs well—accurately diagnosing tumors, generating correct code—it is considered trustworthy. Trust becomes a function of performance metrics like reliability, robustness, explainability, and fairness.
This paradigm treats trust as an optimization problem: improve the metrics, and trust will follow.
But reducing trust to metrics risks producing systems that are statistically safe but socially dangerous.
2. The Relational Paradigm: Trust as Agency and Social Contract
From this perspective, trust is not a property of the system—it’s a relationship.
It emerges through participation, shared norms, and human agency. Trust depends on whether systems respect users as participants in a social contract, not just as end-users.
This paradigm highlights risks like:
Loss of human agency
Over-reliance on automation
Anthropomorphism and manipulation
A system can be technically perfect and still untrustworthy if it undermines human capability or social cohesion.
3. The Pragmatic Paradigm: Trust as Pedestrian Verification
Here, trust is institutional.
We trust systems like aviation or electricity not because we understand them, but because they are:
Regulated
Audited
Boring
As described in the paper, trustworthy systems are those that become “pedestrian”—so reliable and normalized that users don’t have to think about them.
The goal isn’t to make individuals trust AI directly. It’s to build expert institutions that ensure safety on their behalf.
4. The Critical Paradigm: Trust vs. Power
This worldview challenges the premise entirely.
It argues that trust is impossible under current conditions, where:
A small number of firms control models, data, and infrastructure
Incentives are structurally misaligned with users
Power asymmetries eliminate meaningful choice
From this perspective, “trustworthy AI” is not a design problem—it’s a political and economic one.
If users cannot opt out, they are not trusting—they are complying.
Why This Debate Feels Stuck
These paradigms don’t just differ—they conflict.
As we note in the paper, attempts to satisfy all four simultaneously lead to incoherence.
Better metrics (Technocratic) don’t resolve power imbalances (Critical)
Stronger regulation (Pragmatic) doesn’t guarantee agency (Relational)
Enhancing user experience (Relational) can increase manipulation risks
So when organizations say they are building “trustworthy AI,” they are often talking past one another—using the same term to mean different things.
Trust as a Boundary Object
The key move in the paper is to treat trust as a boundary object.
This means:
People can disagree on what trust means
Yet still agree that they “trust” a specific system—for different reasons
This flexibility is not a weakness. It’s what makes coordination possible across disciplines and stakeholders.
But it also means that discussions of “trustworthy AI” must begin by clarifying which definition of trust is in play.
Toward a Pluralistic Framework
If trust is plural, then governance must be too.
The paper outlines how different sources of distrust imply different responses:
Poor performance → improve benchmarks (Technocratic)
Loss of agency → restrict or redesign systems (Relational)
Lack of oversight → audits and regulation (Pragmatic)
Concentrated power → antitrust and structural reform (Critical)
There is no universal solution—only context-dependent ones.
The Deeper Insight: Context Matters
One of the more subtle points (often missed in surface readings) is that these worldviews are not fixed.
As discussed in the later sections of the paper, individuals can shift between paradigms depending on context, drawing on different values and frames.
This suggests that:
Disagreements about AI are not just ideological—they are situational
Effective governance requires translating across these perspectives
In other words, the challenge isn’t just technical or political. It’s cognitive.
Final Thought
“Trustworthy AI” sounds like a technical objective.
It isn’t.
It’s a coordination problem across competing epistemologies—each with its own definition of success, its own risks, and its own blind spots.
Until we acknowledge that, we will keep asking the wrong question.
Not “How do we build trustworthy AI?”
But “Which worldview of trust are we operating within?”
Read the full paper: https://rdcu.be/feRMf