Trust Isn’t a Verdict

What Onora O’Neill Can Teach Us About Trusting AI

Onora O’Neill is a philosopher at Cambridge who spent a significant portion of her career thinking about trust with real precision. Her 2002 BBC Reith Lectures, later published as A Question of Trust, are probably her most accessible work, and they contain an argument that feels almost uncannily relevant to where we are now. The core claim: we’ve been obsessing over trustworthiness when we should be thinking about the conditions that make trust intelligent.

Blind Trust vs. Reasoned Trust

O’Neill separates two very different things that often get lumped together. Blind trust is what you extend when you have no good basis for extending it, you just go along, defer, accept. Reasoned trust (she sometimes calls it “intelligent trust”) is what you extend when you’ve actually done some epistemic work: you’ve looked at the track record, assessed the competence, considered the incentives, noticed where the accountability mechanisms are.

The problem, she argues, is that a lot of what gets called “trust” in modern life is actually the blind variety, masquerading as the reasoned kind. We say we “trust” institutions or sources without doing the work that would justify that trust, and conversely, we sometimes withhold trust from things that actually deserve it, because we’ve been hurt before and we’ve self-protectively over-generalized.

The same dynamic seems to be playing out with AI tools right now. Some people extend something like blind trust; they take outputs at face value, don’t push back, don’t verify. Others have gone the opposite direction and dismiss the whole thing, which is its own kind of unreasoned response. Neither is what O’Neill is recommending.

The Conditions Question

Where her framework is genuinely useful for anyone thinking about AI and epistemic anxiety is her recommendation to examine whether the conditions for intelligent trust are present. O’Neill doesn’t say “figure out if the source is trustworthy and then act accordingly.” Those conditions, as she works them out, involve three things: competence (can this thing actually do what it claims?), reliability (does it perform consistently?), and honesty, which she understands pretty specifically to include not just avoiding false statements but not deceiving, and not deliberately obscuring what you’re doing.

That third one is where things become complicated with AI systems. O’Neill, writing primarily about media and institutions in her 2002 lectures, was already worried about opacity. She argued that trust requires the possibility of checking; that if you can’t in principle verify what’s being claimed or how conclusions were reached, you lose the conditions that make trust rational rather than blind. A language model that produces confident-sounding text without exposing its reasoning, its training data, its error rates, or its limits would fail this condition pretty badly. Not because it’s necessarily untrustworthy in some deeper sense, but because the epistemic infrastructure for evaluating it isn’t available to the person trying to decide whether to rely on it.

This is distinct from the question of whether the model is accurate. You can have an accurate system that you still shouldn’t trust in O’Neill’s sense, because you have no way of knowing when it’s accurate and when it isn’t.

What She’d Say About the “Is AI Trustworthy?” Question

O’Neill’s later work, particularly in Linking Trust to Trustworthiness (2018) and her collaborations on trust in digital contexts, extends this argument in a helpful direction. The question “is this source trustworthy?” puts all the burden in the wrong place. It asks us to make a global, categorical judgment about an entity – is it good or bad, reliable or unreliable – when what we actually need is a more fine-grained question: trustworthy for what, in what domain, with what kind of verification available to me?

This is almost exactly the question that clinical users of AI tools need to be asking, and almost nobody is framing it this way. The conversations tend to run either toward generalized enthusiasm (”it’s so helpful!”) or generalized suspicion (”how do we know it’s not hallucinating?”), when the more useful question is: what would it take for me to have a reasonable basis for relying on this particular output in this particular context? O’Neill would note that answering that question requires something from the AI’s developers, not just from the user. Accountability structures have to actually exist. There have to be mechanisms for catching and correcting errors. The system has to be, in her term, “assessable”, which means transparent enough that checking is possible, even if you personally don’t do all the checking yourself.

The Epistemic Anxiety Connection

What’s interesting about applying O’Neill here is that she gives a name to something that drives a lot of the anxiety in this space: the problem isn’t uncertainty per se, it’s the absence of the conditions that would let you navigate uncertainty well.

Epistemic anxiety, as I’ve been using that term in this series, often isn’t really about not knowing things. It’s about not being able to figure out what to know, or not having the right tools to evaluate competing claims. O’Neill’s analysis suggests that this is a structural problem, not just a psychological one. If the conditions for intelligent trust aren’t present: if the accountability mechanisms are opaque, if the error correction systems are invisible, if the competence claims can’t be checked, then the anxiety is actually quite a rational response to the situation.

A lot of psychoeducation around AI anxiety tends to frame it as a matter of helping people feel more comfortable with uncertainty (which is fine, as far as it goes). However, O’Neill’s framework suggests there might also be a legitimate epistemic task here: not just tolerating not-knowing, but actively asking whether the conditions for reasoned trust are present, and noticing when they aren’t.

That’s a different kind of intervention. It’s less about soothing discomfort and more about helping someone become a better epistemic agent, which, as it turns out, O’Neill would probably say is the point.

A Note on What She Doesn’t Solve

O’Neill is refreshingly concrete for a philosopher, but her framework has limits. She was writing primarily about human institutions such as media organizations, governments, professional bodies, and the conditions she describes for trustworthiness (transparency, accountability, track records) map imperfectly onto systems that don’t have intentions, don’t have reputations in the ordinary sense, and don’t face the same kinds of consequences for deception.

There’s also a distributional question she doesn’t fully address: even if in principle the conditions for intelligent trust could be established for AI systems, not everyone has equal access to the tools or expertise required to do the evaluation. The cognitive work of assessing competence and reliability takes resources including but not limited to: time, background knowledge, critical infrastructure. A framework built around individual epistemic agency might underestimate how much of that agency depends on social and institutional scaffolding that isn’t evenly available.

These are extensions of her framework, not refutations of it. The basic move of shifting from “is this trustworthy?” to “do I have enough to trust intelligently?” is exactly the kind of reframing that makes the anxiety more navigable, not by eliminating it, but by giving it somewhere productive to go.