1. Introduction
Imagine: two colleagues in your new department provide clashing testimony about how best to navigate a tricky interpersonal situation; a close friend disputes the trustworthiness of another, perhaps motivated by jealousy; a romantic partner provides testimony you have some antecedent reason to doubt – but then again, you’re no expert and feel pulled to give them the benefit of the doubt. Complex testimonial exchanges like these abound. In them, we find ourselves facing the task of interpreting, weighing, and integrating multiple pieces of evidence – sometimes including higher-order evidence – some of them conflicting, sometimes with high stakes, and guided (perhaps licitly, but not always) by our preexisting relationships of trust.
One might expect epistemologists of testimony to have developed theories that help us think through these complex cases. But traditional accounts of testimony and testimonial justification have been developed, defended, and debunked on the basis of simple, highly idealized cases in which speakers offer clear, unambiguous testimony to listeners who understand it perfectly and either clearly possess or clearly lack defeaters for believing it. In such cases, little interpretive work is required of hearers, who need only decide whether to believe what they’ve been told.
While this methodological approach is genuinely valuable in many ways, helping us isolate and test theoretical principles, the results do not always generalize to real-world testimonial exchanges, which are rarely so clean. By abstracting away from the complexity and dynamics of real-world trust and communication, traditional theories leave open important questions about responsibly navigating the complexity of testimonial exchanges in non-ideal settings. As Lackey (Reference Lackey2021: 224) puts it, “Rather than theorizing only about the giving and receiving of testimony in ideal situations, more work needs to be done on how recipients of testimony should navigate the world as it actually is.” This paper takes up that project, using the tools of Bayesian epistemology to think through decisions about whom to trust and how much.
After laying down conceptual foundations and raising what I call the complexity challenge (Section 2), I argue for testimonial underdetermination: testimonial exchanges can underdetermine whom a recipient of testimony ought to trust and how much (Section 4.1). To do so, I employ a formal model of source reliability (Section 3) from Olsson (Reference Olsson2011; Reference Olsson2013), which helps illustrate the difficulty of making determinations of speakers’ reliability in complex cases where the reference class problem looms large. I then contend that testimonial underdetermination supports a kind of synchronic intrapersonal permissivism about trust: in the face of underdetermination, an agent’s evidence at a single time may permit her to adopt any of several incompatible degrees of trust in another (Section 4.2). This might suggest that trust decisions are hard to make. Thankfully, I argue, practical considerations may help out, guiding one’s choice among one’s epistemically permissible options (Section 5). By way of conclusion, I highlight how the resulting approach to testimonial trust opens up new avenues for inquiry about the diachronic aspects of trust and the causes of persistent disagreement and polarization.
2. The complexity challenge
As is standard in the literature, I consider trust a three-place relation between a truster, a trustee, and a domain about which the trustee is trusted.Footnote 1 In this paper, I focus specifically on testimonial trust, which we exhibit when we believe something on another’s say-so. The following is paradigmatic: Steve tells Bill that the Seahawks won last night, and Bill takes Steve’s word for it.
Typically, we trust speakers across a range of thematically or conceptually related propositions. I trust my doctor, for instance, about a wide range of medical propositions – “Tylenol will help you feel better,” “I’m sorry, but we can’t save your toe,” and “that drinking’s really going to catch up to you someday” – but not about the tree species in my backyard, quantum mechanics, or monetary policy. We’ll call a set of propositions that hang together in these ways a domain. Sometimes, as with our loved ones, we trust others across a range of domains, not just across the propositions within one domain.
Putting these points together, here is how I understand testimonial trust.
Trust: A trusts B within domain d if and only if A is disposed to give high credence to or form outright beliefs in propositions in d on B’s say-so.Footnote 2
While trust is great, trusting rationally is even better.
Rational trust: A rationally trusts B within domain d if and only if (i) A trusts B within d and (ii) A has sufficient epistemic reason to regard B as sincere and competent within d.
Speaker reliability is standardly glossed as domain-relative competence and sincerity, so we can abbreviate Rational Trust as follows: A rationally trusts B as to d just in case she has sufficient epistemic reason to regard B as reliable about d.
With these conceptual foundations in place, let’s consider the following case.
Colleagues: Mark and Sarah are colleagues who have developed a friendship. Over several months, they’ve discussed work regularly. Mark has consistently expressed satisfaction with his job, good rapport with his team, and confidence in his abilities. He has never struck Sarah as insincere or incompetent – he’s always forthcoming with his thoughts and reasonable in his self-assessments. So, she trusts Mark about this. A couple of weeks later, however, another colleague, Anna, complains to Sarah about Mark’s work performance. “His work is sloppy,” Anna says. “And honestly, I’m not sure how much he cares about this job.” Sarah is caught off guard: Anna’s report doesn’t match Mark’s own reports at all, but as far as she knows, Anna has always been honest with her. However, Sarah also knows that Mark recently received an award for which Anna was in contention, too, so she wonders whether Anna’s report stemmed from jealousy toward Mark. That said, Anna doesn’t really seem like the jealous type…does she?
This is a tough case. Sarah’s preexisting trust in Mark seems to provide reason to believe what he says, but her preexisting trust in Anna seems to provide reason to believe Anna’s countervailing testimony. The weight of Anna’s testimony might be undercut by higher-order evidence suggesting that she has a reason to be dishonest – but Sarah is unsure how weighty that higher-order evidence is. So, Sarah may reasonably be stumped about whether to continue trusting Mark. In this respect, she faces a trust decision: a choice between continuing to trust Mark or reducing or withdrawing her trust. This is, in essence, a choice between different interpretations of the available evidence. On the one hand, Sarah could view Anna’s testimony as an expression of jealousy, allowing her to preserve trust in Mark. Alternatively, she could treat Anna’s testimony as a valuable perspective that highlights possible blind spots in Mark’s self-assessment, prompting her to reduce her trust in Mark.
Cases with these features are ubiquitous. Consider how often we receive conflicting testimony that requires us to determine which party is correct and whether one party is biased, deceptive, or incompetent: evaluating conflicting news coverage from different outlets, assessing competing reviews of products online, and assessing contradictory accounts of an event from different friends. Think of how often we have higher-order evidence about a source’s reliability and must determine its weight – for example, a family member warns that a particular news outlet is biased, but we’re uncertain about the soundness of their assessment. Take the case of vaccine hesitancy among marginalized communities with reasonable distrust of medical institutions, who must interpret physicians’ advice against that historical background (Goldenberg Reference Goldenberg2016). Finally, consider cases of expert disagreement where competing authorities offer different interpretations of the same evidence (Goldman Reference Goldman2001). In all of these situations, agents face difficult trust decisions like Sarah’s.
What these cases have in common is testimonial complexity. Roughly, a testimonial exchange is complex to the extent that it involves conflicting reports, ambiguous evidence, unclear communication (think: slurred speech, poor hearing, and crowded rooms), higher-order evidence that casts doubt on the trustworthiness of a speaker, or some combination thereof. Testimonial complexity, then, comes in degrees: some testimonial exchanges are quite straightforward – for example, receiving the morning forecast from a reliable meteorologist – while others, like those mentioned above, are far more complex. While developing a full account of testimonial complexity would require saying much more, we have enough detail here for our purposes – namely, thinking through the upshot of testimonial complexity. In complex testimonial exchanges, we find ourselves needing to interpret, weigh, and balance various considerations and pieces of evidence that push and pull in different ways. And it may be unclear how to do so.
To the extent that we must regularly grapple with testimonial complexity, we regularly face the following:
The complexity challenge: The challenge of determining whom to trust – and how much – in the face of testimonial complexity.
The reason this is a challenge is that testimonial complexity can induce what Carr (Reference Carr2020) calls normative uncertainty: uncertainty about what one’s evidence supports (cf. Dorst (Reference Dorst2019) on higher-order uncertainty). More importantly, I will argue below that testimonial complexity can generate full-blown underdetermination about whom we should trust and how much.
But complexity isn’t just a challenge for us – it’s also a challenge for traditional approaches to testimony. Epistemologists of testimony are centrally concerned with the permissibility question: Under what conditions is one epistemically permitted to believe another’s testimony? There are two main camps in this debate. On the one hand, nonreductionists hold that one has defeasible, default epistemic permission to believe a speaker’s testimony whenever one understands her testimony and lacks defeaters for believing it (Reid Reference Reid1764/Reference Reid1997; Burge Reference Burge1993; Simion Reference Simion2021). Crucially, this is a default entitlement: one need not have any evidence about a speaker’s reliability at all in order to enjoy this entitlement. Reductionists, however, argue that one is epistemically permitted to believe a speaker’s testimony just in case one has sufficient undefeated non-testimonial positive reasons to regard the speaker as reliable (Hume Reference Hume1748/Reference Hume1999; Fricker Reference Fricker1994). Other theorists have advanced hybrid views combining elements of these theories (Lackey Reference Lackey2008; Faulkner Reference Faulkner2011).
I think that neither reductionism nor nonreductionism is of much assistance in helping us determine how to make trust decisions in the face of complexity. To see this, let’s think about how they would advise Sarah. Nonreductionists would hold that Sarah has a defeasible default entitlement to believe Mark and a defeasible default entitlement to believe Anna. What she should do is ultimately determined by whether either of these entitlements has in fact been defeated – that is, whether reasons to suspect that Anna is jealous defeat Sarah’s entitlement to believe her or whether Anna’s testimony defeats Sarah’s entitlement to believe Mark. But the whole problem is that Sarah does not know which, if either, of these is the case. Reductionists, meanwhile, would argue that Sarah is justified to believe, and thus trust, whichever speaker her non-testimonial evidence on balance supports believing is more reliable. But again, this is exactly what’s at issue. I say this not to malign either of these theories; they were developed for different purposes than addressing real-world complexity. Given that our purpose here is to address real-world complexity, though, we need a different approach entirely, one that has more promise of offering guidance in situations like Sarah’s. Let’s turn to this now.
3. The model
I’ll argue that Sarah’s evidence underdetermines what she ought to believe about Mark and thus the degree of trust she should place in him in this domain. If this argument succeeds, it will establish the following:
Testimonial underdetermination: Testimonial exchanges can underdetermine whom a recipient of testimony ought to trust and how much.
Making that argument is the task of the next section. To do so, I will draw on Olsson’s (Reference Olsson2011, Reference Olsson2013) formal model. The reason to go formal here is that we’ll need to think about degrees of confidence and how precisely determinations of a speaker’s reliability should shift our opinions and our levels of trust. Formalism, while risky, provides helpful tools for thinking about inherently graded phenomena like these carefully and rigorously. That said, I want to emphasize that the formalism is intended to lay bare a problem we all navigate in testimonial exchanges, if implicitly or subconsciously, and help us theorize about it. I do not intend to suggest that the model perfectly represents explicit thought processes. Moreover, the response I’ll develop to the problem does not require the formalism. The model neither creates nor solves the problem; it just makes its structure apparent and easier to theorize about.
Let’s model Sarah’s doxastic attitudes as degrees of belief – credences – understood as subjective probabilities satisfying the standard axioms of probability theory. Let A be the set of propositions toward which Sarah has doxastic attitudes. We’ll define a credence function
$P\left( \cdot \right):A \to \left[ {0,1} \right]$
, mapping each proposition in A to her credence in it.
Prior to receiving Anna’s testimony, Sarah had high credence in h: “Mark is competent at and enthusiastic about his work.” This credence is based on Mark’s testimony. Sarah also knows that Anna recently lost an award to Mark. Let’s encode Mark’s testimony and Sarah’s knowledge about the award within Sarah’s background information, which we’ll designate i. Sarah later receives new testimony, e, from Anna that Mark’s work is sloppy and he seems apathetic. Given her background information, Sarah is unsure how to revise her credence in h in light of this testimony and, thus, manage her trust in Mark within the relevant domain.Footnote 3
According to Bayesian orthodoxy, agents should revise their credences by conditionalizing on new evidence. If
$P\left( \cdot \right)$
is an agent’s prior credence function and e is new evidence, then her updated credence in some proposition p—her posterior credence in p—should be
$P^\prime\left( {{p}} \right) = P({{p|e}})={{{{P}}({{e|p}})\cdot{{P}}\left( {{p}} \right)} \over {{{P}}\left( {{e}} \right)}}$
. So, Sarah’s updated credence in h, upon receiving e and given her background information i, should be:
$$P(h|e,i) ={{P(e|h,i)\cdot{{P}}(h|i)} \over {P(e|i)}}.$$
The critical term is the likelihood,
$P(e|h,i)$
, the probability that Anna testifies that Mark is sloppy and apathetic, given that he’s actually competent and engaged; he said as much, and Anna recently lost an award to him. Given this setup,
$P(e|h,i)$
is a function of Sarah’s assessment of Anna’s reliability. If she judges Anna highly reliable, then
$P(e|h,i)$
should be low. This is where the complexity of the case bites, because Sarah is uncertain of Anna’s reliability: her recent loss to Mark makes it salient that her testimony might be biased, but Sarah is unsure what specific value in
$\left[ {0,1} \right]$
Anna’s reliability R takes (where 0 amounts to perfect anti-reliability and 1 amounts to perfect reliability). Lacking a precise value for this, Sarah cannot determine what specific value
$P(e|h,i)$
should take.
This kind of uncertainty calls for a model that can represent both an agent’s credence in a proposition and her uncertainty about the reliability of the source on whose word that credence is based. Fortunately, Olsson (Reference Olsson2011, Reference Olsson2013) has developed a model that is up to this task. In his model, the recipient of a report does not assign a precise reliability to a source but instead maintains a probability distribution over possible values their reliability could take in the interval
$\left[ {0,1} \right]$
. I’ll call this distribution a reliability distribution.
Let
$\tau \left( x \right)$
denote a probability density function representing Sarah’s reliability distribution for Anna. As a shorthand, I’ll call any probability density function representing a reliability distribution a reliability profile. For any interval
$\left[ {a,b} \right] \subseteq \left[ {0,1} \right]$
, the probability that Anna’s reliability falls within that interval is given by:
$$P(a \le R \le b) = \int_a^b \tau (x)dx.$$
The mean (or expected value) of Sarah’s reliability distribution for Anna,
$\mathbb{M}\left[ {\tau \left( x \right)} \right]$
, is given by:
$$\mathbb{M}[\tau (x)] = \int_0^1 x \tau (x)dx,$$
which we may think of as Sarah’s best guess as to how reliable Anna is.
With these details in view, we can now look at two key equations in Olsson’s model.Footnote 4 First, when an agent receives testimony e in favor of some proposition h, her prior credence in h is updated according to:
$${P_{t + 1}}\left( h \right) = {P_t}(h|e) = {\mathbb{M}{\left[ {{\tau _t}\left( x \right)} \right]\cdot{P_t}\left( h \right)} \over {\left[\mathbb{M} \right[{\tau _t}\left( x \right)\left] {\cdot{P_t}\left( h \right)} \right] + \left( {1 - \mathbb{M}\left[ {{\tau _t}\left( x \right)} \right]} \right)\left( {1 - {P_t}\left( h \right)} \right)}}$$
where
${P_t}\left( h \right)$
is her prior credence in h,
${P_{t + 1}}\left( h \right)$
is her posterior credence in h after receiving e, and
$\mathbb{M}\left[ {{\tau _t}\left( x \right)} \right]$
is the mean of her reliability distribution for the speaker at time t.Footnote
5
Simultaneously, the reliability profile itself is updated as follows:
$${\tau _{t + 1}}\left( x \right) = {\tau _t}(x|e) = {{x\cdot{P_t}\left( h \right) + \left( {1 - x} \right)\left( {1 - {P_t}\left( h \right)} \right)} \over {\left[ \mathbb{M}\right[{\tau _t}\left( x \right)\left] {\cdot{P_t}\left( h \right)} \right] + \left( {1 - \mathbb{M}\left[ {{\tau _t}\left( x \right)} \right]} \right)\left( {1 - {P_t}\left( h \right)} \right)}}\cdot{\tau _t}\left( x \right)$$
where
${\tau _t}\left( x \right)$
is the agent’s reliability profile for the speaker prior to receiving evidence e, and
${\tau _{t + 1}}\left( x \right)$
is the posterior profile reflecting the agent’s revised judgment about the speaker’s reliability after receiving e.
What matters most for our purposes is that the model shows precisely how uncertainty about a speaker’s reliability affects the evidential force of their testimony, since it allows us to compute the likelihood
$P(e|h,i)$
as an average over possible reliability values:
$$P(e|h,i) = \int_0^1 P (e|h,i,x) \cdot \tau (x|i)dx,$$
where x represents a possible value Anna’s reliability could take. In Olsson’s model, though, reliability is defined as the probability of reporting truthfully, so, more carefully, x represents a possible value for the probability that Anna reports truthfully.Footnote
6
Since e is a report that Mark is incompetent, Anna’s report is false if h (“Mark is competent”) is true. So, the probability of her giving that report,
$P(e|h,i,x)$
, is
$1 - x$
. Substituting into the integral above yields:Footnote
7
$$P(e|h,i) = \int_0^1 {(1 - x)} \tau (x|i)dx = 1 - \mathbb{M}[\tau (x|i)].$$
This result shows that Sarah’s assessment of Anna’s reliability – captured by the mean of her reliability profile for Anna – directly determines the evidential force of Anna’s testimony. The model, then, helps clarify the mechanics of Sarah’s uncertainty. That said, it also leaves something out. While it tells us how to revise our credences based on someone’s testimony, given a reliability profile for them, it does not tell us how a reliability profile should be constructed – that is, how one should determine what values a speaker’s reliability might take and distribute one’s credence across those values. I’ll argue below that there may be no single correct answer to this question. Let’s turn there now.
4. Testimonial underdetermination and permissive trust
4.1. From reference classes to testimonial underdetermination
To say that a source is reliable or unreliable is to say something about how often her reports are correct, among other things. Since we rarely have complete information about any speaker’s reliability, such judgments require generalizing from limited evidence: observations of her past behavior, observations of others in similar circumstances, and relevant background information. Because of this, judgments of a speaker’s reliability – whether precise or rough, qualitative or quantitative – are made relative to reference classes, sets of testimonial exchanges taken to be similar or representative.
Importantly, what counts as “similar” or “representative” can be tricky, depending on which properties of the speaker, the audience, the environment, or the speech act are treated as relevant. For example, we might assess a speaker’s reliability as a coworker, as someone with a personal grievance, or as someone with a strong track record in unrelated domains. Each of these assessments invokes a different reference class. Sometimes, it may be unclear which reference class among several plausible candidates should be used to assess a speaker’s reliability. When multiple reference classes are applicable, we face the well-known reference class problem: the challenge of determining which reference class is appropriate in some context (Reichenbach Reference Reichenbach1949, Fetzer Reference Fetzer1977, Hájek Reference Hájek2007).
Before we continue, two terminological notes are needed. In what follows, I will occasionally say that reference classes are “plausible,” “left open by the evidence,” “ruled out by the evidence,” and “supported by the evidence.” Of course, strictly speaking, reference classes cannot be plausible, nor can evidence support them or rule them out: they are sets of testimonial exchanges, not propositions. These expressions are merely useful shorthands for more cumbersome expressions, such as “Sarah’s evidence rules out regarding the features that define the class as causally or statistically relevant” and the like. Second, I will occasionally talk about “endorsing,” “adopting,” or “prioritizing” a reference class. These should also be interpreted as shorthands for unwieldy expressions like “forming a high credence that a reference class is appropriate and using it in one’s deliberations.”Footnote 8
Thinking back to colleagues, Sarah must have some reference class in mind to construct a reliability profile for Anna. But there seem to be multiple plausible reference classes, including at least: (i) the set of exchanges involving people discussing workplace issues generally; (ii) the set of exchanges involving people complaining about a colleague; (iii) the set of exchanges involving people complaining about a colleague to whom they recently lost an award; and (iv) the set of Sarah’s own past exchanges with Anna. In turn, these different reference classes generate different reliability profiles, which feed into how she should update on Anna’s testimony.
Let’s consider in slightly more detail just two of these plausible reference classes:
Competitors: This class consists of testimonial exchanges involving negative reports from speakers who recently lost a competition to the subject of their testimony. The underlying thought is that competitive dynamics might skew such a speaker’s reliability: losers may exaggerate faults or withhold credit. While there’s little hard data on workplace gossip like this, familiar social dynamics make it plausible that direct competitors tend to be less reliable sources than neutral parties. If Sarah were to assess Anna’s reliability with this reference class in mind, she might reasonably arrive at a reliability distribution with mean 0.3.
Track record: This class consists of Sarah’s past testimonial exchanges with Anna across a variety of contexts. The guiding idea is that a speaker’s reliability across domains in the past underwrites generalizations to new domains: a strong track record justifies a presumption of continued reliability. If Anna has previously been highly reliable in other contexts, Sarah might regard her as likely to be highly reliable in this context, too, adopting a reliability distribution with mean 0.8.
The exact numbers are not important here. What matters is that different plausible reference classes generate very different reliability estimates, with different consequences for Sarah’s posterior in h and for her trust in Mark and Anna.
I think Sarah’s evidence does not determine which of these reference classes is most appropriate, which she should prioritize in her deliberations. Instead, I’ll argue, she faces what philosophers of science call transient underdetermination – a situation in which one’s available evidence leaves open multiple responses to it. To see this, let’s get clearer on underdetermination.
Transient underdetermination: An agent’s available evidence e transiently underdetermines how she rationally ought to respond to it if and only if:
-
i. there exist multiple doxastic responses,
${r_1},{r_2}, \ldots, {r_n}$
to e, -
ii. each response is incompatible with any other response,
-
iii. each response is consistent with e, and
-
iv. e does not favor one response over any other.Footnote 9
Sarah’s situation, I argue, meets all of these conditions.Footnote 10
We have already seen that condition (i) is met. Let’s turn to condition (ii), incompatibility between responses. The relevant sense of incompatibility depends on the type of response at issue. When we’re considering full beliefs, incompatibility is a matter of logical inconsistency. For credences, incompatibility is a matter of probabilistic inconsistency. In Sarah’s case, the incompatibility between prioritizing the competitors reference class and prioritizing the track record class is straightforward: the former would lead Sarah to assign Anna an expected reliability of 0.3, while the latter would lead her to assign an expected reliability of 0.8. Obviously, Sarah cannot simultaneously assess Anna as both highly reliable and unreliable.
Are these two responses both consistent with Sarah’s evidence? It is clear that Sarah’s evidence constrains the reference classes she can reasonably adopt: Her past experience with Anna prevents her from dropping her credence in Anna’s reliability too low, but the possibility of bias compels her to refrain from assigning too high a reliability to Anna. That’s good: we want Sarah’s evidence to constrain how she is permitted to respond to it. But her evidence does not rule out either of the reference classes above as appropriate. The competitors reference class remains viable because Anna genuinely did lose an award to Mark, providing legitimate grounds for considering how competitive dynamics and interpersonal friction affect reliability. The track record reference class remains viable because Anna genuinely has been highly reliable in past interactions, providing legitimate grounds for continued trust. Both reference classes fall within the range of options that Sarah’s evidence leaves open, whereas some other reference class consisting solely of testimonial exchanges with known inveterate liars, for example, is ruled out: Sarah has no good reason to think the features of that class bear on Anna’s reliability.
Finally, does Sarah’s evidence favor either response over the other? To answer this question, we must consider what it would take for evidence to favor one reference class selection over another. Sarah’s evidence would favor the competitors reference class if it provided a clear indication that competitive dynamics are the dominant factor affecting Anna’s reliability in this context – for instance, patterns in Anna’s behavior suggesting she struggles with jealousy or disappointment. On the other hand, her evidence would favor the track record reference class if it provided a clear indication that Anna’s past reliability is the best predictor of her current reliability. But these are exactly the things about which Sarah is uncertain. So, while Sarah’s background information and evidence provide relevant information and constrain plausible reference classes, they fail to favor either of our two central reference classes. The upshot: She faces genuine underdetermination, demonstrating the truth of the testimonial underdetermination thesis.
4.2. Synchronic intrapersonal permissivism in trust decisions
It is common to hold that underdetermination generates leeway in scientific theory choice (Longino Reference Longino2004; Biddle Reference Biddle2013; Turnbull Reference Turnbull2018). Following Jackson and LaFore (Reference Jackson and LaFore2024) and Jackson (Reference Jackson2025), I think the same is true in cases of underdetermination in ordinary situations we confront in our daily lives. To see why, consider the following highly plausible principle:
No arbitrary requirements: Epistemic rationality cannot require an agent to adopt one response to her evidence over another when her evidence does not discriminate between them.
If this principle were false, then rationality could demand that agents make epistemically arbitrary choices, requiring that agents adopt one belief over another without any evidential basis for doing so. That picture is difficult to square with the very idea of epistemic rationality, which is supposed to track reasons given by our evidence. Accordingly, epistemic rationality cannot require Sarah to prioritize one reference class over another: from the standpoint of epistemic rationality, both competitors and track record are live options; she is permitted to regard either as appropriate and to employ it in her deliberations.Footnote 11
Recall that the reliability distribution Sarah accepts for Anna will in turn determine the value of the likelihood
$P(e|h,i)$
via the relationship established in Equation (7) above. This has knock-on effects for Sarah’s posterior credence in Mark’s workplace performance and for her trust in him. To see this, consider two different but rationally permissible ways Sarah might assess Anna’s reliability.
Trust-reducing interpretation: Suppose that Sarah has a prior credence of 0.8 in h, reflecting her high trust in Mark. She then prioritizes the track record reference class and constructs a reliability distribution
${\tau _a}(x|i)$
for Anna with mean
$\mathbb{M}[{\tau _a}(x|i)] = 0.8$
. This would yield
$P(e|h,i) = 1 - 0.8 = 0.2$
. With this value, Anna’s negative testimony would reduce Sarah’s credence in h to 0.5, indicating reduced trust in Mark in this domain.Footnote
12
Trust-increasing interpretation: Suppose instead that Sarah prioritized the competitors reference class and constructed a reliability distribution
${\tau _b}(x|i)$
with mean
$\mathbb{M}[{\tau _b}(x|i)] = 0.3$
. This would yield
$P(e|h,i) = 1 - 0.3 = 0.7$
. With this value, Anna’s negative testimony would actually increase Sarah’s credence from h to 0.9, indicating increased trust in Mark in this domain.Footnote
13
Both interpretations represent equally legitimate, rationally permissible ways for Sarah to construct a reliability distribution for Anna and proceed to revise her credence in h and her trust in Mark. In other words, Sarah could rationally accept either a trust-increasing or trust-reducing interpretation of Anna’s testimony, establishing a species of synchronic intrapersonal permissivism, owing to the difficulty of reference class selection in testimonial exchanges. To the extent that colleagues is representative of complex, real-life cases, rational trust may often be quite permissive in this way.Footnote 14
4.3. Objections and replies
A natural response to the above is that when an agent’s evidence is indeterminate, rationality requires suspension of judgment. If this is right, and Sarah’s evidence fails to determine which of these competing reference classes is appropriate, she should not adopt or prioritize either to construct some new assessment of Anna’s reliability on its basis. Instead, she ought to remain unmoved, holding fixed her attitudes about Anna and her posterior credence in h until more evidence comes in.
The suspension objection has intuitive force, but its force arises, I think, from trading on unclarity about what “indeterminacy” means. We can distinguish different kinds and sources of indeterminacy – and when we do, the suspension objection has much less force in this case. Consider the following cases:
No reasonable interpretation: Miriam is a scientist trying to discern whether Hypothesis Y or rival Hypothesis X is more likely to be correct. She needs to process a batch of raw data that might bear on these hypotheses. There are two models Miriam might use to process the data, each of which is well-established and well-regarded in her field: Model A, a regression with plausible covariates, and Model B with a different specification. Miriam tries both, and it’s a wash: neither model reveals that the data clearly support Hypothesis X or Hypothesis Y. The hypotheses seem to be tied across the acceptable models. No reasonable interpretation of the data, then, yields a determinate conclusion.Footnote 15
Multiple reasonable interpretations: The setup is the same as above, except Model A now determinately supports Hypothesis X over Y, while Model B determinately supports Hypothesis Y over X. Multiple reasonable interpretations of the data yield different conclusions.
In the first case, Miriam faces evidential indeterminacy, while in the second she faces what statisticians, scientists, and philosophers of science sometimes call model uncertainty – uncertainty about which model is appropriate for interpreting one’s data or evidence when multiple models are available.Footnote 16 We must clearly distinguish between these. It is highly plausible that agents facing evidential indeterminacy are rationally required to suspend judgment between responses to their evidence when the evidence fails to discriminate between them on any reasonable interpretation of it. Any response besides suspension would outstrip the evidence.
It is far less clear that suspension is required in the face of model uncertainty. Proponents of suspension here would counsel Miriam either to suspend between these rival hypotheses, since there’s no cross-model winner, or to suspend between rival plausible models entirely. But this would come at a steep cost. Scientists routinely face model uncertainty. If rationality required suspension in these cases, vast swaths of responsible scientific practice would be irrational. Instead, the rational response is not the suspension of judgment about the hypotheses under consideration. Rather, it is to adopt and provisionally rely on a model or average the results of multiple models, guided by various theoretical and practical considerations, proportioning one’s posterior credence in the hypothesis to one’s credence that the model selection or averaging technique was appropriate.Footnote 17
I think that colleagues is relevantly analogous to multiple reasonable interpretations. Both Miriam and Sarah must settle on which features of the phenomena of interest are relevant, which relationships and patterns in the data should be prioritized, and which interpretive framework – whether a formal model, as in Miriam’s case, or a judgment about reference classes and reliability, in Sarah’s – should be employed to interpret the data. And in both, multiple principled interpretive frameworks yield different determinate conclusions, with indeterminacy residing in the choice of framework rather than at the level of evidential support following this choice. The structural analogy between the cases suggests that they require the same normative treatment. If this is correct, it would follow that Sarah, like Miriam, need not suspend judgment between her competing interpretive frameworks and thus need not suspend judgment about Anna’s reliability.
Moving on from suspension of judgment, one might instead object that Sarah should average over the reference classes her evidence leaves open, constructing a weighted combination of the corresponding reliability distributions. The objector might claim that this approach would generate an appropriately cautious assessment of Anna’s reliability. But following this strategy would require Sarah to assign weights to competing reliability distributions, and determining appropriate weights is precisely what is at issue: her evidence provides no principled basis for weighting a distribution constructed from the competitors class more or less heavily than one constructed from the track record class. The averaging strategy thus pushes underdetermination up one level rather than resolving it.Footnote 18 In summary, then, Sarah’s case is genuinely permissive: Owing to underdetermination, multiple incompatible responses to her evidence are epistemically permissible, each pointing to a different posterior level of trust in Mark.
5. The pragmatics of trust decisions and arbitrariness
If underdetermination renders incompatible responses to testimony permissible, how should agents actually make trust decisions? While developing a complete theory of the pragmatics of trust decisions would take us too far afield here, I suggest that while purely epistemic considerations determine the bounds of epistemically permissible trust, practical considerations may guide one’s choice between competing epistemically permissible options when the evidence fails to determine a unique response.Footnote 19
Two kinds of practical considerations are especially salient in testimonial contexts. The first is the value of existing trusting relationships. In the face of underdetermination, agents like Sarah may allow the value of their preexisting trusting relationships to shape their trust decisions.Footnote 20 One must also consider the risks of misplaced trust – personally, professionally, and emotionally. Sarah might worry, for example, that if she continues trusting Mark and he wronged his team members, she risks becoming complicit in a pattern of workplace injustice. Moreover, one must take into account one’s tolerance for epistemic risk. Just as we differ in our willingness to take practical risks, we also differ in our comfort with the risk of error, striking different balances between James’s (Reference James1897) edicts to believe truly and avoid believing falsely.Footnote 21 Of course, not all practical considerations may appropriately guide trust decisions. If I must choose between preserving or reducing my trust in another, and I’m guaranteed a slice of apple pie if I make the latter choice, reducing my trust for that reason would be wrong because it manifests a failure to respect the significance of trust.
This broadly pragmatic approach to trust decisions helps blunt the force of a final objection to the permissivism developed above – namely, that it licenses problematically arbitrary trust decisions.Footnote 22 Following Quanbeck and Worsnip (Reference Quanbeck and Worsnip2025), we can distinguish between what is epistemically permissible or obligatory and what is all-things-considered permissible or obligatory, where the latter incorporates practical, in addition to purely epistemic, considerations. Even if one’s evidence renders epistemically permissible two or more incompatible trust decisions, practical considerations may further constrain what trust decisions one all-things-considered ought to make – perhaps even determining a uniquely permissible response at that level. In such a case, one would be forbidden from the kind of arbitrary flip-flopping between epistemically permissible options that concerns White (Reference White2005) and other proponents of uniqueness. While it may be that the choice between epistemically permissible options is, in some sense, arbitrary at the epistemic level, this arbitrariness is constrained at the level where it matters most – the all-things-considered level.
There is, of course, much more to be said – more than can be said here – about fitting reasons for trust and distrust, about weighing various practical considerations against one another and against epistemic considerations when making trust choices, and about the relationship between the practical and the epistemic. I cannot pretend to be able to resolve these issues here. That said, these remarks, while provisional, point toward valuable new lines of inquiry into how practical and epistemic considerations jointly shape our trust in others – lines I leave for future work.
6. Conclusion
In a wide range of testimonial exchanges, we find ourselves confronted with complexity. When we are, we need to make difficult decisions about whom to trust and how much. These decisions will be sensitive to our assessment of the reliability of those whose testimony we’ve received. And assessing someone’s reliability is no mean feat: these assessments are made relative to reference classes, and it may be genuinely difficult to discern whether any particular, seemingly applicable reference class is most appropriate. Sometimes, multiple reference classes may be appropriate. When they are, we face genuine underdetermination, our evidence leaving open a range of options for how reliable to find someone and how to apportion our trust.
Recognizing that rational trust may often be underdetermined and permissive has several important implications. First, it opens up new hypotheses for the causes of widespread persistent disagreement, polarization, and apparent resistance to evidence, which may in some cases arise due to rational leeway in the interpretation of testimonial evidence, owing to the difficulty of assessing others’ reliability. Moreover, if rational trust is often underdetermined and permissive, this raises new questions about how trust should be managed responsibly over time. My trust decisions right now will shape how I interpret my future evidence and revise my trust. If there were always a uniquely rational trust decision, this diachronic dimension would be trivial. But if there is latitude in trust, then different yet rationally permissible trust decisions today may set me on very different epistemic trajectories tomorrow – something the purely synchronic focus of many traditional accounts of testimonial trust may lead us to overlook.
The most important upshot, however, is this: navigating testimony responsibly in the real world is not always a purely epistemological affair. When our testimonial evidence is ambiguous and underdetermining, as I suspect it often is, practical considerations – relational, moral, social, and political – may play a necessary role in shaping our trust decisions. Epistemologists must not bracket these considerations if we want accounts of testimony and trust to capture their full nature and significance in actual human lives.
Acknowledgments
For helpful comments on earlier drafts of this paper, I thank Sam Fullhart, Taylor Koles, Travis McKenna, Conner Schultz, and the audience at the Fall 2025 UNC PPE Faculty Workshop. Special thanks go to Marina DiMarco, whose generous discussions with me about underdetermination and values in science meaningfully improved the paper.
Competing interests
The author declares no competing interests.
Appendix
This appendix collects details referenced in the main text. Proof 1 establishes the relationship in Equation (7):
$P(e|h,i) = 1 - \mathbb{M}[\tau (x|i)]$
, while Calculations 1 and 2 work through the trust-reducing and trust-increasing interpretations of Anna’s testimony that are referenced in Section 4.2.
Proof 1 We establish the relationship
$P(e|h,i) = 1 - \mathbb{M}[\tau (x|i)]$
, where h represents the proposition that Mark is competent, e represents Anna’s countervailing testimony about Mark, i represents Sarah’s background information, and
$\tau (x|i)$
is the probability density function representing Sarah’s reliability distribution for Anna given background information i.
By the law of total probability for continuous random variables:
$$P(e|h,i) = \int_0^1 P (e|h,i,x)\tau (x|i)dx$$
where x represents a possible value of Anna’s reliability.
Reliability in the model is defined as the probability that a speaker provides truthful testimony. We are computing
$P(e|h,i,x)$
, the probability of Anna’s testimony e given that h is true. When h is true, Anna’s testimony e is false. Thus, the probability that Anna provides this testimony is the complement of her reliability:
$$P(e|h,i,x) = 1 - x.$$
Substituting this relationship into our integral:
$$P(e|h,i) = \int_0^1 {(1 - x)} \cdot \tau (x|i)dx.$$
We split the integral:
$$\int_0^1 {(1 - x)} \tau (x|i)dx = \int_0^1 \tau (x|i)dx - \int_0^1 x \tau (x|i)dx.$$
Because
$\tau (x|i)$
is a probability density function over
$\left[ {0,1} \right]$
, we have
$\int_0^1 \tau (x|i)dx = 1$
. By the definition of expected value,
$\int_0^1 x \tau (x|i)dx = \mathbb{M}[\tau (x|i)]$
.
Therefore,
$$P(e|h,i) = 1 - \mathbb{M}[\tau (x|i)].$$
Calculation 1: Trust-reducing interpretation. Suppose that Sarah’s prior credence in h is 0.8, manifesting high trust in Mark. We are given that Sarah adopts reliability profile
${\tau _a}(x|i)$
with mean
$\mathbb{M}[{\tau _a}(x|i)] = 0.8$
. She then receives e, Anna’s negative testimony about Mark. From the relationship established in Proof 1,
$P(e|h,i) = 1 - \mathbb{M}[{\tau _a}(x|i)] = 1 - 0.8 = 0.2$
. Recall that Anna testifies against h. When
$\neg h$
is true (Mark is incompetent), reliable testimony would correctly report this. Therefore,
$P(e|\neg h,i) = \mathbb{M}[{\tau _a}(x|i)] = 0.8$
. By the law of total probability,
$P(e|i) = P(e|h,i) \times P(h|i) + P(e|\neg h,i) \times \!$
$P(\neg h|i)$
. So
$P(e|i) = 0.2 \times 0.8 + 0.8 \times 0.2 = 0.16 + 0.16 = 0.32$
. We now apply Bayes’ theorem:
$P({{h|e}},{{i}})={{{{P}}({{e|h}},{{i}}) \times {{P}}({{h|i}})} \over {{{P}}({{e|i}})}}$
. Thus
$P(h|e,i) = {{0.2 \times 0.8} \over {0.32}} = {{0.16} \over {0.32}} = 0.5$
.
Calculation 2: Trust-increasing interpretation. Suppose that Sarah’s prior credence in h is 0.8, manifesting high trust in Mark. We are given that Sarah adopts reliability profile
${\tau _b}(x|i)$
with mean
$\mathbb{M}[{\tau _b}(x|i)] = 0.3$
. She then receives e, Anna’s negative testimony about Mark. From the relationship established in Proof 1,
$P(e|h,i) = 1 - \mathbb{M}[{\tau _b}(x|i)] = 1 - 0.3 = 0.7$
. Recall that Anna testifies against h. When
$\neg h$
is true (Mark is incompetent), reliable testimony would correctly report this. Therefore,
$P(e|\neg h,i) = \mathbb{M}[{\tau _b}(x|i)] = 0.3$
, reflecting the bias Sarah attributes to Anna. By the law of total probability,
$P(e|i) = P(e|h,i) \times P(h|i) + P(e|\neg h,i) \times P(\neg h|i)$
. So
$P(e|i) = 0.7 \times 0.8 + 0.3 \times 0.2 = 0.56 + 0.06 = 0.62$
. We now apply Bayes’ theorem:
$P({{h|e}},{{i}})={{{{P}}({{e|h}},{{i}}) \times {{P}}({{h|i}})} \over {{{P}}({{e|i}})}}$
. Thus,
$P(h|e,i) = {{0.7 \times 0.8} \over {0.62}} = {{0.56} \over {0.62}} \approx 0.90$
.
Open access
