Epistemic Deference to AI

Abstract. When should we defer to AI outputs over human expert judgment? Drawing on recent work in social epistemology, I motivate the idea that some AI systems qualify as Artificial Epistemic Authorities (AEAs) due to their demonstrated reliability and epistemic superiority. I then introduce AI Preemptionism, the view that AEA outputs should replace rather than supplement a user’s independent epistemic reasons. I show that classic objections to preemptionism – such as uncritical deference, epistemic entrenchment, and unhinging epistemic bases – apply in amplified form to AEAs, given their opacity, self-reinforcing authority, and lack of epistemic failure markers. Against this, I develop a more promising alternative: a total evidence view of AI deference. According to this view, AEA outputs should function as contributory reasons rather than outright replacements for a user’s independent epistemic considerations. This approach has three key advantages: (i) it mitigates expertise atrophy by keeping human users engaged, (ii) it provides an epistemic case for meaningful human oversight and control, and (iii) it explains the justified mistrust of AI when reliability conditions are unmet. While demanding in practice, this account offers a principled way to determine when AI deference is justified, particularly in high-stakes contexts requiring rigorous reliability.

Introduction. AI systems1,2 increasingly outperform human experts.3 For example, in medicine, AI systems can analyse images such as X-rays, MRIs, and CT scans more quickly and, in some cases, more accurately than doctors, detecting diseases like cancer at earlier stages and with greater precision.4 Similarly, in finance, AI-driven trading algorithms execute high-frequency trades, optimize portfolio management, and forecast market trends with a level of speed and data-processing capability that surpasses human traders.5 As these AI systems continue to improve, a central question becomes: when should we rationally defer to AI’s recommendations, particularly when they are more accurate than those of human expert authorities?

This question is not merely theoretical but has significant real-world consequences. Blind deference to AI could lead to over-reliance on systems that may harbour biases or hidden errors or that might compromise important epistemic virtues.6 Conversely, irrational scepticism towards AI could lead to rejecting better judgments in favour of less reliable human reasoning. Striking the right balance requires a principled approach to epistemic deference to AI.7 In this paper, I develop a total evidence account of AI deference. According to this view, AI outputs should function as what philosophers call contributory reasons rather than preemptive replacements for human judgment. This approach situates the debate about when to rely on AI outputs in expert domains in the context of recent work in social epistemology, particularly discussions on deference to epistemic authority and expert testimony. I examine the implications of a total evidence account of AI deference, showing that it has three key advantages that align with existing concerns about the ethics and responsible use of AI: (i) it mitigates expertise atrophy by keeping human users epistemically engaged, (ii) it provides an epistemic rather than purely moral rationale for meaningful human oversight and control, and (iii) it explains justified epistemic mistrust of AI.

The article proceeds as follows. Section 2 formulates and motivates the case for Artificial Epistemic Authorities (AEAs) and AI preemptionism. I show that classic objections to Preemptionism – such as uncritical deference, epistemic entrenchment, and unhinging epistemic bases – apply in amplified form to AEAs, given their opacity, selfreinforcing authority, and lack of failure markers. Against these shortcomings, section diagnosis, algorithmic efficiency, and materials discovery. However, it still struggles with complex reasoning, competition-level mathematics, and strategic planning. 4 See Liu et al. (2019) for a comparison of AI and healthcare professionals in disease detection, Lebovitz et al (2021) for a critique of AI training and evaluation based on expert knowledge, and Han et al. (2024) for a review of AI in clinical practice. For a recent critical review of claims about AI outperforming human experts, see Drogt et al. (2024). 5 For example, AI-driven algorithmic trading systems adjust to market information more quickly and may generate higher profits around news announcements due to their superior market timing ability and rapid execution (Bahoo et al., 2024; Frino et al., 2017). 6 There are different questions about epistemic authority and deference. My focus here is the deference question: how we should rationally assign special epistemic weight to the views of an epistemic authority (once it is recognized). Other questions concern the i) nature and ii) identification of epistemic authority and iii) the transmission of epistemic goods (Jäger, forthcoming). 7 The literature on epistemic deference to AI remains sparse, with Wolkenstein (2024) and Hauswald (2025; forthcoming) being among the few notable contributions that explicitly draw on social epistemology.

Related work. We rely on others’ expertise all the time. We trust doctors to diagnose and treat our illnesses, lawyers to interpret complex legal statutes, and auditors to ensure financial accuracy and compliance – often without understanding most or even any of the exact reasons and details. In doing so, we grant these experts a special kind of authority, recognising that their specialised knowledge places them in a superior epistemic position relative to us.8 This view holds that A is an epistemic authority (EA) when a person S correctly believes that A has superior knowledge in a specific domain D at a given time t, relative to an epistemic goal G such as gaining knowledge or forming true beliefs. This means A must not only possess epistemic superiority but also be capable and willing to help S Given the above definition, we can ask whether artificial epistemic authorities (AEAs) could exist: AI systems that play a similar epistemic role to human epistemic authorities. There are good reasons to think that some AI systems might qualify as AEAs (Hauswald, 2025) – especially if we entertain the idea that epistemic authority concerns superior epistemic positioning and reliability rather than necessarily beliefs and communicative intent.11 One answer is AI Preemptionism, which holds that when an AEA is a reliable truthtracker, we should not merely add its outputs to our independent reasoning but replace our own reasons with its verdicts.

Though AI Preemptionism has an especially strong appeal, it inherits the classic objections to Preemptionism (Jäger, forthcoming). Indeed, I think that these objections are amplified in certain cases.

Method. Total evidence views of epistemic authority may offer a promising alternative (Dougherty, 2014; Jäger, 2016; Lackey, 2018; Dormandy 2018).17 Rather than advocating for complete pre-emption, total evidence views of epistemic authority hold that while authoritative testimony should carry significant epistemic weight, it should not completely replace an agent’s independent reasons but rather be integrated as a contributory reason with an agent’s own epistemic resources, allowing for aggregation rather than outright replacement of reasons.

This definition contrasts with AI Preemptionism by allowing U to retain and weigh their own epistemic reasons rather than fully substituting them with AEA’s output.18 There are different ways to spell out a total evidence view in detail. I here want to focus on a rendition that answers the concerns that emerged in the previous section. These can be formulated in terms of three desiderata.

Total Evidence View of AI Deference Critical Deference with Oversight: Assume U is faced with the decision to belief p or non-p in domain D, then 1. U withholds of revisits deference to AEA if i. Domain Mismatch: p lies outside A’s validated domain. ii. Reliability Undermining: Evidence suggests that A exhibits systematic bias or recurring errors. iii. Conflicting Authority: A comparably reliable human or AI EA disagrees. iv. Novel Evidence: U holds independent reasons that it is plausible that A did not consider. 2. Otherwise, U defers to AEA.

Discussion. Moreover, this account prevents epistemic entrenchment by ensuring that AI deference remains open to revision. If users defer to AI uncritically, they risk being passively shaped by AI-driven knowledge ecosystems, making it difficult to recognise alternative perspectives, challenge embedded biases, or identify systematic epistemic distortions. To counter this, the model incorporates defeaters19 and overrides that allow users to withdraw deference when clear signs of systematic bias, domain mismatch, or contradictory epistemic authority testimony arise.

Finally, this account also ensures that users do not become severed from their own epistemic bases. The account presented here avoids this by allowing for partial aggregation rather than full replacement. If a user identifies relevant, independent reasons that the AI might not have considered, those reasons can be reintroduced into the evaluation process, either by prompting the AI for an updated assessment or by weighing the user’s own reasoning against the AI’s conclusion. This ensures that epistemic progress is made rather than lost, maintaining a user’s active role in belief formation rather than reducing them to a passive recipient of AI outputs.

One worry about deferring to AI is that human experts – doctors, lawyers, engineers – may gradually lose their domain-specific competences.20 If a specialist consistently defers to an AI’s outputs, they might become a passive conduit for algorithmic decisions rather than an active decision-maker, ultimately losing the skill to check or challenge the AI’s recommendations. This risk of ‘expertise atrophy’ could hence render practitioners less capable of responding effectively if the AI performs poorly or faces unexpected scenarios outside its training scope.

A second advantage of the AI Deference Account is that it provides an epistemic case for maintaining human control, independent of any moral or political values. While normative arguments for human oversight emphasise accountability, responsibility, and the preservation of autonomy (Amoros et al., 2020), the proposed accounts highlight the value of human input in identifying epistemic defeaters. Because the AI’s recommendations are only default-accepted under conditions of demonstrated reliability, situations can and do arise where its authority is overridden, which necessitates a human arbiter to resolve conflicts, domain errors, or contradictory signals.

A third benefit of the AI Deference Account is that it sheds light on why some users epistemically distrust certain AI systems.22 According to this view, the rationality of trust in AI is grounded in the AI’s capacity to meet the conditions for recognised epistemic authority (demonstrated reliability, domain alignment, transparency about its limitations, etc.). If these conditions remain unmet – for instance, due to opaque training data, a poor track record, or repeated biases – then it is epistemically warranted for users to withhold deference and maintain scepticism about the system’s outputs.23 This account clarifies that user mistrust can be reasonable from a purely epistemic perspective, rather than constituting a knee-jerk rejection to technological innovation or status quo bias. If the AI cannot demonstrate that it holds a reliably superior vantage point in the relevant domain, or if it fails to accommodate defeaters that a competent human recognises, then deferring to it becomes irrational. The proposed account thereby validates and explains users’ reluctance when conditions for justified deference are not met, providing a structured epistemic rationale for distinguishing between justified acceptance of AI outputs and rightful epistemic mistrust.

Conclusion. In conclusion, I have examined whether and how we might justifiably defer to AI outputs rather than relying on our own or others’ human judgment. I began by outlining AI Preemptionism, according to which an AI’s recommendation replaces all of a user’s existing reasons. Although Preemptionism appears compelling, classic objections to epistemic Preemptionism are amplified in AI contexts, given opaque algorithms, perceived infallibility, and the risk of reinforcing epistemically corrupt environments. To address these concerns, I proposed a total evidence view of AI Deference that treats AI outputs as contributory rather than pre-emptive reasons. By requiring ongoing reliability assessments, recognising defeaters, and retaining partial user agency, this approach avoids expertise atrophy, provides an epistemic rationale for human oversight, and explains users’ justified distrust when AI systems fail to meet necessary epistemic conditions.