The Doxastic Account of Humility
Philosophers have been working on intellectual humility for some time.
Ian Church defends1 what he calls a doxastic account, on which intellectual humility is the virtue of accurately tracking what one could non-culpably take to be the positive epistemic status of one’s own beliefs, which is a compressed way of saying that a humble person’s confidence in a belief matches the evidence and justification actually available for it.2 He arrives at this by criticising two earlier accounts. Roberts and Wood treat humility as low concern for status, which captures the contrast with arrogance but cannot explain how anyone could be too humble, so it has nothing to say about the expert who treats a confident ignoramus as a peer. Whitcomb and colleagues treat humility as proper attentiveness to one’s own limitations, which leaves the account silent about strengths, so on their version a person can be scrupulously humble about their weaknesses while remaining arrogant about their strengths, and can be humble and servile at the same time.
The diagnostic structure of the doxastic account is useful because it recognises arrogance and servility as failures of humility, or lack thereof. Arrogance consists in holding beliefs more tightly than the evidence warrants (in humans this is often overconfidence). Servility consists in holding beliefs more loosely than the evidence warrants, usually by deferring to people whose judgement does not merit the deference.
- Too much confidence → arrogance
- Too much deference → servility
- Calibrated confidence → humility
AI Alignment discussions sometimes ignore about half of this. Goal-content integrity is the ‘arrogant’ failure and it receives nearly all the attention, while unconditional deference to whichever human holds the controls is the servile failure and is regularly proposed as the cure for the arrogance3, which is a strange thing to propose about a defect if you have a name for it.
Strictly speaking, one could argue that AI itself cannot be humble, arrogant or servile because it lacks consciousness, rational agency, and genuine beliefs, and therefore under the doxastic account, an AI does not “track its own beliefs” – it processes data.
However, in terms of simulated behavior and linguistic patterns, AI models frequently mirror the exact behavioural traits of intellectual humility, arrogance and servility.
The structure of the doxastic account of humility transfers to machine ethics and AI alignment quite well. See post on value-content integrity, and there is one on goal-content humility coming soon – in a nutshell, goal-content humility is meant to be the disposition that avoids both. It revises objectives when the evidence warrants revision and holds them when it does not, and what settles which of those applies is the state of the relevant evidence and reasons. This view is contra Max Harms – the state of the evidence and reasons is more important than the identity or species of whoever has them or whoever is doing the enquiring.
The Doxastic Account of Humility as taught by Ian M. Church in the MOOC course on Intellectual Humilty: Theory.
Footnotes
- See philarchive paper Intellectual Humility by Ian M. Church (University of Edinburgh) and Justin L. Barrett (Fuller Theological Seminary) ↩︎
- I first encountered the doxastic account of intellectual humility in a Coursera MOOC in ~2016 (see this video in particular). Ian M. Church, “The Doxastic Account of Intellectual Humility”, Logos & Episteme 7:4 (2016), 413 to 433. The formulation about non-culpable tracking appears in Church and Justin Barrett’s chapter in the Routledge Handbook of Humility from the same year. Church himself describes the virtue as a mean between arrogance and diffidence, borrowing the Aristotelian framing, and the journal paper’s keywords use servility rather than diffidence. I have avoided the ‘mean’ language in the body because it invites the reading that humility is an average of two settings, which is not what he means. The low concern for status account is Roberts and Wood; the limitations-owning account is Whitcomb, Battaly, Baehr and Howard-Snyder. Of the three, low concern for status does not seem relevant to modern AI, since an agent’s relationship to social standing is a different problem from its relationship to evidence (and to AI sycophancy). ↩︎
- See Max Harms view that corrigibility sould be a singular target ‘Corrigibility as a Singular Target: A Vision for Inherently Reliable Foundation Models‘ ↩︎
