Justified true belief

Justified true belief (JTB) is the classical analysis of knowledge: S knows that p just in case p is true, S believes that p, and S is justified in believing that p. Its ancestry is usually traced to Plato, whose Theaetetus considers, and rejects, the proposal that knowledge is true belief with an account, and whose Meno distinguishes knowledge from true opinion by its being tied down by reasoning. By the mid-twentieth century the tripartite analysis was widely treated as standard, and Gettier’s counterexamples showed it to be insufficient.

Each condition has its own literature. Truth is rarely disputed. Belief has been questioned: Lackey’s selfless assertion cases suggest that testimony can pass on knowledge from a speaker who does not believe what she says, though they leave the hearer’s belief in place. Justification is where internalists and externalists divide, and where reliabilists propose to replace it with the reliability of the belief-forming process.

The analysis is a useful checklist for the epistemology of AI, provided one asks the right question of it. It is tempting to ask whether a language model has justified true beliefs. Its outputs are often true. Whether it has belief-like states at all is a live research question, on which Herrmann and Levinstein propose standards of accuracy, coherence, uniformity and use, the last requiring that a representation govern the model’s outputs. And justification is unclear for a system whose explanations of its answers are further outputs of the same process.

I think the more productive question is about the user. When she believes what a model says, her belief may be true and justified, and what needs asking is what kind of warrant she has and whether it was transmitted or generated. On my account it is generated by her use, as a reader’s warrant is generated by a well-validated instrument. The source-side condition that matters is not whether the system believes but whether the truth of what it asserts figures in why it asserts it.