Computational reliabilism

Computational reliabilism is the view that reliance on an opaque computational process can be warranted by verification and validation methods, robustness analysis and a history of successful implementation, rather than by insight into the process itself. Durán and Formanek developed it in answer to the problem of epistemic opacity in computer simulation: if no one can survey the process, warrant must come from evidence that the process is reliable.

I treat computational reliabilism as an ally rather than a rival. Its proponents have not needed to reclassify opaque computational sources as testifiers to explain why relying on them can be rational, and that is the shape of my proposal for language models: users who learn from them stand to them as competent readers to an instrument, and their warrant is generated by their use.

Burge reached a structurally similar position by another route. Discussing computer proofs, he allows that a recipient who cannot think through a source’s reasoning is reduced to induction over its checked successes, and so to treating the source as an object. Establishing a record under conditions of use and projecting it to the case at hand is the structure of validating an instrument.

What my argument adds is a reason the validation demand is not a formality. For a sound but opaque instrument, validation confirms a connection between output and world that exists whether or not anyone can inspect it. For a system whose production is connected to the truth only through proxies, there is no further connection behind the record to confirm. The record a user projects is all the connection there is.

The picture changes in one place. Where a model’s training signal comes from a deterministic verifier, such as a proof checker or a unit test, production in that domain is regulated by a procedure whose function is to constrain output to truth, and I think that is original satisfaction of the Tracking Minimum. Whether competence acquired under verification transfers to domains where nothing verifies is open, and the burden lies with whoever asserts the transfer.