Epistemic disclosure

Provide insights on the influences, and validity, of the information and data that cause an AI system to reach a particular decision.

This requires presenting the contextual information that shapes or impacts the output. For example, capturing what the model was trained or fine-tuned for, or what the bias metrics are of the resulting model

In some contexts the epistemic disclosure will be more important, for example in public sector services that use AI systems. To achieve epistemic disclosure, the technical architecture will need to be designed in such a way that AI systems are intentionally composed as smaller machines, rather than one singular version of the truth. Examples of this kind of emerging architecture are called compositional chaining.

For example, when using a council advisory service, the user should be able to see more information about what data was used to train it, how it was collected and who funded its development.

IF thinks this pattern will become more important in the future, as technology and the content we consume become less trustworthy, or useful, with the increase of misinformation and AI generated content.