Does the Machine Know?

Since Gettier’s 1963 paper, epistemologists have known that justified true belief is not sufficient for knowledge: you can be justified, and right, by luck. The theories that followed — reliabilism, safety, virtue epistemology — each add a condition to rule luck out. None of them was written with a machine in mind.

This experiment lets you construct a case by setting its features, then runs it through four accounts of knowledge. Some cases every account passes. Some every account fails. The instructive ones are those where the accounts disagree.

What to notice. Reliabilism looks the most generous to machines and virtue epistemology the least. But a broken thermometer in a room held at constant temperature is reliable too, and no one learns the temperature from it. In “Assertion Without a Speaker” I argue that what the accounts share is a demand that the truth of an output figure, through the source’s own operation, in why the output was produced — and that a working barometer meets that demand while current language models do not.