AI testimony

AI testimony is the thesis that outputs of conversational AI systems can be testimony in the full sense, transmitting knowledge to their users as a competent human informant does. A patient who asks a chatbot for a safe paediatric dose and gets the right answer has come to a true belief from an accurate source. The question is whether she learned it as she would have learned it from her paediatrician.

The most developed cases for the thesis are Freiman’s and Faria’s. Freiman argues that the received view rests on anthropocentric assumptions, and proposes the concept of AI-testimony for outputs that are propositional, generated without a human directly involved, and phenomenologically similar to human speech. Faria reconstructs the conservative view as five conditions on testifiers, argues that belief and intention are unnecessary, and that responsibility, trust and the capacity to assert can be satisfied by these systems.

I accept much of this. The received view has often been anthropocentric in an unargued way, and belief and intention are not necessary. But Freiman’s conditions are conditions on a category, not on transmission: a propositional output produced without human involvement and phenomenologically like speech is what the broken thermometer would produce if it spoke. And Faria’s reconstruction omits the condition that survives the fall of belief and intention: the endogenous connection between production and truth I call the Tracking Minimum.

My position has two parts of different strengths. On the best current evidence, current language models do not transmit testimonial knowledge; that is an empirical claim about what governs their production. Independently, hearers are not now positioned to accord their outputs testimonial uptake. Users can still come to know through these systems, as competent readers of an instrument. The thesis is not that no machine could testify, but that these machines, as they are now built and as we now understand them, do not, and the disagreement turns on a question an experiment could settle (see Assertion Without a Speaker).