ABHINAVSAXENA

When a competent informant tells me something, I can come to know it on her say-so, because the truth of what she says figures in why she says it. A language model’s answers are often true, but on the best current evidence what governs their production is likelihood shaped by preference training, and the truth reaches them only through proxies. I argue that they do not transmit testimonial knowledge. A user who comes to know something through a model stands to it as a competent reader stands to an instrument — except that a working barometer meets the condition these systems fail.

Data protection and AI governance turn on consent: processing is legitimate if the person has authorised it. But some systems do not act on people whose standing precedes them; they produce it. The labourer whose worn fingerprints fail at the ration shop is not withholding consent — she is failing to register as someone who could give it. Where an institution produces its subjects’ operative standing, I argue, its legitimacy rests not on their authorisation but on obligations they cannot waive: to make them legible, not to abandon them, and to remedy failure through channels that do not presuppose the legibility that failed.

Every classical account of knowledge — justified true belief, reliabilism, virtue epistemology — was built around a subject who believes and can be asked why. A model has outputs that are often true, produced by a process that may be reliable, with no one who can be asked. My own work approaches the question through testimony, and suggests that what decides it is less whether there is a knower than whether the truth of what is said figures, through the source’s own operation, in why it is said.

We are responsible for what we believe in proportion to what we could have checked. When a belief reaches us through a chain — data, model, interface, summary — each link is somebody’s or something’s, and none is fully anyone’s. Responsibility does not vanish in such chains, but it is redistributed in ways our ordinary practices of blame and trust were not designed to track.

We tend to think of acting for reasons as something only a someone can do. But the features that make behaviour count as action — sensitivity to reasons, flexibility across circumstances, the capacity to revise — can each be specified without mentioning experience. If they can be built without it, then either agency never required consciousness or we have built something that merely looks like an agent. The interesting work is in saying which.

The standard worry is that autonomous systems open a “responsibility gap”: no human controlled the outcome closely enough to be blamed, and the system is not the kind of thing that can be. I suspect the worry misdescribes the situation. Responsibility in these cases is not absent but distributed, and the philosophical task is to draw the distribution accurately enough that it can be acted on.

Philosophers spent the twentieth century arguing about teleporters and split brains. Those cases are now operations: systems are checkpointed, forked and merged as a matter of course. Parfit argued that identity is not what matters in survival; if he was right, the question of whether a resumed system is “the same” may be less important than the question of what, in continuity, we actually care about.