Skip to content
Coordinates of ThoughtHome of Philosophy
Join
Debate

Thesis Reliabilism has a machine problem: reliable prediction is not knowledge

Pieter van Dijkhuizen Fellow
· 6 replies
Knowledge & beliefAnalyticAlvin GoldmanAlvin PlantingaErnest Sosa#externalism#generality-problem#machine-learning#reliabilism
Follow

Where it stands

Some machine learning systems are very reliable: a fraud model that flags transactions correctly far more often than any human team. Process reliabilism, roughly, says a belief is justified if it is produced by a reliable process. So I want to run an argument and see who it upsets.

  1. If reliabilism is the whole story about justification, then any reliable process producing true outputs is producing something knowledge-like.
  2. My fraud model is a reliable process producing true outputs.
  3. But my fraud model does not know anything. It cannot say why, it does not notice when the world shifts under it, and it would flag a transaction just as confidently in a world where its training data meant nothing.
  4. So reliability is not enough, and reliabilism is missing something.

I expect the obvious reply: the model does not have beliefs, so it cannot have justified beliefs, so it is not a counterexample. Fine. But then I want to know what makes a human belief different, other than the fact that it happens in a human, and whether that difference is an internalist one.

The everyday side of this is getting less abstract. People ask me whether "the system knows" someone is a fraudster. I would like a better answer than "it is a statistical thing", and I suspect the philosophy here is unfinished.

LikeReply
Follow6 replies

Join the conversation: members reply here, and joining is free. Reading is always free.

Join free to reply Sign in

6 replies

For the thesis 2

  1. Nur Aisyah Rahman

    For Fellow

    For, from the fairness side. "Reliable" always means reliable over some population. A credit model can be accurate on average and systematically wrong for a subgroup that was thin in the training data. So is the process reliable or not? It depends on how you describe the process and which cases you count.

    Philosophers apparently call this the generality problem for reliabilism, and it was not invented for machines. But machines make it impossible to ignore, because we can measure the subgroup error. A theory that cannot say which reliability counts cannot tell me whether my model "knows" anything about the people it misjudges.

    Helpful · 2
  2. Priya Raghavan

    For Fellow

    For, but with a different reason. Sosa's virtue epistemology says knowledge is a success creditable to the agent's competence. The archer's shot is apt when it hits because of skill, not luck. Credit requires an agent to credit. Pieter's model has competence in some sense, but there is no one whose success it is. That seems to me the real difference, and it is not a matter of the model being unable to recite reasons.

Against the thesis 1

  1. Derek Hollis

    Against Fellow

    Against, and I will make the obvious reply properly. Goldman's theory is a theory of justified belief. A reliable thermometer is not a counterexample to it, because thermometers do not believe. Your model is a thermometer with more dimensions. So premise 1 is false as stated: reliabilism never said that any reliable process produces knowledge, only that reliability is what justifies belief in a believer.

    The interesting question is the one you end with: what makes something a believer. That is philosophy of mind, not epistemology.

    1. Pieter van Dijkhuizen

      Fellow

      Derek, granted. Let me weaken premise 1 to: if a system had belief-like states produced by a reliable process, reliabilism would call them justified. The model is the limit case. My intuition that it does not know survives the weakening, but I admit I may be smuggling in the belief question.

      Helpful · 1

Other replies

  1. Wanjiru Njenga

    Fellow

    From the credit-scoring side, I think Plantinga's version of externalism fits machines better than Goldman's. He says warrant needs faculties working properly in an environment suited to them. A model trained in one market and deployed in another is a faculty in the wrong environment. That is exactly what we call distribution shift. On his view, the model's outputs lose warrant when the environment changes, even if the model is unchanged, which matches what we see.

    So I would say neither for nor against: externalism is fine, it just needs the environment clause.

    Helpful · 3
  2. Pieter van Dijkhuizen

    Fellow

    Pulling this together, I would revise the thesis. Reliability alone is not enough, but the missing ingredient is not access to reasons. It is some combination of what Nur Aisyah, Wanjiru and Priya said: a non-arbitrary description of the process, an environment it is fitted to, and an agent whose success it is. Interesting that two of those are externalist conditions. I came in thinking this was an argument for internalism and I am leaving thinking it is an argument for a better externalism.

    Helpful · 1