Why AI interpretability and reliabilist epistemology are stuck on the exact same question: what counts as the process?
When someone asks whether a language model or a generative visual engine is "reliable," they are usually asking a question that cannot be answered until a much older philosophical problem is solved first.
In epistemology, externalist reliabilism holds that a belief is justified if it is produced by a reliable cognitive process—one that generates a high ratio of true outputs to false ones. In 2000, Alvin Goldman extended this directly to machines in his work on "telerobotic knowledge." Yet reliabilism has always faced a devastating structural objection, articulated most sharply by Earl Conee and Richard Feldman in 1998: the generality problem.
How Narrow Is the Mechanism?
Suppose you look out the window at twilight and see a silhouette you believe is a stray dog. Was the belief-forming process reliable? To answer that, we have to assign the single historical event (the token) to a broader category of processes (the type).
If we type the process broadly as human vision, it is highly reliable. If we type it narrowly as vision at dusk, at forty metres, by a tired observer through rain-streaked glass, it is deeply unreliable. If we type it narrower still—down to the exact millisecond and firing pattern of your specific retinal ganglion cells—the sample size drops to one, and the process becomes trivially 100% or 0% reliable depending on whether there actually was a dog.
Without a principled, non-arbitrary rule for how broadly or narrowly to type a process, "reliability" is an empty metric. You can make any system look like an oracle or a roulette wheel simply by redrawing the boundary of the mechanism.
Interpretability's Epistemic Twin
Wait long enough in philosophy of mind, and an abstract puzzle eventually reappears as an engineering bottleneck. Today, mechanistic interpretability in AI is the generality problem in a lab coat.
When we evaluate a generative model, what is "the process" that produced the output? Consider the candidate types:
- The Model Level: Treating a 96-layer transformer as a single cognitive faculty. At this level, reliability is a blunt statistical average across everything from Python debugging to seventeenth-century poetry.
- The Layer or Block Level: Isolating the interaction between linear multi-head attention matrices and non-linear feed-forward networks.
- The Circuit or Induction Head Level: Using activation patching to locate specific emergent pattern-completion circuits that fire on particular syntactic structures.
- The Prompt-Trajectory Level: Typing the process by the exact distribution of Query, Key, and Value vectors induced by a specific user's phrasing at a specific temperature.
Just as in Conee and Feldman's critique, if you type the AI process at the model level, the verdict is too broad to justify trusting any tricky edge case. If you type it down to the exact weight-activation path of a single forward pass, every token generation is a unique event with a sample size of one.
Why Opacity Forces an Evidentialist Shift
This is where Conee and Feldman's own alternative—evidentialism—bites hard. Under evidentialism, what a user is justified in believing depends strictly on the evidence the user actually possesses right now about the system's trustworthiness. Because deep neural networks are epistemically opaque, ordinary users do not have access to circuit-level reliability data. They cannot see whether a prompt triggered a robust induction head or a brittle statistical hallucination.
In creative and commercial production, this opacity is fatal. A generative tool that merely approximates a "cinematic look" operates at the broadest, most unfalsifiable process type. When it gets a period detail wrong, there is no boundary showing which sub-process failed.
This is why we built NeuraScene around explicit, hard-coded ontological constraints rather than end-to-end prompt guessing. When you lock a scene in NeuraScene to a specific historical register—requiring, say, an 1890s Century Universal Camera and a Petzval lens—you are manually fixing the process type. You are narrowing the generative mechanism to a falsifiable domain that a human cinematographer can inspect, verify, and anchor against real optical physics. You cannot solve the generality problem in the abstract, but in an engineered system, you can force the process to declare its own boundaries.
Method note. This piece was built in NeuraScene's Advertising Mode using an explicit psychological frame:
- Persuasion trigger: Anchoring
- Target affect: Saudade
- Gestalt principle: Law of Continuity
For a sceptical, technically literate readership, Anchoring the evaluation of AI systems to explicit, falsifiable parameters overcomes the distrust caused by black-box opacity. Visually, the Law of Continuity—grounded in modern natural scene statistics of edge co-occurrence—guides the viewer's eye seamlessly from the inspector's cards into the wrought-iron architecture, while the twilight Saudade register evokes the threshold between human observation and mechanical measurement.
