C:\CONNY_KNIELING\HOMEPAGE>

Hi, I'm Conny — a PhD candidate in Philosophy at the University of Pittsburgh, advised by Bob Batterman. My dissertation asks how scientists can make use of AI models they can't fully see inside of.

My interests are in general philosophy of science, the philosophy of AI, and the philosophy and history of mathematics. More specifically, I work on questions about AI models used in science, the issue of generalization in deep learning, and trust and values in science.

"DLMs in Science: Epistemic Legitimacy without Representation"

Science has come to depend on AI models no one can fully inspect. Deep learning systems predict protein structures and read medical images accurately, while remaining silent about what their predictions rest on. The standard response to this opacity issue is to demand representational insight: recover what the model has learned to check whether it learned the right things. I argue this is both unnecessary and insufficient: even a fully transparent model would not tell us whether what it learned supports the inferences scientists want to draw from it. Legitimate deployment rests on epistemic practices surrounding the model, not on its internals. I argue that deep learning systems in science are better treated as instruments to be validated than as models to be interpreted, and I set out the conditions under which their scientific use becomes legitimate.

My dissertation consists of four chapters: (1) The Legitimacy Gap, a three-level opacity analysis; (2) Against Representation, a critique of DLMs-as-representational-scientific-models; (3) A Diagnostic Taxonomy of DLM deployments; (4) Workflow-Conferred Legitimacy, my positive account.

[Add courses, teaching philosophy, and syllabi or evaluations here.]

"Deep Learning Models in Science: Reliability Without Representation"
My job-market paper — a single-arc defense of a reliability approach, anchored by AlphaFold. Available on request.

Under review

A paper on AI-assisted science.

A paper on generalization and deep learning.

In progress

Working paper on mechanistic interpretability
Argues that mechanistic interpretability cannot deliver mechanistic completeness, since no available criterion distinguishes the mechanism that a network implements from descriptions that merely fit its behavior; proposes a purpose-relative account of interpretive success in its place.

Working paper on AI in mathematics
Argues that automated methods in mathematics succeed or fail by the non-epistemic values of the practice they enter. A claim supported by the September 2026 declaration in which twenty-five Fields Medalists objected, on largely non-epistemic grounds, to AI companies' use of mathematical problems as benchmarks. Builds on "The Value-Free Ideal in Mathematics" (draft available), which identifies the ideals currently governing mathematical proof and argues that proof practice necessarily involves non-epistemic values.

"The Value-Free Ideal in Mathematics: Proofs and Non-Epistemic Values"
Identifies the ideals currently governing mathematical proof, shows that several are value-free ideals closely analogous to those long contested in philosophy of science, and argues that proof practice necessarily involves non-epistemic values, so that the question is not where to draw the epistemic/non-epistemic boundary but which roles values may legitimately play.

"The Moral Uncanny Valley: On Moral Mimicry and Value Alignment" (essay for public audience, under review)
Argues that value alignment treats moral behavior as mimicry, producing systems exceptionally good at imitating moral agents without meeting any account's conditions on moral agency, and that this carries risks the program has not reckoned with.

"The Ghost Work in the Machine: On the Moral Exploitation of Data Workers" (with Anthony Nguyen, Cornell)
Identifies a distinctive wrong in data labor: exploitation targeting individuals in virtue of their capacity for moral decision-making.

"Procedural Injustice and Automated Decision-Making"
Argues that subjects of automated decisions hold procedural rights irreducible to claims about outcome fairness.

[CV coming soon.]

cok22 [at] pitt [dot] edu