The Epistemic Alignment Problem of Machine Learning
The value alignment problem of Machine Learning is the problem of properly aligning the objectives we put into {ML} systems with human values. I argue that deployments of Machine Learning in research contexts create an analogous, Epistemic Alignment Problem. Given a distinction from epistemology between epistemically final and instrumental values, this problem can be seen to have two levels: In level one, an {ML} system is consciously misaligned with an epistemically final value to prioritize an instrumental one. In level two, it is inadvertently misaligned with an epistemically final value. I argue that only level two should truly worry us.
- Published in:
30th Biennial Meeting of the Philosophy of Science Association (PSA2026) - Type:
Inproceedings - Authors:
- Year:
2026 - Source:
https://philsci-archive.pitt.edu/29796/
Citation information
: The Epistemic Alignment Problem of Machine Learning, 30th Biennial Meeting of the Philosophy of Science Association (PSA2026), 2026, https://philsci-archive.pitt.edu/29796/, Boge.2026a,
@Inproceedings{Boge.2026a,
author={Boge, Florian J.},
title={The Epistemic Alignment Problem of Machine Learning},
booktitle={30th Biennial Meeting of the Philosophy of Science Association (PSA2026)},
url={https://philsci-archive.pitt.edu/29796/},
year={2026},
abstract={The value alignment problem of Machine Learning is the problem of properly aligning the objectives we put into {ML} systems with human values. I argue that deployments of Machine Learning in research contexts create an analogous, Epistemic Alignment Problem. Given a distinction from epistemology between epistemically final and instrumental values, this problem can be seen to have two levels: In...}}
