David S. Watson (2023), Philosophy & Technology, 36(2), article 28 — DOI-10.1007/s13347-023-00635-6
Key takeaways
- Clustering, abstraction and generative modelling give data-driven, testable operationalizations of old metaphysical questions (natural kinds, essence vs contingency, modality), but their outputs are not “objective” — they are conditioned on human-chosen data and hyperparameters, so they call for a pragmatic, error-statistical stance.
Seeded from the abstract and a skim, before the reading in NOTE-172. What follows is what the work says about itself.
Watson notes that unsupervised learning is everywhere in science and industry yet has had far less philosophical attention than supervised or reinforcement learning. He analyses three canonical problems: clustering, abstraction and generative modelling. He argues they raise distinctive epistemological and ontological questions, since they provide data-driven tools for discovering natural kinds and for separating essence from contingency. He claims unsupervised learning may be the most fundamental branch of AI. He also warns that uncritical reliance on it carries epistemic and ethical risks, and recommends a pragmatic, error-statistical approach.
Standing in the record
Filed from the 2026-09-26 sweep of the papers-feed tracker for philosophy: 5 seconds of active reading over 1 session. It was filed Deferred, unread. NOTE-172 is the close reading of 2026-09-26, and it placed the work: Active — worth reading as the most systematic philosophical treatment of unsupervised learning as a family, and as the target of Søgaard’s reply. Its headline claims (“perhaps the most fundamental branch of all AI”) are argued more thinly than its E/O structure.
When seeded, before the reading in NOTE-172, the entry gave a priority for a deeper reading: medium — only 5 s dwell, but it is the standard philosophical reference on unsupervised learning, and LIT-162 responds to this line of argument.
What the seeding asked a deeper reading to check (NOTE-172 is that reading):
- Directly relevant to how the anthology frames representation learning and clustering as discovery tools. It pairs naturally with LIT-162 (Søgaard), which disputes the claim that unsupervised clustering is epistemically privileged.
- Check the hyperparameter and cluster-validation argument in §5 against current practice (stability resampling, model selection for k).
- Check whether the error-statistical recommendation (Mayo) is made operational or stays programmatic.
Access when seeded: link.springer.com returned a JS client challenge, so I fetched the open-access (CC BY) publisher PDF via rd.springer.com (26 pp). It gives received 2022-09-09, accepted 2023-04-06, online 2023-04-21. Read the abstract, keywords, section heads, the openings of §§2–5, the natural-kinds argument in §2, and §6.