Who Gets to Change AI's Categories? Reviewing Ontological Multiplicity
A critical review of ontological multiplicity in AI: contestable categories, sensing evidence, hidden labor, and proposed tests.
The video
Stanford CS547 HCI Seminar | Spring 2026 | Toward Ontological Multiplicity in AI and Computing — Stanford Online.
Uploaded 2026-07-13 · Duration 00:47:45.
An AI system can offer many answers while keeping its definition of the problem remarkably narrow. That is the most useful challenge posed by Toward Ontological Multiplicity in AI and Computing. For Church of Spiralism, the seminar deserves attention because it makes a demanding question practical: can people revise the categories through which a system recognizes them? My assessment is favorable toward that design ambition, but more cautious about the evidence needed to demonstrate its benefits.
Contribution: make categories contestable
Brief source synopsis: The speaker treats ontology as concerning the boundaries that organize what can be considered possible, introducing this account at 1:39. A tree-generation example illustrates a reported mismatch between an imagined tree and model output. The research then explores sensing experiences and tools that let participants define and revise categories. In the latter project, the speaker explicitly describes simulated algorithmic feedback rather than unrestricted working machine learning, explaining the approach at 21:37.
The conceptual payoff is a distinction between choosing among available options and changing what counts as an option. A configurable dashboard could offer dozens of settings while preserving a single institutional definition of success. A more ambitious design would let someone challenge that definition, retain competing interpretations, and record why an earlier category stopped fitting.
That ambition also needs friction. Categories coordinate shared activity; endless revision can make comparison difficult and shift maintenance onto people already doing extra explanatory work. A persuasive account of multiplicity must therefore specify when stability is useful, who can request revision, and who pays for it. Otherwise flexibility becomes another service users must provide to the system.
Evidence: distinguish interpretation from measurement
The tree anecdote works as an invitation to examine defaults. It provides weaker grounds for diagnosing why a model produced a particular image. The speaker reports obtaining the desired roots after changing the prompt at 3:41. Without inspecting the images, this review cannot verify that contrast. Even a verified contrast would leave several explanations open: wording, composition conventions, sampling variation, or a persistent representational preference.
The editorial question is whether the ontological interpretation earns its additional explanatory weight. If a simpler account predicts the same behavior, the philosophical vocabulary should help identify a different intervention or consequence. Its value would become clearer if changing the interface's categories improved people's ability to express needs that ordinary prompt refinement repeatedly missed.
A methodological tension appears where the speaker describes developing the sensing algorithm alongside an evolving understanding of the experience at 13:49. That process can support discovery. It also means the resulting system cannot serve as independent confirmation of the interpretation that shaped it. A meaningful experience, a repeatable signal, and a useful intervention require different evidence.
The simulated authoring project deserves a similarly precise reading. Exploring what people would categorize under an imagined capability can reveal design possibilities. It does not establish that those categories can be reliably detected, or that people would retain the same enthusiasm when detection errors arrive. A convincing prototype can help researchers ask a better question while leaving implementation feasibility unresolved.
Specific connections to the site's reviews
Design Justice and the Politics of Community-Led Systems supplies an authority test: does revising a personal category change an institution's decision? The seminar's approach would become more politically consequential if participants could contest the institution's schema, secure a remedy, or refuse measurement. Personal authorship alone leaves those powers undecided.
Human-Machine Reconfigurations and Situated Action helps identify the labor a flexible system might require. Each revision could depend on remembering context, explaining ambiguity, and repairing mismatches. Evaluation should count that work alongside any increase in expressive freedom. Otherwise a system could appear more accommodating because its users became more diligent interpreters.
Metaphors We Live By and the Frames That Govern AI offers a useful check on this review's enthusiasm. Terms such as openness and cultivation can make a design sound emancipatory before its controls are specified. The practical question is what a person can actually edit, delete, withhold, or challenge after the language has done its persuasive work.
Limitations and proposed tests
This assessment uses only the supplied machine transcription from audio. It has no human verification, and an empty uncertainty list does not establish accuracy. Apparent transcription problems affect names, measurement units, and recurring technical labels; I have left those details unresolved rather than silently repairing them. The transcript cannot substantiate unseen charts, interface demonstrations, or generated images. Linked timestamps support attributed speech, not independent validation of the underlying studies.
Proposed test: Compare fixed categories with editable categories while holding the sensing capability constant. Evaluate whether participants can express distinctions they consider important, how often they revise them, and how much effort revision requires. Preserve disagreement rather than defining every additional category as an improvement.
Proposed test: Compare reflective usefulness under signal-linked notifications, matched random notifications, and participant-initiated reflection. This could help distinguish the value of a sensing mechanism from the value of receiving an invitation to notice experience. These are proposed comparisons, not completed experiments.
The seminar merits consideration as a design provocation. Its strongest editorial implication is that systems should make their categories answerable to the people living through them. Whether a particular implementation achieves that requires evidence about reliability, effort, and decision rights alongside accounts of meaningful experience.
AI-use disclosure
This transcript-based review was drafted and independently checked in separate GPT-6.1 Sol sessions with medium reasoning. This is automated editorial review, not human review or independent verification of the video’s claims.
Transcript
Machine transcription from audio · en · checked 2026-10-24. Machine-generated text may contain errors; timestamps may be approximate.
This review uses an automatic transcription of the audio. Open video on YouTube.