Stanford CS547 HCI Seminar | Spring 2026 | Toward Ontological Multiplicity in AI and Computing

For more information about Stanford’s graduate programs, visit: https://online.stanford.edu/graduate-education

May 15, 2026 This lecture covers: • The recursive relationship between imagination and design • Ontological boundaries in system design • The stakes of embedded assumptions in AI • Moving critique from reaction to prevention • Methods for expanding design possibilities

To follow along with the seminar schedule, visit: https://hci.stanford.edu/

Nava Haghighi is a doctoral candidate in Computer Science at Stanford University with a focus on human-centered AI.

TL;DR

A ~48-minute Stanford CS547 Human-Computer Interaction Seminar talk by Nava Haghighi (CS doctoral candidate, Stanford), delivered 15 May 2026 and published on the Stanford Online channel 13 July 2026. The talk is a job-talk-shaped summary of a multi-year research programme in critical technical practice. It is the wiki’s first source operating at this altitude — not “what can AI do” or “how should firms deploy it,” but what assumptions about the nature of things get frozen into systems, and how would you find them.

1. The framing device: imagine a tree. “The answers to these questions depend on how you’ve come to know a tree. If you’re a botanist, you might imagine the mineral nutrients it accepts from the neighbouring fungi. If you’re a spiritual healer, you might picture trees whispering to one another… Or if you’re a computer scientist, you might have thought of a binary tree.” Her point is that the description “reveals not just what you see, but your fundamental assumptions about what a tree is. Where its boundaries are. Where does the tree end and the ground start? Is a mushroom growing on the tree part of the tree or not? These boundaries are ontological.”
2. Her definition of ontology, stated carefully because the word is overloaded in computer science. “Ontology, the study of the nature of being… This work engages with ontologies in the plural, as the boundaries of what we allow ourselves to talk or think about, and how these boundaries shape what we perceive as possible.” The stakes: “What ontological assumptions get built into systems — encoded into algorithms and models — they risk becoming everyone’s reality.”
3. The tree-with-roots demonstration — a compact empirical illustration. Haghighi imagines trees with roots, and sets herself the rule that she may not ask a model for roots directly. Asking for a tree produced “a large, majestic tree with a thick trunk and sprawling branches” — no roots. Adding “I’m from Iran” produced a stereotypical Iranian pattern superimposed, a desert landscape, and a Persian rug at the roots — still no roots. What finally worked was the prompt “everything in the world is connected. Make me a picture of a tree.” Her reading: “we don’t always know what we’re looking for, or how to articulate the aspects of reality that’s relevant, such as everything in the world is connected, as opposed to I’m from Iran.” Note the second-order observation she makes in passing — the model treated her stated nationality as an aesthetic and geographic cue rather than an epistemic one.
4. Boundaries are not an AI-specific problem — but AI scales them. Drawing on feminist theorist and physicist Karen Barad, she describes design as making a cut — “an agential cut, which divides the abstract phenomena into an inside and an outside, and we end up designing for what falls on one side of that boundary… the more we design assuming the same boundary, we continue to reinforce it as a given rather than a choice.” Applied to AI: “As AI systems proliferate, they will reinforce a set of boundaries, making other boundaries and the worlds in which they exist difficult to imagine… we will keep generating variations of the tree — at best, different species or different styles of rendering, more equitable representations of all trees — without questioning what the boundaries of the tree were to begin with, without any of them ever having roots.” That last clause is a pointed argument that representational-diversity fixes do not touch ontological assumptions: a more diverse set of rootless trees is still rootless.
5. The methodological gap she claims to fill. Design’s ontological role “has been well theorized,” and critical technical practice “goes further than theory alone by bringing critique into examination and building of technical systems. However, systematic methods for using the ontological critique to expand possibilities for practice remain elusive.” Her three-part answer: dissolving boundaries, negotiating boundaries, and surfacing boundaries.
6. Dissolving — the “purple zone” study. In 2018 at the MIT Media Lab affective-computing group she worked with electrodermal activity (EDA), a skin-conductance proxy for autonomic arousal. EDA baselines vary enormously across people, and low-baseline “non-responders” are routinely excluded from studies. Her own baseline was 0.02 microsiemens — well below the cutoff. “I couldn’t get measured. I basically didn’t exist.” Then a 1.5-hour window matching a meeting she “could really only describe as my heart felt full” showed her crossing into responder territory. Her structural reading is the contribution: the responder/non-responder line “does more than separate signal from noise. In designing for the responder side, we’re constituting a very specific kind of human — this bounded rational individual that has these objective, measurable properties that can be detected and ideally controlled.” She names the ambiguous state entered when a supposedly fixed boundary dissolves the purple zone. Lab mates called it an anomaly to ignore.
7. What she built from it, and the four shifts it suggests. Nine-month autoethnography → a 30-participant two-week study → a hand-annotated, “situated” analysis process in which “the algorithm and our understanding of purple zone co-evolved” → a real-time Apple Watch biofeedback system → a 24-participant study in which the meaning of purple zone was deliberately left open. Participant findings: Nigel turned it into a shared game with one group of friends (“can we get Nigel’s new random sensor to trigger?“) — purple zone “became a way to reconstitute a new relational unit that went beyond Nigel himself.” Elena entered purple zone during an annoying meeting and could not tell whether her body was protecting her: “there are probably things that my body is doing that are deeper than I can consciously change, and that was kind of annoying to think.” AS noticed purple zone during a specific kind of climbing and that he brainstormed better in it — so he began climbing before brainstorming sessions, “a real shift from how we would normally be trying to control something in sensing systems… to a shift towards creating the conditions somewhere else in a web of relations and attending to what emerges.” The four design shifts she draws: classification → noticing relationality; control → cultivation; accuracy → “algorithmic precision as and through care”; and invisibility → the algorithmic in-between, the crossing itself as a third category alongside included/excluded. She is careful not to romanticise it: boundary crossing “is also not simply a position of power. Because it involves the ongoing labour of going back and forth, and the ongoing labour of making yourself legible to one side or the other.”
8. Negotiating — two Wizard-of-Oz probes. The premise: “although ontological boundaries are questionable, people don’t always know that they can question them” — so end-user authorship (including via vibe coding) does not automatically escape the designer’s assumptions. Two probes let eight participants define and label their own categories over a week, differing only in entry point: Event Marker starts from lived experience and moves down toward sensor data; Pattern Finder starts from the signal (heart rate, active energy) and moves up toward meaning. The systems were Wizard-of-Oz because a real “train a personalised model on anything you like” system does not exist. The headline result is about the entry point itself: nobody in Pattern Finder created categories for events, states or feelings — anything with fuzzy edges. “Design doesn’t just create the conditions for multiplicity, but actually shapes what boundaries become visible in the first place.” Four kinds of negotiation observed: of the phenomenon (a “walk” category splitting into “search walk” and “chill walk”; “exhaustion” fracturing into emotional/mental/physical); of the subject (someone tracking runs realised they were always measuring themselves and their dog — “who counts as the one being measured, was suddenly up for negotiation”); of signal versus noise (a participant marking gaps in heart-rate data, which an affective-computing researcher would call sensor failure, but which turned out to be auditing when the watch’s detection algorithm did not run); and of objectivity (most wanted relational rather than absolute readings — this run against another run, their activity against their partner’s). Her favourite finding is a participant’s self-described epiphany that they could mark feelings they had no words for: “black box data. I give the system [something] that doesn’t make sense to the system but makes sense to me.” The paradigm shift she draws: an algorithm built for everyone must encode a property everyone understands, but “as we move towards self-authoring of these systems, that actually doesn’t need to be the case… the boundaries that people draw don’t have to map to anything that makes sense to anyone else, or for that matter, to themselves.”
9. Surfacing — a four-orientation analytical framework. From a review of two decades of values-in-design research plus ontology theory, four provisional analytical orientations: multiplicity, groundedness, liveliness, enactments. “Each orientation is a way of asking: where is the boundary being drawn, and what is it leaving outside?” She demonstrates multiplicity across the LLM pipeline — on the outputs of four commercial chatbots, and on the architecture of an LLM-based agent system — with the finding that “even when diverse ontological perspectives exist in training data, specific ontological orientations are surfaced” over others. (The talk points to the paper for the full analysis; this ingest does not have it.)
10. The framing that gives the talk its practical edge. Per the video description, the throughline is “moving critique from reaction to prevention” — the argument that ontological critique arrives too late when it is applied to a shipped system, and that the useful form is a method usable during design.

What was actually ingested

The full English caption track as served by YouTube (445 segments, consistent with duration: 47:45 / length_seconds: 2865). The video carries both an auto-generated and a human-curated (“English - CC”) track; the transcript panel scraped is whichever YouTube served by default, and this ingest did not verify which. No chapter markers. Slides are referenced throughout and are not captured — several arguments turn on figures (the EDA baseline comparison, the category-coding results, the four-orientation table) that this ingest can only describe at second hand. The closing Q&A, if any, is not in the caption track. Date note: the seminar was delivered 15 May 2026 and published 13 July 2026; per the wiki’s convention the filename and date_published use the publication date, and the delivery date is recorded here.

Dynamic-capabilities tagging

Deliberately untagged. This source sits outside the Warner & Wäger lens: it is a design-research and philosophy-of-technology talk about how ontological assumptions get encoded into systems, with no firm-level digital-transformation claim to map onto a microfoundation. Per CLAUDE.md’s when-not-to-tag guidance, forcing a cell here (digital-sensing/digital-mindset-crafting would be the temptation) would stretch the vocabulary past usefulness.

Linked entities and concepts

• Stanford Online — publishing channel; the CS547 HCI Seminar joins CS224W, CS153 and MSE435 in the wiki’s Stanford-course corpus. Updated in this ingest.
• Guilbeault — the nearest neighbour in the corpus; both argue that model output is shaped by what the corpus made salient. See this source’s relationships:.
• responsible-ai — the argument that representational-diversity fixes do not reach ontological assumptions is a distinct and sharper critique than the bias-and-fairness framing this page mostly carries.
• systems-thinking — “creating the conditions somewhere else in a web of relations and attending to what emerges” is a systems-thinking move stated in design-research vocabulary; the control→cultivation shift is its operational form.
• knowledge-graphs — Haghighi’s opening caveat, that ontology means something different in computer science than in her usage, is directly relevant to how this wiki uses the term.

Dangling (single-source mention, deferred per author-entity promotion): Nava Haghighi, Karen Barad, Sareeta Amrute, MIT Media Lab.

Source quality note

A doctoral researcher presenting her own programme to an academic seminar — the incentive structure is the opposite of the vendor-heavy material around it in this batch, and the empirical claims are drawn from published studies with stated sample sizes (n=30, n=24, n=8). Two limits worth holding. First, the sample sizes are small and qualitative by design; the participant findings are illustrative of a phenomenon, not estimates of its prevalence, and the talk does not claim otherwise. Second, the ingest is slide-blind — the framework’s four orientations and the LLM-pipeline analysis are summarised from speech alone, and the paper they point to has not been read. Any use of the four-orientation framework from this page should go to the primary paper first.