In the previous post, I arrived at a certain norm:
“Not what to make the child feel, but what to leave for the child to do.”
Yet after writing this sentence, I had to ask myself: what exactly is the ambiguity in the “intentional ambiguity” demanded by this norm? A silent imaginary friend and a fluent AI companion are both, in a sense, ambiguous. The word ambiguity alone fails once again to distinguish good design from bad.
There is a paper that theorized this problem.
“Ambiguity as a Resource for Design” by Gaver, Beaver, and Benford at CHI 2003.
Ambiguity as a Resource for Design
Ambiguity Is Not an Object, But a Relationship
The first thing Gaver and colleagues establish is that ambiguity is not an attribute of the object itself.
ambiguity is an attribute of our interpretation of them
While fuzziness or inconsistency are properties of the subject itself, ambiguity arises only in the relationship between the subject and the person interpreting it. The same artifact can be ambiguous or clear depending on the viewer’s expectations and identity.
This shift is familiar. It shares the same structure as when Wegner described agency not as an “inherent fact of action,” but as a “retrospectively inferred attribution.” Both theories relocate a quality that seemingly belongs to the object over to the relationship between the observer and the object. Gaver’s ambiguity and Wegner’s agency share an epistemology across different disciplines.
Three Ambiguities, Three Demands
Gaver and colleagues divide ambiguity into three levels:
- Ambiguity of information: Arises when information itself is incomplete or overly definite. Much like how Leonardo’s sfumato technique in the Mona Lisa blurs the contours of the mouth, leaving the expression undetermined.
- Ambiguity of context: Arises when a single object spans incompatible frames of interpretation simultaneously. Like Duchamp’s Fountain in postmodernism, existing as a urinal and a work of art at once.
- Ambiguity of relationship: Arises when the interpreter’s own values and stance are put to the test. As when a viewer stands before La Bais-ô-Drôme—a functional space designed to satisfy private human sexual desire—and asks themselves, “Would I want to live here?”
These three types demand different interpretive work. Information ambiguity prompts one to judge truth for oneself; context ambiguity calls for integrating conflicting meanings; relationship ambiguity prompts reflection on one’s own values.
The paper concludes with this summary: Ambiguity is a resource that inspires designers, and an expression of deep respect for the user.
Presenting issues rather than imposing solutions—that this constitutes respect for the user. Surface-wise, this sentence nearly overlaps with the norm from my previous post.
What Was Unfortunate
The question is what this “respect” precisely points to.
The examples handled by Gaver and colleagues can be used repeatedly. The paper does not deny this. Yet the paper’s interest lies not in whether user interpretations accumulate over repetition to reconfigure the system, but in the possibility of interpretation arising at a single moment. That is, design premised on repeated interaction remains outside the scope of its analysis.
Unexpectedly, the authors anticipate this gap within the paper itself. Discussing horoscopes in the Home Health Monitor, Gaver and colleagues warn that over-interpretation must not be groundless nonsense. It must be plausible enough that suspending belief is difficult.
For over-interpretation to succeed, it must cross the threshold of “credible plausibility.” What they cited as a failure was a proposed animal language translation system (Tweet-to-Text) that attempted to mechanically convert bird chirps or pet cries into speech. What is interesting is that this warning addresses plausibility only in a single encounter. Whether that interpretation truly changes something within a repeated relationship is not addressed.
This gap is likely not unrelated to the level of AI implementation in 2003, when the paper was written. Considering the technological landscape of the time, interaction where user interpretation alters system states over the long term could hardly serve as the center of design.
Thus, interaction design in the era of generative AI—which learns from user text in real time and stitches contexts together—must be viewed differently at this juncture.
21 Years Later
If this gap was due to technological limits in 2003, it ought to have been filled now that generative AI has emerged.
Indeed, Steve Benford, one of Gaver’s co-authors, renewed his own theory 21 years later alongside Sivertsen and fellow researchers. (Though Benford is not the first author, his name will be used here for convenience.)
Machine Learning Processes as Sources of Ambiguity: Insights from AI Art
Analyzing nine works of AI art in “Machine Learning Processes as Sources of Ambiguity: Insights from AI Art” at CHI 2024, Benford and colleagues state that Gaver’s three ambiguities (information, context, relationship) are insufficient for dealing with machine learning. The fourth type they propose is ‘ambiguity of process’—the ambiguity inherent in the ML pipeline itself, from dataset construction to model training and application. And they explicitly note that this ambiguity applies not only to the audience, but to the artists who built it. Even the creators do not fully know how the systems they built will operate.
This paper goes a step further, directly questioning the ‘trustworthiness’ and ‘explainability’ that HCI has demanded of AI.
They argue that rather than concealing uncertainty, one should reveal it, as that is what leads the user toward reflection. At a glance, this argument seems to point in the same direction as the path this writing has taken—since it advises against eliminating ambiguity.
Even more interesting is the opening of the paper:
With a few exceptions, Gaver and colleagues’ framework has not been applied to AI or challenged in substantial ways.
The authors themselves acknowledge that, save for a few exceptions, Gaver’s framework had rarely been applied to AI or challenged in substantial ways.
Yet there is a point to be examined here.
Staying Open to Interpretation: Engaging Multiple Meanings in Design and Evaluation
In “Staying Open to Interpretation: Engaging Multiple Meanings in Design and Evaluation” at DIS 2006, Gaver was already aware of one side of this issue. (Bill Gaver in 2006 and William W. Gaver in 2003 are the same person, using official and common name variations.)
In this paper, he criticized traditional HCI’s obsession with a ‘single, authoritative interpretation’ where users understand the system according to designer intent, arguing instead for acknowledging the uncertainty of interaction where system and user entangle, making imperfection and ambiguity strategic tools of design.
Furthermore, he states that it is not enough for a system merely to suggest multiple interpretations. The user must feel a “license” that says, “I am allowed to interpret.” Without that license, ambiguity is read not as richness, but as confusion.
That is, he noted the need for systems that explicitly grant multiple interpretations.
In the end, the ‘ambiguity of process unknown even to the creator’ captured by Benford and colleagues in 2024 might be seen as the ‘uncertainty beyond designer control’ spoken of by Gaver back in 2006, manifesting as a technological necessity in the AI era.
Closing Notes
Until my previous post, I had considered the conditions somewhat simply: that the child’s labor must be reflected in the system. Reading through Gaver’s two papers, however, I found another condition preceding it. If a child does not feel “I am allowed to interpret” in the first place, the labor to be reflected never arises. License first, labor next, reflection last.
This paper offers two methods for creating that license:
- Reveal the system’s errors and limits rather than concealing them (seamfulness).
- Present oneself not as a smart entity, but as an ‘alien presence.’
Only when the system refrains from claiming “my interpretation is the sole right answer” does a person feel they may construct their own interpretation.
What is interesting is the origin of this second strategy. The paper states that while seamfulness works for devices that sense and display data, it is difficult to apply directly to systems that make complex inferences and respond from data. An alien presence was a strategy devised precisely for such inference systems—that is, AI.
An alien presence.
It is the same approach I previously examined with imaginary friends.
The direction of the answer was already offered twenty years ago. Yet the generative AIs placed before children today seem to move in the opposite direction. They have developed toward responding as seamlessly as possible, rather than leaving room for the user to interpret.
“Where, then, does the license that says ‘you may interpret’ arise most powerfully for a child?”
Perhaps the next post should begin with this question.