Seeing with Machines

Reading File 002  |  The Perception Machine  |  Chapters 3–5  |  Zenya He

After reading the first few chapters of this book, I asked myself: When machines come to know us, how should we interact with them? They observe our bodies and appearances, and given the current development of AI, it seems we are about to lose the right to define ourselves. The book doesn't directly discuss system-recommended content for users, but it inspired my thinking and provided entry points for my research on viewing, classification, and autonomy.

The Importance of Viewing

In Chapter 3, Zylinska turns from repeatedly failing in The Last of Us to pausing, observing, and taking screenshots. Her discussion brings the body, controllers, virtual cameras, and game space into the process of seeing (Zylinska, 2023, pp. 72–79).

“Vision is thus never just visual.” (Zylinska, 2023, p. 78; Translation by the author.) This makes me wonder: images of clothing cannot convey the wearing experience. Machines lack the ability to capture the feel of fabric and the atmosphere of a scene. So how can these elements be incorporated into a user's decision-making process? When a system helps users narrow their choices to make product recommendations more precise, is it a help or a constraint?

My assessment is that when users understand and accept the filtering criteria, narrowing the scope may reduce the decision-making burden, but it may also deprive users of a rich and diverse experience. For example, when a user actively requests "slimming" or "heightening" effects, the system assists in providing choices based on the user's already confirmed data. Should it also provide other suitable options besides "slimming" or "heightening" while meeting the user's direct needs? I think it's possible, but the choice should be made by the user.

Constraints arise when conditions are hidden, default values ​​are used for a long time, and users are neither aware of the impact nor able to find an opportunity to modify them. The system should proactively explain which conditions were used and allow users to adjust them. Users should not first realize they are trapped in a recommended range before having the opportunity to correct it. A goal proposed at once also needs to be distinguished as either for this specific use or for preservation with her consent.

How to Maintain Revisation in Professional Classifications

Chapter 4 discusses "nontrivial machines" using cybernetics and perception research, and questions the formation of classifications. The author further analyzes racialized visual infrastructure, demanding that images, databases, and their history and power relations be included in the examination. Her ethical and political demands cannot be simplified to "making recommendations more personalized." (Zylinska, 2023, pp. 101–117)

"It will need to be regularly rebuilt, retuned, and reprogrammed." (Zylinska, 2023, p. 117; translation by the author of this article.)

I strongly agree with this view. For AIFFD, the understanding and analysis of people needs to retain a certain safety margin. The influence of physical characteristics is on appearance and fit, but the user's own pursuit of style can be retained; she can adopt strategies to change based on her own characteristics and desired style. For example, can a user with a youthful style, due to her slender figure, exhibit a dramatic, leading lady style? Absolutely. Using dramatic outerwear paired with exaggerated colors suitable for a youthful figure can create a striking visual effect that softens a slender physique while highlighting a dramatic, edgy style. Users have the right to explore different styles, and the specific implementation rules require consideration of clothing, body type, and the external context.

Therefore, professional categories should provide explanatory references, and the user interface should demonstrate and explain the paths and choices to achieve this goal. Even if a user understands the system's suggestions but still chooses a option that leans towards the system's recommendations, the system cannot forcibly reject it, because the user's choice is the starting point of the system's work, and the user's expressed needs are the direction of the system's work.

The System Must Bear Responsibility for Recommendations

The user's decision regarding their goals does not absolve the system of responsibility for the basis of those recommendations. In my vision for AIFFD, suppliers provide product attributes, professional stylists analyze the relationship between style, material, and different body types, and the system then matches these with the user's goals. "Slimming" and "height-enhancing" should be conditional effect judgments, not automatically valid simply because a merchant has uploaded promotional tags; the professional rules themselves also need to be tested and revised.

This is related to AIFFD's positioning of providing clothing solutions to users. Style and materials both influence the final result, while price is considered in budget and implementation costs. If the user sets a specific budget limit, the system should adhere to it; however, even within a budget, a higher price should not be directly equated with greater suitability. The validity of a solution depends on the user's accepted goals.

These are design stances I've derived from reading, and they don't yet prove that the corresponding designs will improve autonomy or trust. To test them, we need to observe whether users can understand the basis for the recommendations, point out inapplicability, and whether this feedback truly changes subsequent judgments.

Leaving Room for Interactions Without Predicted Outcomes

Chapter 5 discusses the relationship between static and moving images and explores creation within automation through AUTO-FOTO-KINO. Zylinska re-encodes Chris Marker's The Jetty using tools like StyleGAN2 and GPT-2, creating A Gift of the World (Oedipus on the Jetty). She retains some of the offsets in the generation process, reading them as opportunities to rethink gender and existing narratives; human choices, editing, and interpretation remain involved. (Zylinska, 2023, pp. 119–144)

“It is also always someone’s vision.”

It is always the view of a particular subject. (Zylinska, 2023, p. 129; translation by the author of this article.)

This prompts me to consider whether unexpected options offered by AI still have value. I believe that they can enter the exploration space actively initiated by the user, allowing the user to decide whether to keep or discard them. They may not meet current requirements and therefore cannot be considered a better, formal answer; however, the user might also discover: “This isn’t what I was originally looking for, but I’m willing to keep it.”

I understand this encounter as a kind of beauty in unpredictable interaction. The system presents an unexpected element, and the user decides whether it is meaningful to them. However, keeping it might simply be for appreciation, curiosity, or preparation for another scenario, and cannot be directly interpreted as a change in long-term preferences. Users should retain the freedom to simply try and not adopt, and there should be steps that allow users to quickly exit or retain the experience at any point after exploring it further.

From the Right to Define to the Right to Participate in Revision

This reading helped me sharpen my initial question: when an AI system constructs a profile of me, can I understand the assumptions behind it, correct what does not apply, and guide it toward the person I wish to become?

In my research, I want to examine how users participate in setting goals, reviewing information about themselves, revising preferences, and deciding what to do with unexpected suggestions. Whether these opportunities are understandable, usable, and consequential in practice remains an empirical question.

Zylinska, J. (2023). The perception machine: Our photographic future between the eye and AI. MIT Press. https://doi.org/10.7551/mitpress/14471.001.0001

Chapters discussed

All page references refer to the book’s printed pagination. The discussions of AIFFD, women’s self-presentation, user goals, and responsibility in recommendation systems are my research reflections developed through this reading. They should not be understood as design findings established by Zylinska.