# Depiction and Synthetic Images  ## Metadata - Author: [[170]] - Full Title: Depiction and Synthetic Images - Category: #articles - Summary: insert summary - My notes: - Document Tags: [[refereeing]] - Summary: Philosophers struggle to explain how synthetic images depict objects because these images often lack the maker's clear intention. Existing theories like intentionalism, non-intentionalism, and conventionalism all face serious problems when applied to synthetic images. This challenge suggests that synthetic images may require a new, special theory of depiction or might not depict anything at all. - URL: https://readwise.io/reader/document_raw_content/348110408 ## LLM Chats ## NotebookLM ## LLM Audio ## Highlights > Depiction and Synthetic Images ([View Highlight](https://read.readwise.io/read/01k4hwvach7jxqb6db74g6jb4m)) > what goes on ([View Highlight](https://read.readwise.io/read/01k4fm34bwpe3dmnqrc9hxs6q4)) - Note: not the best phrase > Many of these philosophers hold that one necessary condition for markings on a surface to depict an object is that the maker of the markings must have held some relevant intention toward that object being depicted. ([View Highlight](https://read.readwise.io/read/01k4hq2tq9rc375drc1j0dx076)) > 2. The Intentional Standard of C ([View Highlight](https://read.readwise.io/read/01k4hqwhcg57yb5vqabg412ctc)) - Note: could this section be cut down a bit? > Francesco Pierini ([View Highlight](https://read.readwise.io/read/01k4hqzrvf7a5jd3779j8pjh24)) > The markings in the image depict a philosopher, which is, of course, what I intended. What is the issue? > The issue is that the picture features objects that resemble more things than merely a philosopher. ([View Highlight](https://read.readwise.io/read/01k4hr1xr35dhy01vk4ecee0s3)) - Note: cool > His solution proceeds as follows. The extras found in an AI-generated image, while not specified in the maker’s prompt, are nevertheless a result of the maker’s intention. ([View Highlight](https://read.readwise.io/read/01k4hr4sbc9cjpcjscnngndhe0)) > Unlike John, I did not intend to delegate control over the objects in the image to ChatGPT. In fact, what I intended to create was an image of a philosopher on a plain, monochromatic background. My reason for not specifying this in the prompt was simply because I did not anticipate that ChatGPT would fill in the background details as it did. ([View Highlight](https://read.readwise.io/read/01k4hr6kmdf2pqhpcrj0cseb77)) > If we try to handle my case like Pierini does for the case of John, we would have to say that I intended to delegate control over the appearance of globes, hourglasses, quills, etc., to ChatGPT. This is implausible. Sometimes, makers of synthetic images might exclude certain details from a prompt because they intend to leave it up to the generator whether to include some features. It does not follow, of course, that everything left out of an image-generation prompt is intended to be left up to the generator. This is evidenced by the fact that synthetic images often require post-generation revisions to correct for features that the generator included that the maker did not intend to delegate to the generator ([View Highlight](https://read.readwise.io/read/01k4hr7ehf7cbntm0tng6gcmqb)) > y DALL-E. ([View Highlight](https://read.readwise.io/read/01k4hr7xym4netpz1hv3ta7ych)) - Note: which version? > One might suggest that, although I had no intention to depict candles while creating Figure 3, I nevertheless had the intention that the viewer would experience the markings as candles when I chose to display them. ([View Highlight](https://read.readwise.io/read/01k4hrp0ajg5h32arn06xj4xcx)) > The standard response to deviant causal chains is to insist that the cause brings about the effect “in the right way,” which, of course, excludes cases of deviance. Naturally, the challenge then becomes explaining what counts as causation “in the right way.” One might worry that, in the case of causal accounts of depiction, the answer ends up being “causation via the maker’s intention,” which brings us back to the intentional standard of correctness. It is the fact that my painting resembling the vase would be an accident that makes it fail to depict the vase. The most paradigmatic way for something to be caused accidentally is for it to be caused unintentionally. > To avoid this return to intentionalism, we need a causal view of depiction that gets us the right results for synthetic images but avoids deviant causal chains without appealing to causation via the maker’s intention. While I do not wish to claim such a view is impossible, I see no obvious way of formulating it as things stand. ([View Highlight](https://read.readwise.io/read/01k4hs3wx7w778z55taq7pz4fr)) > Lopes (1996) further refined this view, emphasizing that depiction requires mastery of a socially taught skill: the ability to interpret markings according to established representational practices. ([View Highlight](https://read.readwise.io/read/01k4hs4hkrrr4sjbx3t0k31hv7)) > The problem with counterfactual hypothetical intentionalism is that it does not distinguish between makers like DALL-E and, for example, Putnam’s ant. An ideal audience would presumably conclude that, if Putnam’s ant had intentions, it would have intended to depict Winston Churchill. And so, the markings in the sand depict Winston Churchill. We have arrived at what many consider to be an unacceptable result. But it is unclear why an ideal audience can ascribe counterfactual intentions to DALL-E in cases like Figure 3 but not to Putnam’s ant.5 ([View Highlight](https://read.readwise.io/read/01k4hshjvcg5ecz2bcyanax6tq)) > Recall Putnam’s lesson of the ant that traces an image in the sand that resembles Winston Churchill. Perhaps AI image generators are like Putnamian ants, and AI-generated images are like the traced markings in the sand that resemble Winston Churchill. In other words, perhaps the ([View Highlight](https://read.readwise.io/read/01k4hsm0x8yhprmk52aark85ps)) > correct takeaway from my arguments is that synthetic images do not depict anything. This would be no small bullet to bite. It would entail that all of the examples of synthetic images thus far presented in this paper are nothing more than pixels on a screen that do not actually have any real depictive content. ([View Highlight](https://read.readwise.io/read/01k4hskwbssj3ej7cmeqxan6gf)) > In short, without appeal to a maker’s intention, there seems to be no way of determining (not just epistemically, but metaphysically) what is depicted between candidates of isomorphic structures. But it would be absurd to say: “Figure 3 is a picture of books, faux books, a book wallpaper, book holograms…” So, if there is no difference-maker between these possibilities, the best explanation is that it depicts none of them. Once generalized, we have an argument for the view that synthetic images never depict anything at all. ([View Highlight](https://read.readwise.io/read/01k4hsp7by5msk0d32sk3xk5jp))