AI-Generated Images?


How do we ought to decode the images generated by AI-based algorithms? In the absence of human agency, what meaning can these works have, if any? Can we just refer back to the input given by the user?

This is what Dall-e (by OpenAI) spits out for the input "English village":

Note: these are four examples from four different generations of images, as I wanted to get a sample reflecting possible different common views of the subject.

It is tempting to read these images as we read human-created ones, taking into account the context, the intention of the author, etc, but that cannot be right in this case. The AI "agent" is void of agency, paradoxically. Where is the meaning coming from then? Why does it output these images and not others? It is easy enough to think of scenes, buildings an objects that are not depicted in any of the examples above. They are clearly reflecting back to us a common view, an average view, of the subject, the "English rural village". Note, for example, that "rural" was not part of my query, but it is very much part of the output. I didn't specify either that I wanted centuries-old buildings, or a lack of people, and cars. I didn't ask for an "idillic" view of an English village, but that's what I got back, because that's also the most common view you get if you search for images on Google for this same subject.

AI generators are not generators at all: they are aggregators, they create nothing. They steal the ideas and images published by humans and give us a kind of summary of them all.

Taking that in mind, I argue that the most useful way of reading these AI-aggregated images is as the results of a poll or some other social research: they give us a glimpse of what the most common view of a subject is. In the examples above, we can see that the idillic, romantic notion of the English countryside is very much alive. We can analyse what is not being depicted: no traffic, no polluted rivers, no supermarkets and coffee chains, no people, no Land Rovers parked in front of pristine gardens, no food banks.

But there is a problem with this reading too: we know little to nothing about the agency of the AI owners. AI agents have no agency, but the companies behind them do have agency and agendas. What data is being fed to this program? From where and in which periods of time was the data collected? What data is NOT being used? What are the variables used in order to select "good" images (and to discard others)?

This is, I argue, the main reading we can make today of these aggregator-images: they are like outputs from a social research study, but a badly run study of which we know nothing about. No authors listed, no declaration of conflicts of interest, no clear goal defined beforehand, no details of the methods used, no analysis of shortcomings and strengths, ...

We have to take these images for what they are: outputs of weak social research.

Create your website for free!