Fairy Note 3
From Operational Images to Human Recognition
Paglen, Barthes, classification and the changing human relationship to images
28 September 2026
Publications discussed in this note
This Fairy Note responds primarily to two recent publications that extend the questions developed in my earlier review of AI aesthetics and image theory.
Paglen, T. (2026). How to See Like a Machine: Images After AI. London: Verso. Published 19 May 2026.
Trevor Paglen’s new book examines the transformation of visual culture brought about by computer vision, machine learning and generative AI. Paglen argues that contemporary images increasingly need to be understood not simply through what they represent to human viewers but through what they do within computational and social systems. Images can classify, predict, identify, activate and intervene. The human viewer is therefore no longer necessarily the centre of visual culture.
Aleksandruk, I., Palchevska, O. and Androshchuck, A. (2026). ‘Artificial Intelligence and the Death of the Author: Originality, Creativity and Generative Art Ontology’. Dia-Noesis: A Journal of Philosophy, 20, pp. 10–34. Published 18 September 2026.
Aleksandruk, Palchevska and Androshchuck reconsider Roland Barthes’s concept of the death of the author in relation to generative AI. They argue that creativity increasingly needs to be understood as distributed across humans, technologies, institutions and cultural contexts rather than located exclusively in the sovereign individual author.
Abstract overview
Read together, these publications describe two related displacements within computational image culture.
Paglen destabilises the assumption of the sovereign human viewer. Images increasingly circulate between computational systems and perform operations without necessarily being contemplated by a person.
Aleksandruk, Palchevska and Androshchuck destabilise the sovereign human author. Generative images emerge through relationships between human intention, language, cultural archives, training data and computational systems.
This creates an important problem for my own research. If computational visual culture is displacing the human at both the point of production and the point of reception, where does specifically human agency remain?
For Phantom Mirror, I propose that one possible answer lies in recognition. The machine can generate, classify, organise and predict possibilities, but the human encounter determines when an image becomes affectively significant.
This develops the argument of the earlier Fairy Notes from AI as a mediated latent medium towards a more specific question concerning the relationship between operation, classification, ambiguity and human recognition.
1. From representation to operation
Paglen’s work asks us to reconsider a fundamental assumption of traditional image theory: that images are principally made by humans to be looked at by other humans.
Increasingly, images exist within computational systems in which machines produce, analyse and act upon visual information. An image may identify an object, measure a change, classify a person, detect a fault or trigger another process.
The critical question therefore shifts from:
What does this image represent?
towards:
What does this image do?
This develops the idea of the operational image, historically associated particularly with Harun Farocki and subsequently developed by artists and theorists including Paglen.
An operational image does not simply depict a system. It participates in its operation.
This is particularly relevant to industrial technologies. Digital twins, sensor displays, process diagrams, predictive visualisations and control interfaces do not merely show infrastructure. They make complex and otherwise invisible systems perceptible, interpretable and actionable.
The image becomes an instrument of thought and action.
2. Representation, operation and activation
For my research it is useful to distinguish three functions.
A representational image asks:
What does this image show?
An operational image asks:
What does this image enable someone or something to do?
An activational image asks:
What does this image cause to happen within a system?
These categories can overlap. A single industrial visualisation may represent a system to a human operator while simultaneously participating in computational monitoring and triggering action.
This is important for my work with industrial technology because it moves the discussion beyond simply making images of technology.
The more interesting subject is the relationship between invisible technological processes, their computational representation and human understanding.
3. Classification versus recognition
This produces an important distinction for Phantom Mirror.
Machine vision frequently depends upon classification. Ambiguous sensory information must be converted into categories that permit subsequent action.
Classification effectively asks:
What is this?
My use of recognition in AI phototherapy means something different.
Recognition is the moment when an image becomes affectively meaningful:
This is what this feels like.
Or:
I recognise something of my experience here.
The distinction might therefore be expressed simply:
Classification asks: “What is this?”
Recognition asks: “What is this like?”
The first seeks categorical resolution. The second may actually depend upon ambiguity.
This gives refusal another important function within AI phototherapy.
When a generated image is plausible, beautiful or coherent but emotionally wrong, I refuse it. I do not allow computational fluency to resolve the inquiry prematurely.
Refusal protects ambiguity from premature resolution.
The inquiry remains open until recognition occurs.
4. The displacement of the viewer
Machine vision also complicates another assumption of traditional image theory: that an image requires a human viewer.
Large numbers of contemporary technical images are generated for machines, processed by machines and passed between machines. Human beings may never see them.
Traditional image theory often assumes:
human maker → image → human viewer
Computational visual culture increasingly includes:
human / machine / dataset → computational image → machine interpretation → action
The human viewer is no longer necessarily at the centre.
This is particularly interesting in relation to Phantom Mirror, because my methodology deliberately returns human affective encounter to the centre of the process:
lived experience → dialogue → machine generation → image → human encounter → refusal / acceptance → regeneration → recognition
The machine produces possibilities, but it cannot determine which image becomes personally significant.
5. Barthes and the displacement of the author
A parallel development occurs at the other end of the image.
Roland Barthes famously challenged the idea that the author provides the final authority over the meaning of a text. Aleksandruk, Palchevska and Androshchuck revisit the death of the author through generative AI, arguing that creative production increasingly appears distributed between humans, computational systems, institutions and cultural contexts.
Generative images emerge through relationships between:
human intention, language, training data, cultural archives, model architecture, platform design, generation and selection.
Authorship therefore becomes distributed.
However, this does not necessarily mean that the artist disappears.
My own practice suggests that the functions traditionally grouped together under “authorship” may instead become separated.
The AI system generates possibilities.
I navigate, respond, refuse, edit and recognise.
The artist’s agency therefore moves away from complete control over the production of the image and towards judgement within an unfolding process.
This is why navigator continues to feel more accurate to me than “prompt engineer”.
6. Two simultaneous displacements
Putting these two publications together reveals something particularly interesting.
Barthes and the contemporary AI-authorship debate destabilise the sovereign human author.
Paglen and machine vision destabilise the sovereign human viewer.
The familiar structure:
human author → image → human viewer
is therefore being destabilised at both ends.
The contemporary computational image can be generated through distributed systems and encountered by other computational systems.
This raises an important question:
Where does the human remain within computational image culture?
For Phantom Mirror, one possible answer is:
in recognition.
The human contribution is not simply the production of an image. It is the embodied, biographical and affective judgement through which one possibility becomes meaningful.
7. Connection to the wider AI aesthetics literature
This extends rather than replaces the framework developed in the earlier Fairy Notes.
Norouzi & Prinz: AI as a mediated latent medium and hypermnesia.
Declos: the medium contains aesthetic biases.
Toister & Zylinska: images participate in thinking.
Musih: language becomes an interface through which images are generated.
Markelj & Kemper: computational images increasingly possess operational aesthetics.
Oliveira: human artistic practice retains embodied judgement through acceptance, transformation and rejection.
Paglen: computational images increasingly act, classify and activate rather than merely represent.
Aleksandruk, Palchevska & Androshchuck: authorship becomes increasingly distributed across human and technological systems.
My developing contribution through AI phototherapy is concerned with what happens when these computational possibilities return to human affective experience.
The resulting structure is becoming:
latent possibility → computational generation → image → human encounter → refusal → navigation → recognition
8. Relevance to Phantom Mirror
This strengthens rather than diminishes the importance of emotion within the project.
The model does not need to experience grief. It does not need autobiographical memory. It does not need to understand dementia phenomenologically.
Its function is to generate possibilities within a computational image space.
The human brings something categorically different: embodied experience, biography, vulnerability, memory and affective judgement.
This is why my phrase:
“I start with emotion and end with emotion.”
continues to matter.
What occurs between those two emotional states is increasingly explicable through contemporary AI aesthetics and image theory.
9. Relevance to industrial technology
The same framework also provides a bridge into my work with AVEVA.
Industrial software makes invisible processes available through representations: pressure, flow, heat, vibration, prediction, risk, failure and system relationships.
These images may be operational. They allow humans and machines to perceive and act upon processes that cannot otherwise be directly seen.
My interest is therefore not simply in illustrating industrial technology.
It is in asking:
How do computational representations make invisible systems thinkable?
And, beyond that:
At what point does an operational representation become meaningful to the human encountering it?
This moves the work towards the relationship between operation and recognition, rather than towards conventional images of technology.
Working proposition
The developing argument across these Fairy Notes might now be summarised as:
Computational images increasingly classify, predict, optimise and activate. AI phototherapy deliberately returns the image to human affective encounter. Through navigation and refusal, computational possibilities remain open until an image produces recognition.
Or, in its simplest form:
The machine generates possibilities. The human determines significance.
That distinction currently feels central both to Phantom Mirror and to the wider question of what remains human within increasingly computational forms of seeing.
Bibliography
Aleksandruk, I., Palchevska, O. and Androshchuck, A. (2026) ‘Artificial Intelligence and the Death of the Author: Originality, Creativity and Generative Art Ontology’, Dia-Noesis: A Journal of Philosophy, 20, pp. 10–34. Published 18 September 2026.
Barthes, R. (1977) ‘The Death of the Author’, in Image-Music-Text. Translated by S. Heath. London: Fontana.
Farocki, H. (2004) Eye/Machine III. Video installation.
Paglen, T. (2026) How to See Like a Machine: Images After AI. London: Verso. Published 19 May 2026.