antonella autuori Annotating Otherwise

01{a techno-determinism counter-programme}

Annotating Otherwise is a set of participatory formats for data annotation and small, situated machine vision models. It proposes a counter-programme to the technocratic agenda of efficiency, speed and invisibility that governs data work and quietly holds up the infrastructure of platform capitalism.

02{invisible by design, flattened by default}

Every machine learned to see through millions of tiny gestures: a box drawn around a face, a label assigned in three seconds, a click paid in fractions of a cent. This labour is invisible by design (Miceli et al., 2020, 2021), distributed across platforms and borders, and largely performed by people tasked with applying ready-made categories dictated by tech companies and policy agendas.

This regime of efficiency extends into algorithmic systems that impose an automated gaze upon the world, one that reduces its complexity to a finite set of categories.

Book spread: Who assigns these labels? COCO alone took over 70,000 worker hours; CelebA's 40 face attributes, usually decided by a few people in a room.

“Seeing comes before words,” wrote John Berger (1972). But machines learn to see through words: labels, categories, lists. If they now see on our behalf, are we still looking at the world, or only at the ways of seeing someone else decided for us?

03{a call for slowness and multiplicity}

Against this economy of reduction, Annotating Otherwise seeks to reclaim a space of pause and slowness in which to look at and perceive the world (Hallnäs & Redström, 2001; Strauss, 2026), resisting algorithmic systems that decide in advance what we are meant to see. If, as Betti Marenko (2025) argues, our categories urgently need to be rethought, annotation is where this work can begin: the small, overlooked gesture through which the world is sorted, named and made computable.

Annotated photos of the dataset, one after another, their masks arriving label by label: two quiet bodies, a small loneliness, salt in the air, the edge of leaving, the sea remembering…

What if we reclaimed its agency and its poetics, recognising, with Kalir and Garcia (2021), that annotation can be “many things, to many people, across many texts [and images] and many moments in time”?

04{situated, embodied, relational data practices}

Drawing on feminist epistemologies and feminist data practices, Annotating Otherwise approaches data annotation as situated, embodied and relational work, and as a form of critical design inquiry.

By returning and opening classification to “others”, those who are usually classified by the AI–state complex, and inviting them to engage playfully with their own labels, we might start to glimpse AI otherwise: through partiality, positionality and material relation, not as a seamless system of capture, but as a fragile arrangement of categories, hesitations and situated decisions.

05{towards a repertoire of otherwiseness}

Annotating Otherwise is an invitation to see from somewhere (Haraway, 1988), to care for differences (Puig de la Bellacasa, 2017), and to stay with the trouble (Haraway, 2016) of established ways of seeing and classifying the world.

Together, we can build a repertoire of otherwise-ness (Marenko, 2025): gestures, labels and stories that open toward plural imaginaries for technologies of classification, ones that trace relations rather than draw borders, and that do not pretend to stabilise the world but make room for its complexity.

06{something more to see}

Technologies that remember, as Merleau-Ponty writes, that it is essential “for the thing and for the world to be presented as ‘open’, to send us beyond their determinate manifestations, and to promise us always ‘something more to see’” (1945/2012, p. 348).

Book spread: How does AI learn to see? Chapter 1, a personal intro to my restlessness with AI's love of neat categories.

07{a starting provocation}

What kind of world would you teach a machine to see — or to unsee?

Teaching machines how to unsee

@Fablab SUPSI, Campus Mendrisio – Switzerland / November 2026 / 9 participants / 9 datasets / 288 annotated images

Surveillance cameras read traffic through categories decided elsewhere: car, truck, person. Every one of those boxes was drawn by someone, and every box drawn teaches the camera to see a little better. This workshop took annotation, the very gesture that feeds surveillance, and turned it against its purpose: not to sharpen the machine’s eye, but to unsettle it. Participants worked on the same Kaggle dataset, annotating it subjectively and reflexively, from somewhere: through their own way of seeing, and through everything that way of seeing carries with it.

Some used fear as a filter through which to see the world: dangerous, get out of the way. Some voiced the social inferences we silently attach to a make of car, guessing who sits behind the wheel: money bags, plain jane. Others listened to their own bodies as they worked, and hunger turned bonnets and bumpers into dishes: cheesecake, costine. None of this is noise. Each label answers to a precise position, and here annotation makes that position visible instead of correcting or erasing it.

Each participant then trained a small classifier on their own reading of the dataset and ran it on the same footage. Side by side, the models ceased to be instruments of detection and became self-portraits: a way of showing how one looks, and of answering the gazes of others. Together, they teach a machine to see as they see, with assumptions, moods and blind spots left in rather than cleaned out.

workshop process

The process: a pre-assembled dataset from Kaggle, shared by all; subjective annotation in Roboflow through instance semantic segmentation; a small ML model for each gaze; live testing on the same footage, where subjective assumptions become visible side by side and each model becomes a probe for interrogating one’s own way of seeing; model presentation as self-portrait; collective discussion, where models become mirrors for reflecting on one’s own labels; and back to annotation. reflecting on one’s own labels kagglevehicle dataset pre-assembleddataset one shared dataset, so thatdifferent ways of seeingcan be compared as theymove through it roboflowannotation tool subjectiveannotation instance semanticsegmentation: eachinstance named inone’s own terms small ML models× 9, one per gaze small, situated models,each trained ona single reading the model as probe:a means of interrogatingone’s own way of seeing sharedfootage live testingon the same footage subjective assumptions,made visibleside by side modelpresentation each model presentedas a self-portrait collectivediscussion who labelled what,and from which position? models as probes andmirrors for reflectingon one’s own labels The process: a pre-assembled dataset from Kaggle, shared by all; subjective annotation in Roboflow through instance semantic segmentation; a small ML model for each gaze; live testing on the same footage, where subjective assumptions become visible side by side and each model becomes a probe for interrogating one’s own way of seeing; model presentation as self-portrait; collective discussion, where models become mirrors for reflecting on one’s own labels; and back to annotation. reflecting on one’s own labels kagglevehicle dataset pre-assembleddataset roboflowannotation tool one shared dataset, so thatdifferent ways of seeingcan be compared as theymove through it subjectiveannotation small ML models× 9, one per gaze small, situated models,each trained ona single reading sharedfootage instance semanticsegmentation: eachinstance named inone’s own terms live testingon the same footage modelpresentation the model as probe:a means of interrogatingone’s own way of seeing subjective assumptions,made visibleside by side each model presentedas a self-portrait who labelled what,and from which position? collectivediscussion models as probes andmirrors for reflectingon one’s own labels

datasets

the trained models

I see food outside my window (Matteo)
I don't like highways (Alice)
Cool Cars? (Jennifer)
Slow-Fast (Giorgio)
Beware the sea trucks (Ayse Ceren)
Too emotional to be a car (Antonella)
Who is driving (Annabelle)