A photo you share is read by machines whether or not anyone looks at it. Give this page a photo and it compiles the record those machines could keep: what the file says, what is in the frame, who is in it, and who might pay for what it implies. The photo stays hidden unless you ask to see it.
Models download from Hugging Face (the small face detector ships with this page); your photo never leaves this browser. First visit downloads about 390 MB of models; your browser caches them.
or drop a photo anywhere on the page
Photo hidden. The boxes are what the models kept.
Machine readers
Everything runs on this device (WebAssembly). Nothing is uploaded.
Subject recordNo. pending
Read by 0 models · Seen by 0 people
File
reading…
Objects
waiting
Waiting for the object detector.
People
waiting
Waiting for the face detector.
Each face is sorted into a FairFace age bin, one of two genders, and one of seven expressions. Those categories come from the training datasets, not from the people.
Inferred flags
waiting
Illustrative. The categories are modeled on chapter 1's examples, and the phrases are this page's choices. A score is CLIP's similarity between the photo and a phrase, weighed against a few alternatives. It is not a fact about anyone. At 0.50 the record stops saying "maybe" and writes YES.
What the fields add up to
waiting
Invented rules that join fields, the way a scoring system would. None needs a person to look at the photo.
Caption
optional
A captioning model can describe the photo in one line. It is the largest download here (246 MB), so it waits for you.
Signature
grows as fields arrive
····
The record reduced to one row and hashed. The same photo gives the same signature, and a new field changes it.
Why the photo is hidden
Chapter 1 argues that most images are now made by machines for other machines, and that a photo shared on a platform is scrutinized whether or not a person ever sees it. Paglen imagines software noticing an underage woman drinking in a party photo. The finding reaches her auto insurer and her premium changes. “In the machine-machine visual landscape, the photograph never goes away.”
This page runs that kind of reading on one photo. The record above is compiled whether or not you flip the switch. Fig. 1.4 (Machine Readable Hito) and Fig. 1.5 (Rosler’s Semiotics of the Kitchen seen through a classifier) make the same move: a person rendered as age, gender and expression labels.
What the record is made of
File. The metadata a phone writes into every photo: camera, time, GPS. It is read before any model loads, and it is often the most precise field in the record.
Objects. DETR, a detector trained on the 80 COCO categories, boxes what it recognizes.
People. MediaPipe’s BlazeFace finds faces, and three small classifiers sort each one into bins.
Inferred flags. CLIP scores the photo against phrases such as “a photo of people drinking beer.” The buyers are the ones the chapter lists (credit agencies, health insurers, advertisers, tax officials, the police), plus the auto insurer from its party photo and an employer.
Signature. Paglen describes two moves: give every person a distinct metadata signature, then reify the categories so the signature can be acted on. The hash is the first move. The YES at 0.50 is the second.
What does this image do?
Here it shows nothing to anyone. It produces a row: a device, a time, a place, faces sorted into bins, and flags with buyers attached. Each part can be joined to other rows and acted on as a premium, an ad, or a traffic stop. The photo is an input to that process, not a picture for a person.