40.2 Ethics of computational photography⧉
Computation does not introduce deception into photography; it widens a gap that was always there. Photography has never been "the truth." The moment you choose a frame, an exposure, and an instant, and choose to keep this shot and discard a hundred others, the image becomes a set of choices rather than a neutral record. Computation simply gives those choices more reach.
- Manipulation and provenance. There is a continuum from a routine retouch, through composites, to outright deepfakes, and no bright line marks where enhancement becomes deception. Two technical responses matter: forensics, which looks for evidence of alteration in image statistics and physics (developed in the Image Forensics chapter), and provenance, which can carry signed claims about origin and editing (C2PA / Content Credentials). Watermarks can help identify content or reconnect it to a record, but are not themselves a signed edit history.
- Fairness and representation. This is not hypothetical. Color film and, later, automatic exposure and auto white balance were tuned for light skin; the calibration reference was literally a photo of a light-skinned woman, the Shirley card, and unequal performance across skin tones remains a concern in metering, white balance, and face detection. The obligation is to design and test for diverse skin tones (the Skin tones chapter), not to inherit the defaults. The corrective is now beginning to ship inside the ISP itself: Google's Real Tone (Google "Real Tone") retunes the stages that carried the bias, including auto white balance, exposure, and tone rendering, using a deliberately wide and diverse set of skin tones, developed with image-makers across many ethnicities, to improve darker-skin rendering. A named feature is not enough. The fix has to live in the defaults and the test set, not in a toggle, but this is the right kind of response: repair the computation that encoded the bias.
- Privacy and consent. Faces, surveillance footage, always-on wearable cameras, automatic face and license-plate recognition all raise the question of the right not to be captured, and of who gets to point a computational camera at whom.
- Generative AI. When any image can be generated from a prompt, the questions sharpen: consent and copyright for the training data, what a synthetic "photograph" of an event that never happened even is, and what photographic evidence can still mean (the Generative AI chapter).
A useful principle for practitioners is to name each capability and its misuse in the same breath, and prefer transparency. Disclose synthesis and substantive alterations when they affect what viewers would reasonably take the image to document; routine computational processing is part of nearly every digital photograph. The goal is neither optimism nor doom; it is a frank account of the tool and its consequences.