Page 94 - Libro 2

ARTICLES ARTICLES

Authentication of Photographs vs Generative AI Images

By Sheldon Bowles, FCAPA

Generative AI can be used to create highly realistic synthetic images, which can present
challenges for photographic organizations who need to distinguish these images from real
photographs. These challenges are prevalent as AI generated images have received awards at
several photographic exhibitions.
AI generated images can be classified as either (a) text-to-image, which are images that have
been created algorithmically from text prompts without any original photographic element, or (b)
artificial intelligence renderings, which are images that have been created or modified using specific
AI generative image in-painting or image out-painting techniques. Image in-painting is when AI
has been used to remove elements or to fill in generated elements that are not captured by the
photographer and are drawn from the AI generated system’s dataset of images scraped from the
internet. Image out-painting is when AI, using generated elements, extend the image beyond the
Slow return of Sandhill Cranes to the night roosting flooded marshes
original image’s boundaries.
Given the rapid advancement in AI generative technology, and the possibility that AI generated
images might be entered in photographic exhibitions, the competition committee for Canada’s
national photographic association (CAPA) felt that it would be helpful to put together a framework
that would help judges and organizations distinguish between AI generated images and authentic
photographic based images.

The framework that I have put together is the result of seven months of research and testing
that was conducted using a diverse array of techniques to identify AI generated images. The results
of these tests are presented in this article. This article offers guidance on detecting AI-generated
images through two methods: Subjective Image Assessment and Objective Technical Analysis.

Subjective Image Assessment relies on human judgement to visually identify AI images. This initial
assessment can be helpful as humans usually excel at identifying errors or logical inconsistencies in
images due to their innate ability to recognize visual patterns. However, as technology progresses,
relying on subjective image analysis to identify AI clues will become progressively more difficult.
Additionally, the performance of humans is quite variable and depends on personal experience
and diligence, and therefore cannot be relied upon in many instances. Nevertheless, recognizing
Late afternoon return of Sandhill Cranes to the night roosting place inconsistencies in an image provides a mechanism to identify images that could have possibly been
Articles AI generated. These inconsistencies include: Articles

1. The image appears excessively spectacular and lacking natural imperfections.
2. The image appears unrealistic.

Unrealistic and Distorted Face

At sunset Sandhill Cranes fly to flooded fields for the night

94 95