Lab Week 4: AI Colorization

Using AI to recolor this photograph of a Carleton student’s dorm room marked a special occasion: my first use of any AI tool beyond the rare Grammarly spell check. While I did approach the assignment with an open mind, I remained skeptical throughout that this was a worthwhile use of artificial intelligence. The process itself using DeOldify was quite simple: open the collaborative notebook, run the Python code blocks, and enter an image URL. The notebook will then produce a colorized version, with options to edit the various settings to increase or decrease color saturation. There is also an option where you can remove the automatic watermark that notes the image as being AI-altered. The creators of the technology state that use of the watermark is strongly encouraged, yet do not require its use, which I found to be an interesting choice. Below is a side-by-side comparison of the original image I used, a photograph of a Carleton dormitory in 1896, and the AI colorized version (which can be found on Omeka). 

After engaging with this tool myself, as well as interacting with various scholars arguing for and against similar uses of AI, I find I still have major ethical qualms that would bar me from using it in the future. One concern I have has to do with the perceived reliability and objectivity of the photographic medium. Because photographs require a subject grounded in reality, they are often seen as representing an objective and real truth that other mediums such as painting and drawing do not. These latter art forms are mediated by the human hand and mind, but photographs are mediated by a machine, lending them credibility. Of course, photographs can be manipulated as much as any other image, especially in the current digital age. However, the assumption of reality holds. People are much more likely to believe what they see in a photograph than what they see in a painting which purports to represent the same thing. Where AI becomes incredibly problematic is that it is altering photographs and reducing their “truthiness” without appearing to do so. As Ted Chiang writes in an anecdote about inaccurate photocopiers, 

“If the photocopier simply produced blurry printouts, everyone would know that they weren’t accurate reproductions of the originals. What led to problems was the fact that the photocopier was producing numbers that were readable but incorrect; it made the copies seem accurate when they weren’t.” 

Ted Chiang, “ChatGPT Is a Blurry JPEG of the Web”

The same logic applies here. Because it is not always obvious that AI has colorized these photographs, thus altering them (because the colorizing process is based only on probability: there is no way to confirm whether the colors added are accurate), the photographs appear to be accurate when they are not. While a photograph may seem innocuous, the problem becomes more menacing when colorization patterns result in lightening all skin tones, for example. And even when issues such as colorism and racism are not at play, the fact of the matter is that AI colorization creates convincing misrepresentations, rather than discovering true realities. In the words of art historian Sonja Drimmer, 

“But this effort to “bring events back to life” routinely mistakes representations for reality.”

Sonja Drimmer, “How AI is hijacking art history”

In the end, while I want to hold out hope that there will be a fully ethical use for AI tools, I have yet to be convinced.