The Open Book

If you're opposed to AI image generators but use Google, you're a hypocrite.

I've been wanting to get this off my chest for a while. Many people are opposed to generative AI for a variety of reasons.  They're entitled to their opinions. But if their opposition is based on an assumption that AI basically copies existing artwork that it "scrapes" from online sources and then plagiarizes it to create facsimile images, your basis is wrong and your opinion is built on a faulty premise.

"Let's stop calling it 'Artificial Intelligence' and call it what it is: 'plagiarism software'," Noam Chomsky, who's never been afraid of going over the top with his opinions, wrote last year. "It's the largest theft of property since Native American lands by European settlers."  

I'm most assuredly not a technology expert. I'm a 70-year-old white guy in the deep south who still remembers the first time he used a Windows-based computer. But I can read and I can learn, and instead of reacting instinctively against what I don't understand, I spent a modicum of effort researching the technology.

To start at the basics, generative AI creates images based on work prompts. After I typed "The Open Book" (today's name in the Universal Solar Calendar) into DALLE-3, it produced the picture above. It can do this because it had previously scoured the internet to find pictures captioned "book" or with the word somewhere in the caption. Then using some incredible technology that simulates a neural network, it "learned" what a book is and isn't. It figured out other words for "book," like "atlas" and "volume" and "paperback." It learned about the many different styles and varieties of books.

It's not unlike teaching a toddler what a horse is.  "No, that's not a horse, that's a cow," "No, that's a sheep," and so on.  Eventually, they reach a level of understanding where they can recognize horses and distinguish them from other animals without being told.  

AI basically learns the same way. What it doesn't do is copy an image of a horse or a book and then later paste it into a requested image and "photoshop" it all together into one seamless picture. What it DOES do is create an original picture of a book or a horse or what have you based on what it learned the prompt looks like, and based on what it "thinks" the prompter is requesting. It then, before the final image is produced, goes through a series of adjustments and corrections to fix discrepancies between the prompt and the draft image, and finally produces what it "believes" is the requested image.

Most objections I've heard to generative AI concern the learning process, or "image scraping." If it learns from an image without the express consent of the artist, they consider it equivalent to plagiarism. That's ridiculous. Does the toddler need the express permission of the farmer to learn what a horse looks like? Do they need the consent of the artist if they learn from an illustration? 

Also, and this needs to be emphasized, "image scraping" is just one category of data scraping or web scrapping, which happens all the time. How do you think Google and other search engines know what web sites are addressing the topic about which you're asking? The search engines engage in almost continuous web scrapping and cataloging what they learn, and provide links to words that match those in your query. Google scrapes this data without explicit permission of the website owners (it could be argued that posting something online is implicit permission for web scraping).

I once read an artist complain that they were looking at an AI image and recognized a lamp in the background that they claim was copied from a painting they produced. The artist claimed that there were details to the lamp, small things most people wouldn't notice, that convinced them that it was "their" lamp, cut and pasted into another picture without their permission. 

Bullshit.  Frankly, the story sounds apocryphal to me, but let's look at the claim for a minute. One "green tiffany lamp on a desk" very well could resemble another, but that doesn't mean the author has copyright over all depictions of green tiffany lamps. What's more, how did the artist learn what a "lamp" looks like?  Did the artist look at other pictures of lamps, and if so, does the artist owe those previous artists royalties or compensation for their lamps? Also, someone, some nameless engineer or technician, designed the first actual green tiffany lamp, and all those artists are actually just copying the intellectual property of that original lamp designer. I know this argument is getting ridiculous, but that's my point. Just because someone thinks a lamp - or a horse or a book - looks like some other lamp, horse, or book, doesn't constitute copyright infringement.

But to be fair, let's look a little broader. Not only is there the matter of object depiction, there's the question of style.  Surrealism, day-glo colors, pastel washes, perspective, lighting - all these elements were "scraped" (i.e., observed) by AI programs. Aren't someone's rights being infringed?

As I said earlier, I'm not a tech expert, and I'm also not an attorney.  But as I understand it, the courts have repeatedly ruled that "style" can't be copyrighted.  Did you know there's such a thing as "art schools," where fledgling  artists are shown works of other, previous artists - without those other artists' permission - and encouraged to "reproduce" (i.e., "copy") their style and "learn" (i.e., "scrape") from previous artworks? 

There's much, much more to this subject, and I'm not saying there aren't unethical types of web scraping (e.g., identity theft) and image scraping (I'm uncomfortable with facial recognition software). There's also the inevitable issues that arise whenever someone's means of livelihood is challenged by a new technology. And the biggest issue is probably the inevitable rise of "deep fake" images and use of AI imagery in propaganda.    

Final point - an analogy to music.  There are only so many notes on a piano and some compositions are inevitably going to sound similar to others. Musicians learn what sounds good or interesting from previously composed music. However, if I were to copy a previous composition note-for-note and call it my own, what's plagiarism. That's copyright infringement. But if I were to use some chord changes and sequences previously used by others, but in a new composition that no one would reasonably confuse with a prior work, that's not plagiarism.  Even if I were to use recordings of other music, but manipulated or spliced in such a way that it's a new work, that's "fair use," not copyright infringement.

If generative AI were used to create an exact copy of another image - a pixel-perfect reproduction of an Ansel Addams photograph, say - that's clearly plagiarism. If it creates a black-and-white image of a Yosemite landscape with cloudless skies in a style similar to Addams' work, even after having observed his work and learning from his style - that's original. It's original even if someone might later confuse it for an Addams' photograph (as long as it isn't fraudulently promoted as such).

These waters are much, much deeper than these few random thoughts, but I think I've already said enough for one day.