For whatever reason, I tasked myself with writing an opinion piece(s) about AIs, despite the hundreds of sources already covering AI-related stories. Sources better qualified than me.
. . . but sources that are not me, so here goes nothing. Oh, I’ve recently been reminded of the virtue of conciseness, but, you know, I’m still me, and there are many aspects to this discussion.
Therefore, I’m doing multiple posts so I don’t tire my readers. This post, then, is Part 1, covering the background stuff in about 1,900 words, or roughly 7 minutes of reading.
Right — here we go.
- Background: Copyright
- Background: How do AI Art Generators work?
- The Database (images)
- The AI Art Generators
Let me start by saying I’m protective of my photos and writing. I’m not fond of entities reblogging my stuff. Likewise, I’m not fond of entities taking my photos and using them as their own.
Note: I say entities because the issue isn’t just people. Bots and Companies do it too.
At the same time, I’m aware there’s little I can do about it. In this digital age, the moment something is out there, it’s open to being copied and shared. Worse, companies actively encourage people to steal from each other even as they fiercely defend their copyright claims.
For example, Pinterest encourages and facilitates people to take other people’s photos and arrange them in albums under their accounts. As a result, there are thousands (millions?) of Pinterest accounts with nothing but other people’s graphics on them.
Even as they facilitate the practice, Pinterest absolves itself by saying they take “copyright infringement very seriously,” but that responsibility for infringement rests with the users. So far, the courts have agreed with them.
Are users liable? Legally, yes. Every created work is automatically copyrighted, with the creator being the copyright owner. They don’t have to say or display it; it just is. So if you create something, you own the copyright to it, and if someone copies it without permission, they violate copyright law.
There are nuances to this, like, for instance, fair use and research, but those qualifiers are nebulous, and most Internet sharing — including my propensity to post cartoons — doesn’t qualify.
Despite what the law says, practically, most individual users incur little risk when they copy and share something, especially if they’re not doing it for monetary gain (they don’t profit from it).
If you find your content on another site, the first step is to send a takedown request. If the site owner doesn’t comply, a recourse for copyright holders is to sue for damages, which means they have to prove damages (for example, loss of income).
In my case, and for most people, the cost of lawyers outweighs the compensation of damages. Of course, that might be different if my photos and doodles were selling for multiple thousands of dollars.
NOTE: because of the above, I will modify the comics I share to include only those I have permission to do so from the originator.
Background: How do AI Art Generators work?
I did a fair amount of reading on this subject, and I link one of the “clearest” explanations below. But, in general, it can be roughly (if slightly erroneously) explained by the concept of reverse engineering.
“What is reverse engineering?” Well, it’s what you agree NOT to do with every user agreement you blindly accept when you use any software.
“OK, but what is it?” Well, it’s what many companies do. Most famously, Chinese companies, although they aren’t the only ones.
“OK, now you’re pissing me off.“
Let me explain with a couple of examples.
I worked for GM in the dark ages of my life. In the 80s, after I had escaped and worked as a contractor for them, I still had access to a building near their Warren Tech Center dedicated to the Mona Lisa. No, not the painting.
Inside this building were rows and rows of competitors’ cars. Each model showed two versions of each vehicle; one complete as-purchased version and one completely disassembled version laid out on tables, every separate component that made up that car arranged in the order that it was removed.
So, I could walk in, examine the finished product, and then walk along the tables to see every component disassembled. At the end of the table, you typically had the welded bare body. Nothing was torn apart, but if two pieces were originally bolted together, they would be shown separately, along with the bolts that connected them.
The purpose was to “teach” designers and engineers how other companies create their products, the decisions and compromises competitors made, and their choice of materials.
The stated aim was to take “the best” ideas and incorporate them into the designs of GM cars and trucks.
Later on, I’ll speak to the limitations of this effort, but for now, here’s another example.
You go to a restaurant, order something from the menu, and it’s delicious. You ask the restaurant if they would share the recipe, but they say it’s a trade secret . . . so you write the Food Network, and they endeavor to duplicate the recipe for you (LINK). It won’t be an exact copy, but pretty close.
If you have a good palate and can identify the various flavors, you could do the same at home. Even if you don’t have a good palate, you could do it by trial and error.
In both cases, you try to figure out how a finished product was made from scratch and do so without having the original plans or original recipe.
The AI is “trained” in nearly (but not quite) the same way.
The video in THIS LINK explains it with a simple visual example, but let me try to explain it using the reverse of an old joke about how you create a sculpture of a bear.
You show the AI a sculpture of a bear. Then, you keep adding bits of marble to the sculpture until, after a while, you have a block of marble. Meanwhile, every time you add a piece of marble, the AI takes a snapshot.
Then, in the end, you show the AI a block of marble, ask the AI to reverse the process, and you end up with the joke:
“How does an AI make a sculpture of a bear?”
“It starts with a block of marble and chips away all the pieces that don’t look like a bear.”
In this case, it’s no longer a joke; it’s now a process.
Now, if you only show the AI one sculpture of a bear, you’ll always end up with the same bear (ignoring any error you could purposefully introduce with a random noise function).
But, if you show the AI one million bear sculptures in various poses, you could then ask it to make a sculpture of a bear running, sitting, sleeping, climbing a tree, or anything you can imagine a bear doing. The AI would then begin to chip away at the marble as it compares the result to its database. Eventually, it would come up with something resembling what you described. The accuracy of the sculpture depends on how closely items on its database match what you describe. It’s also probable that if the database has a thousand images of a bear, the sculpture it produces might have features that aggregate two or more such images and not an exact match to any.
Videos about AI image Generators
The Database (images)
This is where things get a bit tricky. Again, I’ll provide links to more in-depth articles, but I’ll keep this simple.
From what I understand, much money was invested into creating a database to speed up training software programs in pattern recognition. You might have heard the term ‘machine learning,’ and that’s essentially how the database is used.
The database contains billions of images that are “scrubbed” from the Internet. (Note: I refer to ‘database’ as singular, but there are many databases, some targeting only specific content and some broader in scope.)
My entire WP Media Library is likely included in a database since they have appeared on my blog.
“My, you sure have a big ego to think that!“
Nope! I mean, yes, I have a big ego, but that’s not why I say that.
Bots scour the Internet for images and collect any they encounter. They don’t care (or know) if they come from the popular blog of a popular content generator or if they are from the blog of an annoying old guy with an overinflated opinion of himself.
NOTE: this might account for the disproportionate number of views I get relative to the number of visitors. It’s not unusual to have hundreds of views on a post that only gets a few visitors. Of course, I could be wrong, and it’s just speculation on my part, but it makes more sense than the explanation I got from WP.
Given the above discussion of copyright, you might think this would be illegal.
It would be, but the companies creating the database are non-profit, involved in machine learning research and development, and are ostensively doing this to help advance the field of AI. ‘Non-profit’ and ‘research’ are the two shields they use to ward off legal challenges.
Again, I stress that I’m oversimplifying. If interested, the links below go into more depth, but think of it as using the images in this database to train a program to identify, catalog, and classify said images.
Why? Why do that?
Because it’s generally accepted wisdom that there’s a benefit to developing AIs able to mimic the recognition processes humans employ.
Note: If interested, there is lots of digital ink out there as to the potential — real and imagined — benefits of having AIs capable of recognizing, organizing, and using large amounts of data. Here, I’m only addressing the use of AI Art Generation software.
So, now you have this database of billions of images . . . what can you do with it?
https://www.nytimes.com/2023/01/12/technology/microsoft-openai-chatgpt.html (may be behind paywall)
Microsoft to invest $10 billion in ChatGPT creator OpenAI, report says (cnbc.com)
Machine learning – Wikipedia
OpenAI – Wikipedia
Data mining – Wikipedia
The AI Art Generators
This is where MidJourney, NightCafe, DALL-E, and Dream AI come in. Those are the ones I’m familiar with, but there are others.
Those programs use the database to ‘create’ art based on you, the user, asking for something.
“AI, please show me a picture of a bear flying a kite!“
“Sure thing, Disperser. Let me look at my database and remove everything that doesn’t look like a bear flying a kite!“
Again, I’m simplifying, but that’s essentially what the AI program does. Pedantic readers will complain it’s a bit different than that, but for the purpose of this discussion, that description is close to the mark.
From the database, the AI already ‘knows’ what bears and kites look like, and the AI program has a definition for the word ‘flying’. It can contextualize the word flying in relation to the bear and the kite, and build an image based on its understanding of what I asked it to do.
The likelihood of the database containing an actual image of a bear flying a kite is virtually zero.
But it contains images of bears and images of kites, and thus the AI program will give you a picture of a bear and a picture of a kite.
The picture will be unique but based on what it has learned about bears and kites, and will present them in approximate relationship to each other so that one could charitably assume it’s a picture of a bear flying a kite.
It could be a cartoon (DALL-E)
Or incorporate more realistic versions of the bear and the kite (MidJourney)
So far, so good, and people might wonder, “What’s the problem? It sounds neat!“
I’ll cover that in the next post. Meanwhile . . .
That’s it. This post has ended . . . except for the stuff below.
Note: if you are not reading this blog post at DisperserTracks.com, know that it’s copied without permission and likely is being used by someone with nefarious intentions, like attracting you to a malware-infested website. Could be they also torture small mammals.
Note 2: it’s perfectly OK to share a link that points back here.