A friend once asked me for my opinions on the use of artificial intelligence. I’ve been a big fan of A.I. for most of my life, and it’s been a popular theme in science fiction. But now, we actually have credible A.I. in our social and commercial environment, and it’s time to address the elephant in the room.
Artificial intelligence is a tool, nothing more. One does not condemn a table saw because it can present more teeth to the plywood panel faster than a human could. Such assertions that it is somehow inherently evil are misguided and disingenuous at best.
Generative AI does depend on having been trained by observing the works of artists, and a great many of them. This, however, is also true of human artists, and we do not consider this theft or misappropriation. Those who present this notion as viable apply a double standard. The same is true of the written word. Generative AI learns by observing the work of others. It’s not a copy-paste machine. It does not now, and never has, worked that way, and those who imply that it is somehow “stealing the works of others” clearly do not understand how either artificial intelligence nor human creativity work well enough to make an intelligent comparison.
The areas where generative AI shines are the technical ones, writing code that runs, diagnosing complex networking issues, constructing database applications that perform specific tasks. It can also do miraculous things, like protein folding, and speeding the discovery of new, previously unknown materials.
That said, one does not just lay the wood on the table and press the button, hoping for a replica Louix XIV divan to come out the other side. It’s just a tool. It requires a human being to make the decisions as to where to cut, and why. Artificial intelligence is mostly useless when it comes to creative acts, for it cannot create, except under the express direction of a human being.
Those who rely on generative AI for their writing simply by typing a quick command and pressing a button have removed themselves from the equation, and presenting the output of generative AI as their own without any material guidance, in my view, are charlatans and cheaters. The same is true of those who use pushbutton AI to make images, and then presenting that image as their own work. There is a role for AI in image generation, but deceptively passing it off as one’s own creative work is an unworthy occupation.
The U.S. Copyright Office has clarified its stand on the use of A.I. in creative works. They have said that if A.I. is used as a tool to create elements used in the finished composition, that one may copyright such a work. But if the finished piece was created without the guidance of a human hand, it cannot be copyrighted, for machines may not author anything directly.
I use artificial intelligence when creating graphics, but almost never to create entire images. Instead, I do things like remove people or objects from images, or add missing features. I also use A.I. when coding, because frankly, most of what I have to do is grunt work, and my guidance to the A.I. comes in very precisely defining the task so that I get exactly what I want. It’s like talking to a very literal minded child. I also use it in my writing, but never to create whole works, only to analyze and to help me organize what I have already written. I have tried, a few times, to have it write things for me, but the results are always mud-dumb, lackluster or outright wrong. And, A.I., no matter how hard you try to set up meta rules to combat this, tends to tell you whatever the hell it thinks you want to hear. This is not useful behavior in a creative, critical environment.
As I see it, the primary ethical concern with ChatGPT and its ilk is the abuse of the service from the standpoint of people trying to take shortcuts with it, or claiming its output as their own. This can range, therefore, from merely being sloppy and lacking in thoroughness to being outright unethical.
Tools, in essence, are tools. There is nothing inherently good or evil about them. It is in no one’s interest to anthropomorphize it and assign it moral or ethical behavior on its own. It is what we make of it, nothing more.
So, take what you want from this view, but remember that you can’t just declare something to be true in defiance of fact or objective proof otherwise. See the situation for what it is, and plan and react accordingly. No other approach makes sense.
– 30 –
Did I make the header image for this article by pushbutton A.I.? I wanted something decent, and there had to be an image there, but I didn’t care much what. So yes. But I’m not claiming this as my artwork. It’s just generative graphics.
Write the ending first. Make sure you know how your characters get from one scene to the next for every scene. There’s no such thing as a ‘pantser’. They end up writing a bunch of extra crap and wasting time and going back and working out the outline after the fact, and trust me, it’s twice as hard coming up with an elegant way of resolving your plot and story arcs that way, because you have to retrofit stuff and throw out a ton of crap you thought you needed but don’t.
So, to recap, write the biggest landmarks first. Then fill in the smaller landmarks in between. Then smaller than that. Keep going down in granularity until you’re writing scenes, then write the scenes, and you have a draft.
If you do not start with the broadest strokes first, building your story framework true and solid so you know where all the corners and leddges are FIRST, you WILL FAIL. The same rules apply to creating in EVERY OTHER ART FORM.
You can do the ‘pantser’ stuff when you get down to the level of writing scenes, and if you hold that off till then, it’ll be fun, because you’ll already know what the scene has to do in your story, what your characters need from each scene, and how everything works relative to the chapter that comes before it and the chapter that comes next.
Then you can do a polish pass and make everything look like you planned it all along in the first place.
Then get beta readers at minimum, a story editor if you can afford one, and a copy editor to make sure there are no spelling mistakes, because nothing pisses off a reader more than finding typos in your book and it makes you look like an amateur. Listen to your editor and your beta readers, they’ll tell you things you don’t want to hear but that you’ll have to fix.
And that’s how to write a novel. If you stick to this and you write a thousand words a day, you can write two full length novels a year.
Time after time, new innovations do not wait for human society to figure out how we’re going to integrate it. There’s a reason these things are called “disruptive” in business jargon. They shake the box.
For example, Photoshop has had AI powered tools in it for almost a decade now, and nobody’s making that a front-and-center issue. Grammerly has been around for a long time too, and nobody’s pointing at that in panic either.
But AI that can believably regenerate somebody’s voice, or study a few thousand images and make a new image that resembles them in style, and suddenly it’s important.
It’s not that the tools can do it. It’s a matter of degree. This shows us that the problem isn’t that the AI can do it – it’s that the sudden advances have taken us by surprise, and we realize that as a society we have been so busy trying to figure out how we CAN do it, that we haven’t stopped to think about whether we SHOULD.
Getty Sues Everybody
The lawsuit by Getty Images against the creators of Midjourney and Stable Diffusion claim that these tools store images and paste together parts to make new images like an electronic collage.
This is not even remotely how they work. Instead, a special kind of deep learning neural net is trained on the images, producing what is essentially a complex formula with hundreds of mllions of parameters that the AI generation tools use to create new images.
In my opinion these lawsuits will fail immediately on expert testimony because of this gross basic misunderstanding of the technology. Images are not being copied, are not being stored in the database. If they were, you would need thousands of terrabytes to store the data. As it is, Stable Diffusion can generate images on a dataset as small as 2.7 Gb. They don’t even make SD cards or flash drives that small anymore.
A further complication is that in Europe, as in the United States, datamining is legal, so after the question of copying is set aside (to reiterate, it’s not copying, it’s using the images to train a neural network), then it there’s a very good chance that the law suits will fail on the scanning without permission issue as well, because protecting from analysis is not a legal right any copyright holder anywhere in the world enjoys. If it were, simply observing an image displayed on the internet and having any kind of opinion about it would be a crime.
The images are being reduced to parameters in a very complex equation with hundreds of millions of parameters. Datamining isn’t illegal. Training neural networks on material you don’t own isn’t illegal either. Copyrights aren’t being directly violated, because you couldn’t bring up an exact copy of anything the neural nets are trained on if you tried (though you can get close). And, you can’t copyright a style, or a composition, or a color scheme. All that’s left is Right to Publicity, and the responsibility for that falls on the users of the tools, not the tools’ makers.
That doesn’t leave much meat left on the bone.
It’s Just Going Sideways
And sure enough, this is exactly how the law suits are shaking out. Sarah Silverman et al. tried to sue OpenAI for reading their stuff and incorporating that knowledge into their ChaptGPT model. The only problem was that they couldn’t make ChatGPT spit out exact copies of their manuscripts. The New York Times tried the same thing, and had the same problem . Why does this matter? Because in order for the courts to offer anything to the plaintiffs, first there must be a viable record of wrongdoing. It’s impossible for the courts to proceed on the basis of being butt-hurt alone. There have to be provable damages. The court runs on two things above all else: monetary damages, and proof of injury. The New York Times — and Sarah Silverman, and the handful of artists trying to sue Midjourney — haven’t established either one. Even to argue undue restraint of trade, the “right to publicity” argument, they have to show exactly how they’ve been hurt by the AI’s, and none of them can demonstrate this. These cases have been largely thrown out because of these lacks, and all that’s left is the damages from restraint of trade, which none of them can clearly demonstrate.
In my opionion, the writers and artists suing are the victims of class action ambulence chaser lawyers. If they win, mostly the lawyers get the money. And companies like Getty Images are only suing because they want to make their own generative AI service based on Getty Images licensed images and sell that as a service. When you can download Stable Diffusion and SDXL for free, why would anybody care?
The Right to Publicity
What remains appears to be Right to Publicity violations – the recognizability of artist styles, or celebrity faces, which have traditionally been treated by the courts as the responsibility of the individuals using the tools, and not the makers of the tools themselves. As a user, it is my responsibility not to try to sell AI generated images that simulate the style of Salvadore Dali, Chris Claremont or Michael Whelan and sell them with the claim that they are by the original artist.
Finally, if I happen to produce output that resembles one of those artists, how much can the original artist claim they have been damaged by such a production when human artists imitate the style of other artists all the time? Cases where one artist considers themselves damaged by someone else emulating their style are virtually nonexistent, and I could find no examples. Certainly apart from being grumpy about it, few can actually demonstrate in real numbers that their business is being negatively affected by it, if any. Greg Rutowski comes to mind, and even he is circumspect about it. He’s concerned, but he’s not losing his shit over it.
Sue the Tool User, Not the Tool Maker
Think about it for a moment: if they can stop Stable Diffusion and Midjourney for being able to replicate the style of other artists, then they should be able to stop all word processors for being able to output written pieces that emulate the style of other writers. Oops, I accidentally wrote a story in the style of Roger Zelazny, they’ll be coming for my copy of Windows Notepad now… Saxaphones should be outlawed because it is possible for another player to use one to replicate the style of Kenny G … Do you see the fallacy here? It’s not clear cut at all, and is in fact a matter of degree, which makes it a purely subjective call. In point of fact, those bringing these amorphous law suits not based on any established rule of law fail to inform the court as to why the existing protections against copyright infringement are insufficient and why the makers of tools are suddenly liable when they never were before now.
In any case, it’s too late to stuff the genie back in the bottle. AI powered art tools are here. It’s what we do next, to find ways to understand and integrate the new tools, that will define the new landscape.
It Feels Wrong, But Why?
And yet, one way or the other, we still have the same situation. Stable Diffusion, underlying technology for all the successful AI image generation tools, is open source. That makes it very hard to unmake, and even harder to undistribute. Additionally, while it’s obvious that disruptive technology is generally created for the primary purpose of eventually making money, it’s doing so here without breaking the law in any obvious way.
And THAT’S where the problem lies. The ability to replicate somebody’s artistic style to produce specific results is the part that’s disruptive. It makes it harder (and I know I’m preaching to the choir here) for artists to get paid for their work and to have the value of their work respected. Artists instinctively know this, but they don’t have much of defense for what’s happening to them, and this makes them feel like victims, and in a real way, they are.
Artists gotta eat. And pay rent. And visit the doctor. And initially, tools that do work they can do are going to break things.
But as with the invention of the camera, and the music synthesizer, artists will adapt their workflows to include the new tools, and those that do will have an incredible competitve edge.
And those that don’t — or can’t — will suffer for it, and as with any new technology, there isn’t a lot we can do to change that, except maybe help them avoid having their stuff analyzed for neural networks, or helping them learn how to use the new tools. The legal questions won’t be resolved soon enough to matter.
Nobody likes to be hit in the face with some new career-threatening problem that they didn’t see coming, and it’s hard to say that three years ago anybody saw this as an impending storm on the horizon. That’s why it feels wrong. It’s doing something with people’s artwork and photographs that nobody saw coming and for which the standard rules for intellectual property offer no protection whatever Whatever is going to happen as a result of this new technology is just going to happen, long before we figure out something practical to do about it, if we figure out anything at all..
Can Anything Be Done?
I can’t imagine how one would unexplode the hand grenade this represents, given that it takes ten to fifteen years to resolve landmark cases in court. By that time, the technology will have evolved well beyond its current state and likely built into practically everything.
The Getty lawsuit against Midjourney, Stable Diffusion et al. will likely fail on the merits because they don’t fully understand what they’re suing over, and they appear to be trying to claim rights they don’t actually have, but it’ll take years to even get that far. They can start their lawsuits over again and file new cases, but that starts the clock over from scratch.
Nor can they simply use the DMCA and have the source libraries removed from the web (I can’t imagine on what grounds they would do this, because the DMCA only applies to finished works, not tools for making them). Using DMCA’s on this stuff is like a perpetual unwinnable game of whack-a-mole even if somehow you could make it work.
So, I’m going to estimate ten to fifteen years to see anything on this, assuming there isn’t some sort of settlement. Considering Getty is looking for a couple of trillion dollars in damages, and they know they’ll never get that, it seems to me that they’re trying to just scare the ever-loving crap out of the defendants in court, going after settlement money so as to look good to their shareholders. They don’t give a crap about setting a legal precedent. There will be nothing upon which to base new case law, no judgment to cite, and the end result will be money changes hands (if it even gets that far). Once the lawsuits are over, the tools will just chug along as always, completely undeterred.
And the Getty lawsuits are the best shot at this there is.
A Note about Glaze and Nightshade
Both of these anti-AI image mangler apps attempt to “poison” AI by either adding small non-zero numbers to the latent image before passing it to the diffuser, or by adding “phantom” data to the image to fool the training step for the graphical models into thinking that a picture of a cat is, in fact, a dog. Neither of these really do what they claim to do. Both were developed in the “publish or perish” academic environment, by professors who only understand in general how their anti-AI tools work, and both were built on the efforts of unpaid graduate students who did the actual work. The effectiveness and quality of the results are, therefore, about what you’d expect.
Remember that the point of these tools is not to help artists protect their work. The point of the tools is to advance the reputation and standing of the professors involved, and few have the technical prowess to demonstrate that they don’t, in fact, work, except in a cleanroom setting where the variables of the test can be strictly controlled. They were both built to test against Stable Difussion 1.5, which is at this writing two full generations of technology behind that in most common use today. Moreover, the way Nightshade works focuses on token frequency in LAION and LAION tagging, which has been irrelevant for a while now.
Both rely on adding informational noise to the image to create the impression that the image is in a different style than it really is, or contains a different subject than it really does. Both, however, require that a model be trained on a heavy diet of the adulterated images before the trained model will exhibit the desired properties, i.e., to screw up the art styles or the content portrayal. Trust me when I say this: unless you are one of the most prolific artists in the world, and have the time to adulterate everything you’ve ever done over the years and re-upload adulterated versions of what you’ve made, you’re not going to have any affect at all on the training of new models. Heaven knows, after being trained on literally billions of images, you’re not going to have any affect at all on Midjourney or any of the other similar generative AI systems. That ship sailed literally years ago.
Most importantly, there is no evidence, apart from extremely narrowly defined tests in carefully controlled environments, that either Glaze or Nightshade work at all. I can’t stress this enough. You are far better off learning and growing as an artist and creating new art than you are hoping that magical fairy dust will protect your old work. The time to set all that up was before any of the major models were built, and anybody with a home computer can train a LoRa on your work and completely bypass whatever effects of either of these tools have.
I’m sorry if this is disappointing, but if it’s going to be stopped by the global community, there must be a plan put into motion that works. Intellectual property rights laws and right to access as they stand now simply don’t cover it. The next step is a concensus on what to do, but good luck reaching one. Humans have always acted as individuals. Given a population of a sufficient size, and a given stimulus, they will not choose to do a certain specific thing in response to that stimulus. They will do all the things.
That, to me, is what makes the arguments against generative AI art so frustrating. If AI art can’t be copyrighted, as many claim, then what rights are being taken from actual artists? There’s nothing to recover, because by that definition AI art has no intrinsic value. It’s all doublethink gobbledegook.
Anything that a human can imagine will eventually be made or built or invented, and sometimes by multiple people at the same time. I believe that AI art tools on this scale were inevitable. It’s how we use them and what we do next that matter.
These images, by the way, were all generated, by me, using a Stable Diffusion. I used Google to do image searches for each of them and I can confirm that they are not other people’s images. They’re unique as far as I can tell. If you find any of these images and it’s older than the copy posted here, let me know and I’ll take down my copy and reexamine my life.
They’re meant as computer wallpaper. If you see one you like, click on the image to zoom in, then right-click and “Save As.”
I’m playijng with Stable Diffusion now, and my father expressed an interest in good laboratory backgrounds for his videos that he wouldn’t have to worry about accidentally infringing on somebody’s copyright.
Gene to the rescue. Here you go, Dad.
To everyone else, if you want to use any of these, go right ahead. They are specifically not copyrighted (currently there’s debate on whether images generated by AI can be copyrighted at all), or proprietary, and made for using as background plates in Zoom. Go nuts.
To Use an Image
Click on an image you like, and it will zoom up.
Right click on the opened image, and select “Open in New Tab”.
In the new tab, right-click again and select “Save As” and you’ll be able to download the image to your hard disk.
Remember that these are all copyright free. Use them in anything you like, and don’t bother with attribution.
I have been searching for a reliable way to do this for literally years. I have, surprisingly frequently, found myself needing to move the contents of one WordPress site to another server, under a different URL. There have been lots of solutions for this over the years, most of them demanding not inconsiderable amounts of money, usually on the order of $45 dollars a year for an annual license, or more.
But I am a tightwad when it comes to things like this. I like to consider myself generous in other ways (helping friends with cosplay props, giving people needful things to make their lives better, et cetera) but I cannot abide being made to pay for something that just puts a few pushbuttons on something I know how to do myself.
Migrating a WordPress site, though, is a different animal. Smaller sites you can migrate yourself with the built-in import-export tools. Bigger ones kind of break that, and you have to have a plugin, or use a service, and that’s usually where you part with a couple of Benjamins.
Except that now there’s this plugin for Wordpres called WPVivid Backup, which just handles everything. You install a copy on both source and destination Wordpres sites, get a key from the destination site, paste it into the source site’s interface, and hit the button. It couldn’t be simpler.
Or cheaper, There is a premium version that adds some bells and whistles that I assume would be very useful for somebody who maintains dozens of WordPress sites for a living, but none of them are required for the basic essential function of porting a WordPress site from one domain to another, and WPVivid Backup is a free download, with no surprise “to finish this transfer process, pay the license fee that we hadn’t bothered to tell you about before you started monkeying with this” message in the middle. (Yes, one of the plugins I tried actually did this. Holding my data for ransom? Shame on you.)
In particular, it makes a web site designer’s job a lot easier because you can move a testbench site into its production URL WordPress setup without having to manually monkey with SQL dumps, shell access or any of that. It just freaking works.
If you want to do things like not migrate the entire site at once, you’ll have to pay for that. That’s fair and reasonable, and I may pay for that in the future if my back is ever up against the wall and I need it. For now, though, WPVivid Backup saved my ass.
Recent Comments