Is Generative AI Ethical?

Generative AI (GAI) seems to be taking over the world. For instance, many people seem to have FOMO, as if somehow they’re falling behind even though GAI is pretty much brand new. It’s a point of contention in the WGA labor negotiations. I now get two or three emails per day telling me how to use GAI. I’ve heard GAI described as a productivity breakthrough for creators. Others have described it as the devil incarnate, a technology that will destroy jobs and careers.

?

Given the controversy and the hype, I’d like to explore a question that seems to be at the heart of the discussions, even though I don’t hear a lot of people asking it: is GAI ethical? I’m not sure that I have a straightforward answer to that question. But I do think it’s the right question to ask right now.

?

Here is one thing I’m very confident about. GAI is here to stay. It’s being adopted at a much faster clip than any new technology I’ve ever seen. I don’t believe this is a passing fad. But using GAI in a way that is ethical… Now that’s a topic I don’t see a lot of people talking about.

?

My goal in writing this article is not to tell you or anyone else how to use GAI or whether or not GAI is, in some universal sense, ethical or unethical. Instead, my goal is to help you understand how GAI systems work and how, under certain conditions, you might accidentally violate values I’ll bet you hold dear. Please let me explain.

?

On LinkedIn, I’m connected to thousands of creators. These are hard-working, brilliant and super-creative people who’ve made, or are making, careers that feed their families and their souls. The vast majority of these creators would hate it if someone used their artwork without their permission or without them being paid. In fact, they’d probably take legal action to remedy the situation. Yet, and I put myself at the head of this line, these same folks are probably at-risk of using GAI in ways that may not be ethical.

?

This is why I believe, if you are going to use GAI, (and I do believe most creators should at least be experimenting with it), here are some guard rails to consider. I suggest this for two reasons. First, you certainly don’t want to accidentally use GAI in a way that is unethical. Second, you don’t want your intellectual property to be used in ways that are unethical. The more you understand how GAI works, the better positioned you are to prevent these outcomes.

?

?

How I Came Across This Topic

As a creator, I’m constantly on the lookout for great new ideas for both our Supernatural Sounds music brand and our iLightStudios production company. The pandemic basically shut down operations on iLightStudios. So I was looking for new ideas to showcase our capabilities for music and storytelling. After kicking around several concepts with my team, we decided to produce a new fantasy series dubbed Gods And Warriors Trilogy.?

?

One of the three stories we decided to craft was about Thor, the Nordic god. To provide some inspiration and further direction, we began to look up images of Thor. The results were okay but not great. So then we decided to generate some new images using MidJourney V5 (MJ5). Once we began to enter prompts, the results were pretty impressive, but also a bit worrisome.

?

The first round of images looked at lot like the Australian actor Chris Hemsworth who played Thor in at least three movies. That was an obvious red flag for me because I very much doubted that Marvel studios had authorized their content to be injested by MJ5. So we rejected all the prompts that returned images that we could clearly see were derived from Marvel.

?

Our next round of prompts were far more promising and returned some great images. But I began to get this sinking feeling in the pit of my stomach as we were reviewing image after image. The characters and the composition seemed familiar to me, but I couldn’t put my finger on it. Then suddenly, after coffee one morning, it occurred to me: Molly Hatchet.

?

In the early 1980s my older siblings were in rock bands that covered Molly Hatchet. So their albums were always around our house. I remember holding the cover of Flirtin’ With Disaster and looking at the image of a Nordic (I assume) character. He wears a winged battle helmet and wields a battle axe. He looks like a total bad ass Nordic warrior, kind of like Thor. The creator of this album artwork is a famous fantasy artist named Frank Frazetta. If you haven’t seen his work, it’s definitely worth checking out.

?

Then it hit me. Could MJ5 have been “trained” on Frank Frazetta’s work? Since Frank has passed away, would this mean his estate could potentially be losing revenue to derivative artworks that were directly influenced by his art? To be completely honest, I cannot answer that question because I don’t know, and no one I’ve researched seems to know, what “sources” MJ5 was trained on.

?

But you can see my conundrum. I would never want to leverage someone’s work without their consent. I wouldn’t want anyone to do that to me either? But in the new age of GAI, I might, unintentionally and without my knowledge, generate content that was heavily influenced by, or derived from, a creator who is excluded from having any say in my usage and from any economic benefits I might realize.

?

________________________

?Key Take Away

In the new age of GAI, I might, unintentionally and without my knowledge, generate content that was heavily influenced by, or derived from, a creator who is excluded from having any say in my usage and from any economic benefits I might realize.

?________________________

?

?

What Is Generative AI And How Does It Work?

To help me understand this better, I recently enrolled in Google’s GAI course. It’s a solid overview of how GAI systems are designed, the kinds of problems GAI can solve and how content generation works. I don’t want to get too far down in the weeds on this because you can always take Google’s course for free. So I’ll try to keep this really simple.

?

GAI is a sub-discipline of AI. Whereas traditional AI is great at things like analyzing vast amounts of data and predicting outcomes faster and better than humans, GAI generates content: written text, images, audio and video. GAI systems, like most AI systems, can “reason, learn and act autonomously,” according to Google’s coursework. If you find that statement a bit scary, I think you should. I certainly do. How does GAI do this? In essence, there are five steps:

?

1. System Design. GAI systems are an aggregation of artificial neural networks. These are distributed high-performance hardware and software systems that mimic the human ability to ingest information, analyze it, store it and categorize it.
2. System Training. GAI systems are “trained” on existing content. If a GAI model was developed to generate images, then it will be trained on images. If it was designed to generate written content, it will be trained on written content. The same is true of audio and video GAI models.
3. User Prompts. Once a GAI model has been developed, users can enter “prompts,” which are written phrases. To my knowledge, all GAI systems draw upon LLM (large language model) protocols to interpret user input so the system can “understand” what it’s being asked to generate.
4. Content Generation. Based on the prompt, the GAI model generates content. If the model was asked to generate an image of Thor, it analyzes its database of images it was trained on that have been tagged as Thor. It then recombines elements of those images to generate something new.
5. Optimization. If the prompt generates content a user doesn’t find useful or if only parts of the content are useful, the user can refine their prompt or start over with a new prompt. In this way, the GAI model “learns” because it’s constantly getting feedback from users.

?

Now if you’re like me, you probably are focused on step two above—system training. Here is what I believe we now know about GAI systems. They really don’t create. They recombine elements of what they’ve ingested to generate something new. This means the content the GAI system was trained on becomes the crucial point in all of this.

?

?

What Does It Mean To Be Ethical?

For the record, Google’s GAI course didn’t use the term ethics. It has a small module that discusses “responsible GAI.” The video notes that organizations will use GAI according to their “values” and states that these values will differ from one company to the next. The video describes Google’s values for GAI as:

?

• Built for everyone.
• Accountable and safe.
• Respects privacy.
• Driven by scientific excellence.

?

I find this to be a reasonable starting point for a discussion of ethics. But to my way of thinking, it leaves out centuries of thought, especially in Western Civilization, about a topic that is central to ethical behavior: theft. Taking that which one does not own—stealing—is deemed to be unethical by almost any standard of behavior by pretty much every culture in the world. It’s a no-no.

?

But, you might ask, can a user of GAI systems be held accountable for theft if they didn’t create the system that generates the content they use? That’s a great question and one that I have no doubt will be litigated in the courts very soon. To get our bearings on this topic, I think we need to talk about how things work today for creators.

?

?

Copyright, Title, Licenses, Economics, Public Domain?

In the creative industries today, there are built-in protections for creators to ensure they are compensated (fairly or not I’ll let you decide) and have control over their art. Four very important underpinnings of this system include copyright, title, licenses and economics.

?

1. Copyright has to do with authorship. An “author” of a creative work—which could be a writer, painter, sculptor, designer, composer, photographer and the list goes on—claims that a work is legally and ethically their creation by way of copyright. In the US, the Library of Congress accepts and registers copyright claims from authors.
2. Title has to do with ownership of the art and the copyright. When an individual or organization registers a copyright with the Library of Congress, they are granted title to the work. Holding title is essential to everything else that follows.
3. Licensure happens when the title owner “licenses” someone else to use their artwork in specific ways and for specific dollar amounts. The terms of licenses can be simple and straight forward or incredibly complex. But it’s through the process of negotiating the license that the user is granted rights. Without the license in place, if someone uses protected artwork, they can be held legally and financially liable.
4. Economics has to do with how title holders get paid by those who license their artwork. Most creative industries have a somewhat different approach to this, but they all pretty much follow this process: a user requests a license; the title holder and their representatives negotiate usage and fees; a license is granted; payment follows to the title holder.

?

That’s pretty much how most creative industries work today. The legal system has been structured to ensure that creators have control over their artwork and get to earn income from it. Copyrights do run out. When this happens, the artwork passes from being privately protected by a title holder to being in the public domain. Just because an artwork is in the public domain, this doesn’t mean you can do whatever you want with it. But it does mean that you don’t have to seek a license from a title holder. The public now has the right to the artwork. But no one person can claim ownership of the original artwork or any artwork derived from it.

?

Given my analysis of these three concepts—how GAI works, that theft is certainly unethical and how creators are protected—let’s now turn our attention to how using GAI may, or may not be, ethical.

?

?

Obvious Ways GAI Use Is Ethical

?

If a GAI model was “trained” on 100% public domain artwork, I don’t see a scenario in which users of that model could engage in unethical behavior. The GAI model has no copyright protected material inside of it. Anything the GAI model generates, no matter how unique or interesting, could not be claimed by a copyright title holder. This also means that no title holder is being denied economic opportunities due them.

?

The same applies if the materials the GAI model was trained on came from copyright title holders who gave their consent for their artwork to be included in the training. Most recently, Universal Music Group partnered with AI company Endel to produce just such an arrangement. Usually in these types of scenarios, a blanket license is established with rights holders that authorizes derivative uses.

?

The same might apply for alternative types of copyright, such as Creative Commons (CC) licenses. There are several types of licenses like this where a title holder puts out artwork that can be used by the public, but usually with acknowledgement of the creator. The one scenario I can think of where a CC usage might be unethical is if a CC title holder didn’t give their consent for the GAI to be trained on their artwork.

?

________________________

??Key Take Away

The most obvious way a GAI model might cause a user to unintentionally engage in unethical behavior is if that model was trained on copyright protected works without the consent of the copyright holder.

________________________

?

Ways Generative AI May Not Be Ethical

The most obvious way a GAI model might cause a user to unintentionally engage in unethical behavior is if that model was trained on copyright protected works without the consent of the copyright holder. Any derivative works generated by that system could be based on, essentially, tainted training materials. This not only blocks the copyright holder from having control over their artwork, it also takes money out of their pocket.

?

My grave concern, at this point, is that this might be the norm rather than the exception. As GAI models proliferate, and I have every reason to expect that they will, this problem could become exponential. The courts could become jampacked with copyright infringement cases.

?

To return to the example I started this article with, I tried to discover if Frank Frazetta’s work had been included in the MJ5 training materials. I cannot definitively answer that question today. I see similarities with the artwork our prompts generated. Yet the differences are also quite noticeable. It’s a conundrum.

?

?

What Can You Do About This?

My advice, at this moment in time, is to be very careful about the use of GAI products, especially if you have commercial intentions. If you use GAI products to generate revenue and it can be proven that the GAI model you used contained copyright protected materials (I know that may not be easy) WITHOUT the consent of the copyright holder, you may find yourself staring down a very expensive lawsuit. Buyer beware.

?

But even setting aside the potential legal and financial considerations, my question is this. Do you really want to build your artwork on someone else’s artwork without their consent? Would you feel good about that if someone did it to you?

?

This is one reason I believe most creators should begin to ask hard questions of GAI model makers. These questions might include:

?

• What are your GAI model training practices?
• Will you publish a statement certifying that your GAI model was NOT trained on copyright protected works without the consent of the title holder?
• Will you reveal the sources of your GAI model training?
• Do you control GAI training or do you open it to the general public where anyone can feed the model content?

?

This is indeed a strange new world in which we find ourselves. But the kind of questions I’ve posed above might just help you get the most out of GAI without violating values I’ll bet you hold dear. If you have any questions or comments about what I’ve written here, please know I’d love to hear them.

?