0:00 Picture this scenario. It's the 2022 Colorado State Fair. 0:04 Oh, I know exactly where you're going with this. 0:06 Right. So the digital arts category is being judged, and they announce the winner. It's a piece called Teatro d'Opera Special by Jason Allen. 0:15 And it is a genuinely stunning piece, like this sprawling, luminous, almost Renaissance-style scene. 0:22 Totally. But there is a massive catch. Jason Allen didn't paint it. He didn't even, you know... draw it on a digital tablet. He generated it using an artificial intelligence program called MidJourney. Yeah. 0:37 And when the traditional artists in that competition found out, the pushback was just immediate and fierce. 0:43 Absolutely. Because, I mean, the disruption didn't stop there. Fast forward to the 2023 Sony World Photography Awards. 0:51 Oh, the Boris Eldigsen case. 0:53 Exactly. This German photographer wins a really prestigious category with this haunting vintage portrait. Only, once again, it wasn't a photograph. 1:02 It was an AI-generated image. 1:04 But Eldickson actually did something different, right? Like, he deliberately declined the prize. 1:09 Yeah, he said he submitted the piece purely to force a conversation that the photography industry was just completely avoiding. 1:17 Which really sets the stage perfectly for what we're analyzing today, because we are surrounded by these generative tools at work, in our social feeds, our hobbies. But the fundamental rules governing them are just completely up in the air. 1:29 And that is our mission for this deep dive. For you listening right now, whether you're using these tools to, I don't know, draft emails, generate graphics, or you're just trying to make sense of the headlines, we are going 1:41 to untangle this massive collision between artificial intelligence, human creativity, and copyright law. 1:47 It's a huge topic. We need to look at the underlying mechanics of who actually owns what AI makes and whether an algorithm is even legally or philosophically capable of being creative. 1:59 Right. And to grasp the gravity of that philosophical question, we have to look back a lot further than 2022. 2:05 Way further. 2:08 We actually need to go back to 1843. 2:10 Wait, 1843? Yeah. 2:13 So the mathematician Ada Lovelace, who was studying Charles Babbage's early mechanical computers, she explicitly wrote that a machine has, quote, no pretensions whatever to originate anything. 2:25 Basically saying it only does what it's told. 2:27 Exactly. Her stance was that a machine can only do exactly what we know how to order it to perform. 2:34 It cannot create. 2:35 But then a century later, Alan Turing steps in and argues the exact opposite. He believed machines could and eventually would surprise us. Right. 2:44 Turing anticipated that machines would eventually move beyond just being rigid calculators. And looking at the outputs today, writing poetry, generating photorealistic images, writing software code, that surprise Turing predicted is absolutely here. 3:00 Because it is incredibly difficult for the average person to distinguish an AI's work from a human's work now. 3:06 It really is. 3:07 So to make sense of this modern Turing-Lovely showdown, we need to split our analysis into two distinct halves. Because, well, the media constantly blurs them together. 3:16 They do, all the time. 3:18 So first we have the outputs, like Who actually owns the image or the text that comes out of the AI? 3:24 And then we have the inputs. Is it even legal to feed millions of pieces of human art into an AI to train it in the first place? 3:32 Two very different legal battlegrounds. 3:34 Exactly. So let's tackle the outputs first. We know an AI can win a state fair, but can the person who clicked generate actually legally own that winning image? 3:45 Under United States law, the answer is generally no. Wow. Yeah. 3:49 And the legal bedrock for this actually stems from a highly unusual case that had absolutely nothing to do with software. 3:55 It had to do with a macaque monkey. 3:57 Oh, the famous monkey selfie incident. 3:59 I remember this making waves a few years ago. 4:01 Yes. 4:01 The 2018 federal case Naruto v. Slater. Basically, a crested macaque named Naruto picked up a wildlife photographer's unattended camera and snapped several smiling selfies. 4:12 And they went completely viral. They did. 4:14 But the courts ultimately ruled that a monkey cannot hold a copyright. The core legal requirement in the U.S. is that copyright demands human authorship. 4:24 Human authorship. Right. 4:26 So the logic follows that if an animal cannot hold a copyright, neither can a mathematical algorithm. 4:32 Okay. Let's unpack this because this is where the debate gets really tricky for anyone typing a prompt. If I use an AI generator, I am a human being writing the instructions. Sure. So is using an AI prompt more like ordering a meal at a restaurant? Like I tell 4:48 the chef I want a steak medium rare with peppercorn sauce, but because I didn't actually chop the ingredients or cook the meat, I can't claim I made the meal. 4:57 That's a great way to look at it. 4:58 Or, and hear me out, is it more like using a camera? Because early critics of photography dismissed the camera as just a mechanical tool that did all the work. 5:08 But today we fully recognize a photograph as a legitimate instrument of human creativity. 5:14 That's a fascinating debate. AI proponents rely heavily on your camera analogy. They argue that you are directing the tool, you're framing the concept, refining the output. Right. 5:23 I'm the director. 5:24 Exactly. But the U.S. legal system is currently leaning heavily toward the restaurant analogy. The courts are focusing on the mechanism of control. 5:33 What do you mean by control? 5:34 Well, when you take a photograph... You control the lighting, the aperture, the shutter speed, the exact moment of capture. 5:40 But when you prompt an AI, you are introducing a massive element of mathematical randomness. 5:45 Oh, I see. Because I can type the exact same prompt 10 times and get 10 completely different images. I'm not placing the brushstrokes. The AI's internal randomization is doing the heavy lifting. 5:58 Precisely. And we saw this tested directly in 2025. There was a major federal case, Thaler v. Perlmutter. where a creator essentially tried to list his AI, which he literally called the creativity machine, as the legal author of an artwork. 6:12 Wait, he tried to give the AI the copyright? 6:15 He did. His argument was that the machine itself generated the art. But the D.C. Circuit firmly affirmed the Copyright Office's refusal to register it, stating that human authorship is an absolute prerequisite. The machine gets zero rights. 6:29 But what if I don't try to list the machine as the author? Say I am writing an entire story and I'm just using AI to generate the illustrations. I'm building a broader human project. 6:41 That specific scenario played out in a really nuanced 2023 case involving a comic book, Zarya of the Dawn. 6:47 Oh, I've heard of this one. Yeah. 6:49 So the creator, Kristina Kashtanova, wrote the story and meticulously arranged the pages. But she used mid-journey to generate the individual comic panels. So the copyright office had to analyze where the human ended and the machine began. 7:03 That sounds like a nightmare to untangle. 7:05 It was. They basically decided to split the baby. They granted her copyright protection for the text she wrote and the way she creatively arranged the layout, but they explicitly denied copyright for the individual AI-generated images themselves. 7:20 So the overarching structure is legally hers, but anyone could technically take those specific AI pictures... and use them for something else because they were in the public domain. 7:29 That is the current stance. The Copyright Office actually followed up with a comprehensive 2025 report concluding that text prompts alone generally don't offer enough human control over the final image to claim authorship. Wow. However, if you use AI 7:43 merely as an assistive tool, or if you take an AI output and heavily modify it yourself in Photoshop, you can secure protection for the parts you personally authored. 7:53 That forces a very fine line for creators to walk. 7:57 But we should definitely note this is the perspective of U.S. law. 8:01 For our listeners working internationally, the rules aren't uniform, are they? 8:04 Not at all. It is a complete global patchwork. The U.K., for instance, has a specific provision for computer-generated works that credits the person who made the arrangements necessary for the creation. 8:17 Though I imagine applying a law written decades ago to modern generative AI is pretty messy. 8:24 Oh, it's currently a subject of intense debate there. Meanwhile, a Beijing internet court in 2023 took a very different approach. 8:30 They actually allowed copyright protection for an AI-generated image. 8:34 Wait, really? 8:37 Why? Because the judge determined the user provided significant intellectual input through detailed prompting, parameter adjustments, and continuous iterative refinement. 8:47 They essentially validated that camera analogy you mentioned earlier. 8:50 That is wild. Okay. 8:52 Okay. 8:53 So if the U.S. courts are largely saying the human prompt writer doesn't own the final image, we have to look at the only other entity involved, which is the AI itself. 9:02 Right. 9:02 But the AI didn't invent its painting or writing skills out of thin air. It built them by absorbing millions of existing human creations. Which brings us to the second half of our deep dive. The inputs. Who owns the training data? 9:17 This is where the financial stakes escalate into the billions. 9:20 AI models learn from massive data sets scraped from the open Internet. This includes copyrighted books, news articles, artworks, computer code, everything. 9:29 And rights holders like authors, illustrators, news publishers, they argue this is just unlicensed industrial scale copying, right? 9:36 Exactly. They see tech companies building incredibly lucrative machines entirely on the back of unpaid human labor. 9:44 But the AI companies argue this process falls under the legal doctrine of fair use. Yes. 9:50 They lean on precedents like the 2015 Authors Guild re-Google case, claiming the AI is engaging in transformative learning. The defense I hear most often from them is that it is legally no different from a human being going to a library, 10:04 reading a thousand mystery novels to learn the tropes of the genre, and then going home and writing their own original mystery book. 10:10 Okay. Here is where it gets really interesting, and I have to push back. Is an AI analyzing a text? 10:18 really equivalent to a human being experiencing a book? 10:22 That's a fair question. 10:22 Because a human reads a book, feels an emotional connection, and gets inspired. And AI is just a massive algorithm. 10:31 To me, it feels less like learning and more like a giant zip file mathematically recombining stolen work. 10:37 Actually, that zip file analogy is a very common misconception. 10:41 And understanding why is crucial to their fair use argument. 10:45 Really? How so? 10:46 Well, the AI is not storing pieces of images or text like a digital collage. It is learning the mathematical relationships between concepts. When it reads text, it breaks words down into tokens, which are essentially chunks of words converted into numbers. 11:02 Okay, I'm with you. 11:03 It maps out the statistical probability that the word apple appears near the word tree. It learns what a dog looks like by analyzing millions of photos and identifying the numerical patterns of fur and snouts. It stores those patterns or weights, not the actual photos themselves. 11:20 That makes sense. 11:22 And that is why the fair use argument is so difficult to legally crack. 11:26 I follow the math. But if it's only learning abstract patterns, how do we explain the regurgitation issue? 11:33 Right. 11:34 Because plaintiffs like the New York Times filed that massive lawsuit in 2023, demonstrating that if you prompt a language model in a specific way, it will spit out entire copyrighted articles almost verbatim. 11:46 That's true. 11:47 It can spit out famous book openings or exact chunks of copyrighted computer code. If it's not a zip file, how does it produce an exact copy? 11:55 The AI companies argue that regurgitation is a rare anomaly, basically a bug in the system. They explain it only happens when a piece of data appears hundreds or thousands of times in the training set, which causes the model 12:08 to over-index on that specific pattern. And they point to filters they are building to stop it. 12:13 But I'm guessing the plaintiffs don't buy that bug excuse. 12:16 Not at all. 12:17 Plaintiffs view that bug as undeniable proof that the machine retains the capacity to recreate the original works. And more importantly, it acts as a competing substitute. Right. 12:28 If the AI can summarize or reproduce premium journalism on demand, it actively destroys the market for a newspaper subscription. 12:35 Exactly. The concept of market harm seems to be the linchpin here. And we saw this play out with some very specific judicial logic in a major 2025 case involving the AI company Anthropic and Pirated Books. 12:48 Oh, the Barts v. Anthropic ruling. 12:50 Yes, the ruling by Judge William Alsip. He made a historic distinction regarding the source of the training data. He ruled that training in AI on lawfully acquired books was exceedingly transformative and constituted fair use. 13:03 Let me make sure I understand the logic there. Why does buying the book change the nature of the AI training if the algorithmic process is identical? 13:11 It comes down to respecting the initial market. If you buy a copy of a book, the author and publisher are compensated for that unit. The judge reasoned that extracting data patterns from a legally purchased item falls under fair use. 13:24 Okay. 13:25 However, Anthropic was caught using a data set tainted with millions of pirated books downloaded from shadow libraries. Judge Alsip ruled that bypassing the legal market entirely to build a commercial product is not excused by fair use. 13:39 Which was a massive deal. Huge. 13:42 Because of that tainted data, the case ended in a historic $ 1. 5 billion settlement with authors. That's about $ 3, 000 per book. 13:50 Unbelievable. So the source of the data dictates the legality. 13:55 But even with legal data, there are limits, right? Yes. 13:58 There was another recent case, Cadre v. Mehta. The judge actually ruled for Mehta in that instance, but emphasized a crucial warning. 14:06 What was the warning? 14:07 The judge warned that if rights holders can prove AI training floods their specific industry with competing works that destroy their livelihood, that broad market harm could still render the training unlawful in other contexts. 14:19 So exact copying or using pirated data gets you in trouble. 14:23 But what happens when an AI doesn't copy an exact image but instead perfectly mimics the vibe or the specific style of a creator? 14:33 Now we're moving away from concrete copyright and into the much grayer areas. 14:39 Right, because this is a massive source of frustration for visual artists. Under copyright law, you cannot copyright a style. 14:46 Exactly. You can copyright your specific painting of a fantasy landscape, but you cannot copyright the general technique of, say, painting dramatic fantasy landscapes with heavy shadows. 14:58 Which became an absolute nightmare for artists like Greg Rutkowski. 15:02 Oh, the Polish illustrator. Yeah. 15:04 He has this very distinct, dramatic fantasy style. And before generative AI had guardrails, users realized that adding in the style of Greg Rutkowski to their text prompts made the generated images look vastly better. 15:17 He became one of the most prompted names in AI generators. 15:21 Totally legal, but obviously devastating for him to see thousands of images mimicking his life's work. Right. 15:27 Because his specific expressions weren't being copied. Only the statistical patterns of his brushstrokes. But while style isn't protected, voice and likeness raise entirely separate and much stronger legal issues. 15:40 That makes me think immediately of the 2024 Scarlett Johansson controversy. 15:44 OpenAI released a voice for ChatGPT called Sky that sounded eerily like Scarlett Johansson's character from the movie Her. 15:52 And she threatened legal action, so they paused the voice, even though they claimed it was recorded by a different voice actress. 15:58 That incident highlighted the visceral reaction humans have to voice cloning. 16:02 It did. 16:03 And lawmakers are stepping in. For example, Tennessee passed the Elvis Act in 2024, which explicitly protects musicians' voices from unauthorized AI cloning under right of publicity laws. So while a visual artist's style might be legally vulnerable, a person's 16:20 biometric voice is increasingly recognized as protected territory. 16:24 But since the law is slow and the technology is fast, human creators aren't just waiting around for legislation. I've been reading about researchers at the University of Chicago who developed technical countermeasures called glaze and nightshade. 16:37 Oh, those are fascinating. 16:39 Artists can run their digital images through these tools before posting them online. It's essentially digital camouflage, right? like actively disrupting the AI-style mimicry. 16:48 That is a very apt way to describe it. These tools add subtle, invisible perturbations to the image's pixels. 16:54 What does that actually mean at a mathematical level? 16:58 Well, to a human eye, the artwork looks totally normal because our brains process the holistic image. But an AI scraping bot reads the specific mathematical RGB values of the pixels. 17:10 Nightshade alters those pixel values just enough so that the underlying math tells the AI a completely different story. 17:16 Wait, so what does the AI see? 17:19 A human sees a picture of a dog. But the poisoned data teaches the AI's neural network that the image is a toaster. 17:26 That is wild. 17:27 If an AI scrapes enough of these poisoned images, it fundamentally corrupts the model's ability to understand concepts. It is a form of active data sabotage. 17:37 Artists weaponizing the AI's own mathematical reliance against it. I love that. But on the flip side, we also see creators totally embracing these tools. Concept artists are using them to generate 20 rough layout ideas in five minutes. Musicians are using AI to brainstorm. 17:53 It creates a profound philosophical shift. 17:56 So what does this all mean for you? It reminds me of the distinction psychologist Margaret Bowden makes between the process of creativity and the product of creativity. 18:06 Right. Process versus product. 18:08 Yeah. If anyone can generate a flawlessly beautiful image in five seconds, does the societal value of art shift entirely away from how pretty the final product is and toward the human story of how it was made? 18:21 We are already seeing organized labor try to force that exact distinction. Look at the 2023 Writers Guild of America strike. 18:28 Right. They weren't just striking over streaming residuals. 18:31 Exactly. They were striking to define the role of AI in their profession. And The contract they secured is serving as a real-world blueprint for how other industries might adapt. 18:41 What did they manage to lock down in that contract? 18:43 And it was heavily debated. 18:45 They secured a hard rule that AI-generated material cannot be legally classified as literary material or source material. In Hollywood terms, this means a studio cannot generate a script with AI, hand it to a human writer, and say, just polish this up for a lower re-write fee. 19:02 The human maintains the status of the original author. 19:05 Yes. Furthermore, studios are required to explicitly disclose to the writer if any material they are given was generated by AI. 19:12 They established a contractual boundary that original writing is inherently a human labor category. 19:19 Okay, we have covered a massive amount of ground here, from 19th century mathematicians to billion-dollar data sets and pixel poisoning. 19:28 Let's ground this for you, the listener. What is the practical guidance here? 19:32 It's a great question. 19:33 If you are a creator, a business owner, or just someone experimenting with these tools, how do you navigate this safely? 19:41 First, if ownership matters to your project, say you are designing a logo for your company or creating illustrations for a book you intend to publish, you must add meaning. 19:50 human authorship. Purely AI-generated content is incredibly difficult to protect legally. 19:56 So use the AI to brainstorm, use it for mood boards, but the final protectable elements need your direct human execution. 20:05 It seems like documenting your creative process is more important than ever. 20:08 Absolutely. Keep your rough drafts, your sketches, and your Photoshop layers to prove your human involvement. Next, check the platform terms of service carefully regarding commercial use. 20:20 And always disclose your use of AI where required, right? 20:23 Yes. Whether you are registering with the Copyright Office, submitting work to a publisher, or entering a state fair contest, be transparent about the tools you used. 20:32 And beyond the strict legal rules, there is a necessary ethics check we all have to do because the law is still catching up. 20:39 Right. A reliable ethics check involves asking yourself three questions before you publish. One, was the AI tool I am using trained with consent or licensing agreements from the original creators? Two, am I intentionally prompting the tool to imitate a living 20:53 artist's identifiable style to bypass paying them? 20:57 That's a big one. 20:58 And three, am I actually adding my own creative choices to the final product or am I just clicking a button and taking all the credit? What's fascinating here is that there's rarely a single right answer. But asking these 21:10 questions is what separates thoughtful use from careless use. 21:14 Thoughtful use versus careless use. Which leaves us with one final concept to mull over on your own. We've spent this entire deep dive analyzing AI as an infinite, effortless generator of content. 21:25 It's everywhere now. 21:26 We are rapidly entering an era where perfection, a perfectly lit photograph, a flawlessly rhyming poem, a beautifully structured corporate email, is going to be incredibly cheap and instantaneous. Perfection won't be special anymore. It becomes a commodity. 21:43 When perfection is abundant, its value drops to zero. 21:46 Exactly. So will the future of the creative economy be defined by a human premium? Think about the watch market. 21:55 You can buy a digital smartwatch that is objectively perfect at keeping time, but people still pay 10,000 times more for a handmade mechanical watch. 22:04 Because they aren't paying for accuracy, they are paying for the human craft. 22:08 In a world where AI perfection is instantaneous, perhaps the ultimate luxury status symbol of tomorrow's art won't be its flawless execution. Perhaps the real value will be the proven, verifiable inefficiency of a human being struggling to make it entirely by hand. 22:25 The human struggle as the ultimate premium. It is a compelling way to look at where art goes from here. 22:31 Just like those traditional artists at the Colorado State Fair who were so outraged not because Jason Allen's painting wasn't beautiful, but because it bypassed the blood, sweat, and inefficiency of the human struggle. Something to think about the next time 22:44 you hit generate. Thanks for taking this deep dive with us. 22:47 We'll catch you next time.
Community discussion