CONVERGER #5: How to Copyright AI-Generated Works, From the Lawyer Who Did It for "Critterz"

Also: The A24/Google Deal and Other Team-Ups Between Studios and AI Labs; Hollywood’s New Short Story Gold Rush

Welcome to CONVERGER, a biweekly newsletter mapping the content singularity where AI and the internet collapse all media into one—a connective node where emerging technology, policy, culture, futures thinking and storytelling intersect.

<<Achievement Unlocked: CONVERGER was cited in the New York Times!>>

CONVERGER presents news and views from an AI, internet and media policy expert who is pro-innovation but anti-hype, allergic to both AI panic and AI boosterism, and passionate about supporting rather than supplanting human creativity with new technology.

Some issues may be heavier on media commentary, others on AI policy, others on personal passions like sci-fi’s influence on technology (both for good and bad) or the evolving medium and business of comic books in the digital age. You never know what threads might come together in convergence-space!

I’m Kevin Bankston, your host. You can watch me develop newsletter content in real-time on LinkedIn and the social network formerly known as Twitter, and less often on Bluesky and Instagram. You can also look for my more wonkish takes on AI governance at Elicitation, the new Substack from my AI policy day-job colleague Miranda Bogen of the Center for Democracy & Technology’s AI Governance Lab. (Note that my Substack articles don’t necessarily reflect CDT’s positions.)

In this week’s edition, we’ve got our usual long “Features” and shorter “Fragments” and also a new section of bite-sized “Flashes”! There’s just too much happening for CONVERGER to contain it all, but we try!

So let’s get to it. And please share with your friends and colleagues if you enjoy!

---

TABLE OF CONTENTS

FEATURES (>500 words)

1. How to Copyright AI-Generated Works, According to the Lawyer Who Did It for Critterz (~8000 words, ~30 minute read)

2. Keep Your Enemies Closer? A24 and Lionsgate Ally With Google and Runway as Competition in AI Video Heats Up (~1545 words, ~6 minute read)

3. Hollywood’s New Short Story Gold Rush (~1325 words, ~5 minute read)

FRAGMENTS (<500 words)

1. Take This Job and Shove It Train an AI To Do It

2. Prestige TV Writers Are Really Freaked Out by AI

3. Amazon Dumps Luca Guadagnino’s Sam Altman Biopic

4. Stop! That! Train! Director Who Denied Using AI Admits to Using AI

5. ByteDance’s Seedance 2.5: Now Tom Cruise and Brad Pitt Can Fight in 4K for 30 Seconds

6. Tribeca’s AI Selection Stinks

7. That’s Microseries To You, Sir

8. Hollywood’s YouTube Horror Buying Spree: Dumb Money or Smart Move?

9. Don’t Want to Show Your Face in Your Video Content? Then YouTube Doesn’t Want You

10. What YouTube Does Want: To Train on Your Content But Not Admit It

11. Has Another AI Story Won a Literary Prize?

FLASHES (<100 words)

Too many to list!

---

FEATURES

How to Copyright AI-Generated Works, According to the Lawyer Who Did It for Critterz

I wrote at length in issue #4 about my experience attending the “AI on the Lot” conference at the end of May in Los Angeles’ Culver City, including the story of how I (accidentally) helped get an AI show cancelled, a roundup of all my substantive takeaways, and an interview with my anti-AI Hollywood-executive brother who I dragged to the conference.

However, I saved the best for last: an extensive interview with Lisa Oratz, a senior counsel at the Seattle office of the tech law firm Perkins Coie. As the saying goes, Lisa has forgotten more copyright law than most lawyers have ever known, and she was kind enough to share that expertise first by delivering a talk at the conference, and then by sitting down with me for some deep conversation.

In this interview we address some of the copyright questions that were top of mind for AI creators at the conference: Do I own a copyright in the AI-generated works I’ve created? How can I increase the chances that I do? How do I convince the Copyright Office to register my copyright? How can I avoid copyright liability when I use AI models? Am I at risk even if what I generate doesn’t directly rip off another copyrighted work? Will “ethically-sourced” models reduce that risk?

We tackled all of this and more in a wide-ranging conversation, reprinted in full below! And if you want the quick TLDR, check out the block quote excerpts.

Kevin Bankston: So I guess I’ll just start with: Hi, Lisa—who are you, what do you do, and what brought you to AI on the Lot? And before you answer, a standard disclaimer for our audience: Lisa is a lawyer, and a damn fine one. But she is not your lawyer, this is not legal advice, and you should get your own lawyer if you want that. Lisa, please.

Lisa Oratz: Hi. I’ll add to that as I introduce myself: I also have to say that the opinions I express are my own, not my clients’ or my firm’s. And because we have clients on a lot of sides of these issues, on some of them you may not hear me expressing my opinions as much as just setting out the facts.

In any event—I’m Lisa Oratz. I’m an attorney who works at the intersection of technology, intellectual property, and entertainment law. I do each of those individually, but my fun spot is where they all connect. So something like AI on the Lot, where you have that intersection of AI with entertainment, is an area I focus heavily on.

Kevin Bankston: And you did an hour—a “legal power hour,” it was titled at the conference—walking through a lot of the issues we’re going to talk about today. And I’ve got to say, as someone who teaches copyright, you do it very well. It was really entertaining and enlightening, and I certainly learned a few new things, which I’m hoping to communicate to my readers in this interview.

So, where to start? Obviously there are serious copyright questions about AI training—whether using copyrighted works in AI training is a fair use, in copyright terms, and under what circumstances. There are a lot of cases working their way through the courts. You can take issue with this gloss if you like, but I’d say that, with two courts so far, they’ve basically held that it’s likely a fair use—so long as you didn’t pirate the copies you used, and so long as there isn’t evidence that the outputs are diluting the market for the original works. But many more courts are going to have their say, and probably the Supreme Court, before it’s done.

Are AI-Generated Works Copyrightable?

Today, I wanted to focus more on copyright for creators who are using AI rather than building AI products. And I’d love to start with the question of human authorship, because across the conference I often heard about the importance of ensuring that AI remained a tool for human intention rather than deferring to the AI’s creative decisions—not only to maintain high quality, preserve human authorship of the work, and avoid slop, but also to maintain legal authorship.

From what I saw, there was indeed a lot of authorial intention in the demos and workflows and videos. Creators were using gen AI to tell a story that they wrote, where they took care to make key character and background design decisions, or editing decisions or adjustments—big or small—to the video outputs, whether in terms of camera angle, or what virtual lens they were using, or the color of this, or the size of that, or a particular detail here or there. It certainly made me think that the people online who ding anyone who uses AI as an uncreative hack making slop were just dead wrong. The professionals at the conference, as far as I could tell, weren’t just typing things like “make a scene of Tom Cruise fighting Brad Pitt in an apocalyptic wasteland.” They were often using incredibly detailed prompts—sometimes pages long—for each individual shot. To put it simply, there was a lot of craft going on.

So my question to you is: did those humans actually author those resulting videos, from a legal perspective? Do they have a copyright in that? And I suppose, for that answer, we might want to start with Zarya of the Dawn—but you start wherever you want, and we’ll get there.

Lisa Oratz: I want to start by saying that there’s a lot of information, and a lot of misinformation, out there about copyrightability. You often see these headlines: “Court says AI-generated content isn’t protectable.” It’s a very nuanced area, but a lot of those broad statements just aren’t accurate. What is true is that you need to have human authorship, at least of some form, to have a copyright. That we know. Beyond that, there’s not a lot of law specifically in this area about what exactly is needed for human authorship.

So where I want to start is the one and only case we have that touches AI and copyright, and that is the Thaler v. Perlmutter case [in a decision by the DC Circuit Court of Appeals that the Supreme Court declined to review]. What that case said—and this was the case that generated a lot of those “AI isn’t protectable” headlines—was a very, very narrow ruling. What it said is that if you use AI and there is absolutely no human authorship of any kind, that isn’t copyrightable. And Thaler wanted to challenge the human authorship requirement, so he purposely said there was no human authorship and even named the computer system as the author. So what that doesn’t tell us is what type and how much human authorship is needed. The only real guidance we have on that—

Kevin Bankston: Actually, let’s pause before we get to that guidance, and I’ll explain a little bit more about the Thaler case for our readers. As far as I know, in that case the machine-learning algorithm, which was trained on a bunch of pre-existing art, would generate a piece of art based on you simply saying “generate a piece of art.” You’d basically press a button and it would generate a piece of art. There wasn’t a text prompt of any kind dictating what kind of output it should be—it was all basically the AI’s decision about what it was making. And Thaler, as you noted, was doing this as a bit of an art-hack project to object to the human authorship requirement.

Lisa Oratz: Yep.

Kevin Bankston: And what just happened legally was that the Supreme Court denied review of the appellate court saying there was no human authorship. That’s what led to all those bad headlines saying the Supreme Court decided there’s no authorship in generated AI. So I just wanted to add those details—but let’s move on.

Lisa Oratz: Yeah, exactly. So beyond that, the only other guidance we have is from the Copyright Office. But one thing that’s really important to note is that what the Copyright Office says isn’t law. It’s important, because they get to decide what’s registered, and their views can be influential on courts. But until we have a court decision, it’s hard to say what the law says. The Copyright Office has issued a number of rulings and provided guidance, and the high-level takeaway is that the Copyright Office’s view of human protectability is, I’d say, pretty clear—and pretty narrow, meaning they’ve set a somewhat high bar.

You mentioned Zarya of the Dawn. That was the very first ruling the Copyright Office made. It was about a graphic novel that had human-created text, but the images were created with Midjourney. And like you described with some of these others, the text prompts they used were very detailed—in some cases hundreds of words. But the Copyright Office didn’t think the text prompts alone were sufficient human authorship. So they didn’t find that the images created as a result of the text prompts were protectable—even though, in addition, they regenerated and there was a lot of iteration.

Now, they did find that the graphic novel as a whole was protectable, both because there was human authorship in the text and because of the selection and arrangement of elements. That’s a well-known copyright principle: if you select and arrange unprotected elements, or combine protected and unprotected elements, that in and of itself can be copyrightable. So even though they did not find the actual Midjourney-generated images protectable, they did say the graphic novel as a whole was protectable. And they also said that if you modified the images—at least enough, beyond just a slight amount—that could be protectable. So even within that narrow view, they didn’t say, “Well, this graphic novel has no protection because they used AI.”

Kevin Bankston: So in practical terms, that would mean, for example, that I couldn’t market copies of Zarya of the Dawn myself, because that would include a bunch of copyrightable work the author did. But I could, for example, extract one of the Midjourney images and sell an art print of it, and because there’s no copyright in that, I’d be okay.

Lisa Oratz: Yes—theoretically, if you can figure out what came out of the AI. If the images are unmodified, that’s probably not so hard; you can do that. But if the author modified the image—and this is what we typically advise our clients—if they take those images and modify them, they’ll have copyright in the modified version. Not the original version, but if you don’t tell people what you started with and what you modified, when they see those images in the graphic novel they couldn’t just take them out, because they’d be unable, as a practical matter, to separate out what was human versus what was computer-generated.

So that’s why, when you make those modifications, you should keep confidential what it is you started with and what you added. You want to keep notes on that, because you want to be able to show that human authorship if necessary—but you don’t need to tell people what it is. Now—

Kevin Bankston: So we’re going to dive deep into that practical advice, but I’d like to pause for a minute and talk a little more about the Zarya decision. As a lawyer—analogies are always critically important in legal reasoning—it seems the Copyright Office adopted an analogy where, when you’re working with an AI, your prompts are more like you instructing a contracted artist to make something for you, but that contracted artist isn’t human, so there’s no copyright in that work. Was that the right analogy? You mentioned you thought the bar was high and the decision was narrow. How do you think the line could have, or should have, been drawn differently?

“Zarya of the Dawn was the very first ruling the Copyright Office made. It was about a graphic novel that had human-created text, but the images were created with Midjourney. The text prompts they used were very detailed—in some cases hundreds of words. But the Copyright Office didn’t think the text prompts alone were sufficient human authorship.”

Lisa Oratz: Yeah, I do think they took too high a level of human authorship in terms of what’s required. First of all, you’re right. You could take the position that a prompt—at least a detailed prompt, not a two-word prompt—involves a lot of human input and is sufficient human authorship. But their view was that, in addition to your analogy, which they made—this is just like instructing someone—what they were looking for was control. They said those prompts can influence the output, but they don’t control it. So they had this theory of having to control something.

That seems odd to me, because you don’t always control things that are protectable. For example, if you’re a photographer photographing a football game that’s moving, you’re not controlling the movements; you’re not entirely in control of what you capture. You may be in control of some elements, but this notion of predictability and control is something they’ve added, and I don’t really know where the precedent for that comes from.

But you’re right—they did view this as merely giving instructions. I think, though, that at some level, if you give detailed enough instructions, so that you’re really shaping what that looks like, there can certainly be an argument. And the other thing I’d mention is that the threshold for originality for copyright is very low—you just need a modicum of originality. So I certainly think there’s also an argument that you only need a modicum of human authorship. The view that the human authorship has to really dominate, and that the AI is just a tool—I understand that position, but I don’t necessarily think it’s the right one.

And there is a similar ruling that came out shortly after, involving a picture by an artist, Jason Allen, that actually won an art contest [at the 2022 Colorado State Fair]—same kind of ruling. He’s actually filed suit challenging the Copyright Office ruling, and they’ve now briefed a motion [for summary judgment] there. So hopefully, in the not-too-distant future, we’ll have something that answers the question more specifically about prompts.

Because right now what I hear people say—based on decisions like the Zarya of the Dawn decision—is that the output that comes directly out of the tool isn’t protectable. And that’s true if you assume text prompts aren’t sufficient. But that’s where we really need to see whether a court agrees with the Copyright Office. When you look at old, old cases—there’s one from the 1800s that involves photography, when that was the newfangled invention, and they looked at what a photographer added versus what the camera actually did—it’s a very different analogy. So I think it’ll be interesting to see how a court looks at that, and whether they make the same determination the Copyright Office does.

Kevin Bankston: And, as we discussed earlier at the conference—talking about that Supreme Court case, Burrow-Giles—the court didn’t say, “Well, this bit of the photograph is protected and this bit is not.” They said the photographer clearly made some substantial choices, therefore the photo is protected. But it seems like the Copyright Office, as you mentioned, got kind of hung up on the stochasticity that’s inherent in AI outputs—although if they’d seen the myriad ways I saw artists controlling for that stochasticity at this conference, I think they might have looked at it differently.

Lisa Oratz: No, I think you really put the nail on this. When you look at these other cases, and other situations in general, you think of human authorship as a minimum threshold: if you meet it, it’s protectable. But you’re absolutely right—in the Burrow-Giles case they didn’t pick apart and say, “Okay, you control the angle and the lighting and these other features” (which is what they found the protectable contributions of the photographer were) “and that’s protectable, but all the other stuff isn’t.” Yet for AI, they do.

And, frankly, I think the Copyright Office has been somewhat hostile to AI. Now, we don’t really know going forward what their views are, because there have been changes at the Copyright Office. But under the current view that was expressed, at least in the report [on copyrightability] that was issued, it does seem like they’re a little bit disfavoring AI, because I think they’re taking views they have not taken in connection with other types of work.

Copyright Registration for AI-Generated Works in Practice: The Example of Critterz

Kevin Bankston: One of my big criticisms of Zarya when it happened—and when the Copyright Office reiterated its take in that report on AI copyrightability—was the practical question of how this is supposed to work. Is everyone going to have to start keeping a detailed audit trail of the human input into their generative AI work? Are they going to have to submit that to the Copyright Office? How the heck is the Copyright Office—never flush with cash or staff to begin with—going to meaningfully parse this vague line-drawing exercise, especially when more and more works are going to have AI intermingled into them in some way or another? It just seemed really impractical, especially at scale.

But I did have at least one of those questions answered directly at the conference, because lots of speakers—including, repeatedly, Albert Cheng, the VP of the AI studio at Amazon MGM, which was the studio hosting the event—highlighted the need for both maximizing and fully documenting any human input and control over the generative AI process. Not only to keep it artist- and story-centric and avoid generating slop by defaulting to bad creative decisions by the AI, but also in terms of protecting the copyright in the work.

It was also clear that one of the reasons traditional studios were interested in hybrid approaches to gen AI—like, say, human performances in front of AI-generated backgrounds—was that they’d more clearly have a copyright in the elements, like the recorded human performances, that would be hard to separate out from the other elements in the final product.

I guess my question to you is: is that take correct? Were they saying the right thing? How should creators be approaching this if they want to maximize their copyright protection?

“I do think it’s important internally to document your human authorship, because if you’re challenged and someone claims this isn’t protectable, you want to have good records. However, the good news is that the Copyright Office has been very clear that they don’t want detailed descriptions…. What you don’t need to do—and actively don’t want to do—is get very granular and say, “We took this image, and this is what it looked like, and we changed it.”

Lisa Oratz: I’d say—very lawyerly—yes and no. There are two different points there. In terms of having to keep detailed records: I do think it’s important internally to document your human authorship, because if you’re challenged and someone claims this isn’t protectable, you want to have good records.

However, the good news is that the Copyright Office has been very clear that they don’t want detailed descriptions. When you file a copyright application for work involving AI, you do need to disclose it—you need to describe the human authorship, and you need to disclaim things that aren’t protectable. But they’ve made it clear that that does not mean you have to give a very detailed description.

We’ve now obtained several copyright registrations—probably more than that now—for AI-generated works, and really all we’ve described is a paragraph or two, at a very high level, of the type of human authorship: if there was a human script, that they selected and arranged the images, that they modified the images or the video, just at that level. So it was maybe two or three paragraphs at most.

What you don’t need to do—and actively don’t want to do—is get very granular and say, “We took this image, and this is what it looked like, and we changed it.” The Copyright Office has no interest and no time to parse through all that. So I think it’s important to keep records of that, but I’ve heard lawyers even say, “Oh, you’ll have to submit this to the Copyright Office”—and you just don’t want to do that, and don’t need to.

The second piece of what you said is the hybrid approach. Certainly if you’re combining—which many of my clients do—you may have human voice authors, voice actors, human scripts; it’s always helpful. But even if you have something that doesn’t have those kinds of clear human elements, and you’re just talking about AI-generated video—nobody I know takes one prompt and spits out a whole movie. You’re taking pieces of it, putting it together, compiling it. So the film as a whole, or whatever the content as a whole—I don’t think there’s a lot of concern that it will not be protectable.

I mentioned at the conference that the first registration we got was for Critterz, which has gotten a lot of attention. We think that was the first registration for an AI-animated film. I really wasn’t worried we’d get registration—I knew we’d get registration on the film. The question is what we’d have to disclaim, what might not be protectable. But as we talked about before, in that situation they layered and combined so much, and modified things, that it would be very difficult for anybody to figure out what was protectable and unprotectable. So they have, I think, the same level of protection as if you had the hybrid approach you talked about.

Kevin Bankston: Okay, so let’s dive deep into that. First off, I’m going to run down some facts about Critterz, and let me know if I got any of them wrong. Critterz was an AI-generated short film, now being adapted into a feature. It debuted in April 2023, and y’all filed for registration at the beginning of 2025.

It was written and directed by a fellow named Chad Nelson, an OpenAI creative specialist, working with Nik Kleverov’s AI studio, Native Foreign. In terms of AI generation in particular, the creators used the then-state-of-the-art image-creation model from OpenAI, which was DALL-E, for the creature designs and the background design. Is that all good so far?

Lisa Oratz: Yeah. I’ll note that Chad did this before he went to OpenAI—his joining OpenAI sort of came out of that.

Kevin Bankston: I see. Got it. And in terms of production, I believe everything beyond those initial DALL-E generations was conventionally created. It was a human script, human voices, edited by humans using Adobe Premiere, animated by humans—they animated the DALL-E assets via Adobe After Effects for backgrounds, and then used Unreal Engine for the character animation, including human performance capture. So a lot of humanness.

Lisa Oratz: Yeah, it’s really interesting. Lots and lots of human authors.

Kevin Bankston: And I’m not sure about the music, but I think the music was human as well. So it’s not really a surprise that it was registrable, but I’d like to talk a little more about how you got there. You said you only really had to give them a few paragraphs—basically the things I just said—I guess in an attachment.

Lisa Oratz: Yeah. We combined—basically, explaining that they took AI-generated images, modified them, combined them, and sprinkled them with pixie dust, which is one of the things that was added, which I always loved. So we described, at that high level, what was done—that there were voice actors—and again, I think it was two and a half paragraphs. But the other thing—

Kevin Bankston: And that attachment—the Copyright Office didn’t publish it with your filing form. Was that at your request, or just them not publishing it?

Lisa Oratz: They tell you to put it in the notes, and those don’t get published. Now, it doesn’t necessarily make it confidential, because they have it, and—I frankly don’t know if somebody could get access to it—but it isn’t typically published.

There is a process the Copyright Office has to file confidential information, which is used for things like source code. So if you really wanted to describe something confidential, you could use that process. But it generally isn’t necessary, because we didn’t have to give them any more information than I’d tell anybody when I give these talks.

Kevin Bankston: So, the really interesting thing that came up in your talk was how you disclaimed what you needed to disclaim. Because, as we already talked about, there’s no law on this yet—it’s just the Copyright Office’s opinion—and yet, for them to honor your registration request, they seem to want you to disclaim any copyright in anything AI-generated. How do you do that without disclaiming something a court might later find you had a right not to disclaim?

Lisa Oratz: Yeah, that was exactly the catch-22, because we knew they’d take the position that the images, as they came out of Midjourney, were not protectable based on text prompts. But we also feel there’s a good chance the law will change and they could be protectable. We had to disclaim unprotectable elements and say what we weren’t claiming. A lot of the registrations beforehand just disclaimed all AI-generated images. But we tried to get a little creative and see if we could do something that would flex as the law adjusts.

So what we did was to disclaim the AI-generated content “to the extent not copyrightable.” The idea is that, as the law evolves and determines what’s copyrightable, the disclaimer could move with it—because once you disclaim it, if we disclaimed those AI-generated images and then the law later said, “By the way, those actually are protectable,” we’d have already disclaimed it.

I wasn’t quite sure how they’d react to that, but they accepted it without comment. We’ve taken that approach subsequently and have had success. So that was the way we tried to deal with the fact that we didn’t want to get a refusal and have to spend a lot of time and money arguing in court about protectability—but without having to disclaim something we think quite possibly could be protectable.

That’s the thing I really want to emphasize, because even people who are fairly sophisticated on this—who understand that the human elements are protectable—keep going back to this notion that the AI-generated output isn’t protectable because of the Copyright Office view. But I really do think there’s a good chance that, at least with sufficiently detailed prompts, the court might decide otherwise when they hear the Allen case.

Kevin Bankston: Yeah, I tend to agree. But we’re in a practical position where they’re saying it’s not protected—stuff that came straight from the AI after a text-based prompt. Practically, though, you’ll likely get protection out of selection and arrangement and whatever other human elements got intermixed.

Lisa Oratz: Quite a bit.

Kevin Bankston: And so long as you keep the actual outputs confidential and your filing is vague enough—which they did not push back on—it would be really hard for anyone to copy anything from your work without risking copying something infringing.

Lisa Oratz: Yeah, and that’s the practice.

Kevin Bankston: Is that the practical takeaway?

Lisa Oratz: That’s the practical takeaway. All this stuff is academic—what you care about is that somebody can’t take your work.

The other thing I’ll mention here is that the Copyright Office did rule on—because, as I said, they’ve been pretty limited, but there have been rulings that recognized some very helpful things. One of them was that, as you depart from text prompts and get into what they call expressive prompts, the output can have at least some protectability. There was a ruling on an image called “Rose Enigma.” It started with a human sketch, and then they used AI, and it fleshed it out more—in 3D and color—and added some stuff.

They didn’t take the approach of that minimum threshold, where “there’s enough, and the whole thing is protected.” They did still parse out what was human-created versus what the AI tool created. But they did say that, to the extent the human sketch is reflected in the ultimate output, that’s protectable. So that’s the other thing—where we talk about starting from human expressive inputs as opposed to text prompts—that can really help enhance making sure what you have is protectable, because that will dominate in the output.

Kevin Bankston: So let me try to summarize. In terms of maximizing copyrightability: ideally, you want to use expressive inputs, not just text prompts. If you’re going to use text prompts, they should be as detailed as possible. You should be selecting and arranging and combining and modifying elements with human input and control, and you should be documenting all of that—while also documenting, but keeping confidential, the output. Am I missing anything?

Lisa Oratz: It’s fabulous—you should be giving these talks for me. That’s exactly right.

Kevin Bankston: Well, so let’s take it from the practical to the academic—or at least the policy side of it—because I find myself questioning: what is the world we want here, and what is the world the Copyright Office wants here? I’m not religiously opposed to limiting copyright in prompted AI-generated content. That would mean there’d be an enormous amount of new public creative goods, including a lot of software, and the generation is so—I don’t want to say easy, but it is easier than traditional modes of creation—such that I’m not that worried about preserving the incentive to make things with AI.

But if the point is to ensure that only originally authored stuff is protected, how does this application of that policy actually get us there? Because, as we just discussed, the practical result is that people still can’t really copy the stuff anyway. But if you actually wanted people to be able to do that kind of parsing, it would require the kind of documentation and deliberation at the Copyright Office that wouldn’t scale. So I find myself wondering: what is actually the world we want here, in terms of a policy outcome?

Lisa Oratz: Yeah. I guess I am a little more concerned, perhaps, than you are about not recognizing all of the creative input that goes into using an AI tool. I agree that if you’re creating an image and you say, “Give me a teddy bear on a beach blanket,” and out comes an image, that’s an idea—that’s not really protectable.

But when you look at all the work that went in—and this was detailed in that graphic-novel Zarya of the Dawn case—and how much effort the artists made to shape the output, they kept regenerating images, they had detailed prompts. They really used it as a tool to realize their vision. And I don’t know that it’s very different. It’s interesting, because if you use Photoshop to do certain things—at least traditional, non-AI Photoshop—the Copyright Office much more readily recognizes that as a tool than AI, which they seem to treat a bit differently once you do that.

But seeing how many of my clients are really using this very creatively—even without the expressive prompts, even with very detailed text prompts and regeneration—I really do think that outcome does reflect human authorship. And the line—because, again, you don’t need a lot of human authorship, you don’t need a lot of creativity, for a copyright—the line should be lower. As a society, we’re moving into an era where we’ll be using AI very heavily in a lot of our creative endeavors, and we don’t want to lose protection merely because we’re using these tools.

“All the clients of these AI studios want to approve the prompts before you use them, to make sure you’re not citing someone else’s work or style in those prompts, to try to minimize the risk of infringement. They also want indemnity around copyright issues. So I guess the question is: how serious is the risk of copyright liability if you’re using AI?”

As I said, that you’re merely using them is not the law—but having to really fight and worry about all of this seems like much more hand-wringing than we should have to do, because I’ve seen how much creativity comes from the use of these tools, at least when done right. There can be a lot of slop, too. But I think we should go back to traditional concepts and look at a line of human authorship that’s not quite so detailed—not picking apart this element and that element, but really setting more of a floor. And then, once it’s protectable, not trying to pick apart which elements are or aren’t protectable. It just seems silly to me.

Kevin Bankston: It honestly seems a bit silly to me. And to echo your point about incentives—it was clear from the conversations I had at the conference that one of the things slowing adoption of the technology by the studios was concerns about the copyrightability of what they would make, and whether it was actually worth the capital investment if it was questionable whether they’d actually own the resulting work. The other thing they were concerned about was potential infringement liability.

Copyright Infringement Risk When Generating Creative Works

We could talk a little about infringement risk, because one of the things I heard repeatedly, both on and off stage, was that all the clients of these AI studios want to approve the prompts before you use them, to make sure you’re not citing someone else’s work or style in those prompts, to try to minimize the risk of infringement. They also wanted indemnity around copyright issues. So I guess the question is: how serious is the risk of copyright liability if you’re using AI—or contracting with a creator who’s using AI for you—and how might you minimize that?

Lisa Oratz: Yeah, that’s interesting, too, because you’re right—fear drives a lot of issues here. You said fear of copyrightability, which, again, is overblown, because if you do it the right way you can get protection. I think the same is true with infringement. We’re certainly in an area where there’s not a lot of definitive law about the scope of infringement.

But at the same time, as users, I think the practical risks are not as high, because all these cases focus on training. Some users also do their own training, and there are concerns there that the law is still shaking out in terms of training. But as a user, what you’re concerned about is: does your output reflect—and is it substantially similar to—training data, or something else out there that someone had access to? You see these cases where they’ve demonstrated output being similar, but those were also constructed by very specific prompts that were trying to elicit that similar output. I think there are a lot of things users can do to avoid, or at least minimize, the risk that output will look like existing output. One of them is, as you said—

Kevin Bankston: You mean substantially similar to existing copyrighted works.

Lisa Oratz: Right—copyrighted works. So one of the things you can do is not prompt the AI in a way that’s designed to output something similar. That’s why a lot of people are concerned about not putting in prompts that mention a specific IP, whether it’s an artist, an author, or a movie—or even a style. Even though style itself isn’t protectable, it does increase the practical risk that you could end up with something similar, because the tool sees that it’s pulling in something that may be a little too close. So we usually advise to avoid that. And not even just mentioning it: if you described, in very detailed terms, a copyrighted work—imagine you’re describing a specific piece of art and describing everything about it in detail—well, it may give you that, and it may be similar. So avoid anything that’s really geared that way, whether you mention the IP or just describe it.

The other thing you can do—when you start with those expressive prompts, which also helps with protectability—is that, because that’s going to dominate a lot in the output, the chances that you’ll pull back something similar to somebody else’s IP, based on everything that’s being trained on, are much less. Now, there still can be some risk, depending on how the tools work and how they’re trained. So there’s another set of things you can do to protect against risk as a user, having to do with the tools you use.

The first is that there are indemnities a lot of tool providers are making, at least if you have the enterprise version. There are caveats and limitations to that—some of those carve-outs are, for instance, if you don’t prompt things in a way intended to elicit that kind of content—so you need to look at that, because they’re not all the same, but it’s something you can look for.

“While I can’t say 100% there’s no risk there, I think it’s highly unlikely that, as a user, you’re going to be liable for infringement because the tool was trained on some [copyrighted] data, if that doesn’t show up in your output somehow. I think the argument for some sort of liability based on derivative works is pretty far-fetched.”

There are also some tools that are, frankly, trained better than others, because there are things toolmakers can do to filter, or to make sure they’re deduplicating. We won’t get into a lot of the detail here, but there are a lot of things you can do to try to limit the risk that you “regurgitate” training data somehow—which is the term—because that’s not how it’s supposed to work. If it does that, it’s not doing what it’s supposed to; it’s really just supposed to be learning from the training data. The other thing you can do—when you modify output, which again you want to do for copyrightability—that can help, too. Now, as you know, modifying an infringing work doesn’t make it non-infringing, but if you modify it enough that it doesn’t look like something else, then, as a practical matter, will somebody see that? The original image may have been too close.

So, all that to say: I think, as a user, if you’re doing it correctly and following some of the advice we’ve talked about, it’s not zero risk, but the risk that you’re likely to pull up something close enough to something else to be substantially similar, and meet that threshold, is less likely.

The other thing I’ll mention is that there are tools out there that talk about being trained on licensed content—being “safe.” I’m not going to say whether that’s necessary; that’s still shaking out, and I don’t really want to get into my opinion on whether or not you need to, or should, do that. But it’s something some users are doing, both from a copyright standpoint and sometimes because, ethically, they like the fact that it’s trained only on licensed or public-domain data. The challenge is that, because of the vast amounts of data you need, some people are concerned that those models may not do all the same things, and may not be as robust as other models. So those are some of the counter-concerns. But there are a lot of things a user can do to avoid and limit infringement risk.

Kevin Bankston: On the ethically sourced models, one thing I’m keeping in mind is that even if they’ve licensed all of the content for training—which you, as a user, may or may not care about legally, but may care about ethically—there’s still a chance you could elicit something that would infringe on those works. The fact that they licensed it for training doesn’t give you a license to use it in your work.

Lisa Oratz: Yeah. Well, that’s right—that’s exactly the question: does their license extend to output? We don’t always know the extent of those licenses. So that’s certainly one way to go, but I don’t necessarily think you need to go there, if you try to make sure you’re prompting in a more general way and choosing tools you feel comfortable have taken steps to limit that risk and to provide decent indemnities.

Kevin Bankston: What about if your output isn’t substantially similar to a work that was trained on? Do you see a plausible argument that there’s potential liability for creating a derivative work from those works that were trained on, even if there’s not a substantial similarity?

Lisa Oratz: Yeah, that’s an interesting question. Some of the cases have tried to take the view that, somehow, by definition, if you’re creating output that’s derived from the training data, it’s a derivative work. Again, we’re very, very early in this litigation—there hasn’t been a significant substantive ruling on that. But from preliminary rulings [see, e.g., this Northern District of California ruling from 2023 on a motion to dismiss in the Kadrey v. Meta case], the judges haven’t really given a lot of weight to those arguments, and they’ve suggested that, no, it has to be substantially similar.

“One of my goals… is really to help spread good information and limit the fear. So what I’d say is: go forward based on being educated and knowledgeable, and don’t act on fear, because there’s just so much fear going around. I really think that is limiting adoption.”

So, while I can’t say 100% there’s no risk there, I think it’s highly unlikely that, as a user, you’re going to be liable for infringement because the tool was trained on some data, if that doesn’t show up in your output somehow. I think the argument for some sort of liability based on derivative works is pretty far-fetched, but—

Kevin Bankston: I mean, I hope you’re right, because otherwise there are a lot of infringers out there—inadvertently.

Lisa Oratz: Yeah. We’ve actually done a lot—very early on, for some of our clients in this space who are tool providers, we did a lot of detailed analysis on this, which I won’t get into here. But when you look at other arguments, other cases, and other situations, those arguments seem pretty weak substantively.

Kevin Bankston: Okay, well, Lisa, you’ve been enormously generous with your time and expertise, and I can’t say how much I—and the readers of CONVERGER—appreciate it. I hope I’ll see you next year at AI on the Lot. I certainly got enough out of it that I plan to attend next year. But before we go: do you have any closing thoughts or bon mots?

Lisa Oratz: I don’t know about a bon mot. But AI on the Lot is a great conference, and it’s always fun, because one of my goals, besides the fact that I think it’s fun to talk about this stuff, is really to help spread good information and limit the fear. So what I’d say is: go forward based on being educated and knowledgeable, and don’t act on fear, because there’s just so much fear going around. I really think that is limiting adoption—although one of the differences this year from last year, and that I’ve seen over the last year, is a bit of a move away from the fear and into the inevitability of doing this. That’s why I think it’s important to take a practical approach.

And the last thing I’ll say is that, at least for organizations of any size, we talk about either creating or updating your AI use policy. You want to make sure that, if you’re implementing some of these tips for limiting your risk, that it makes its way throughout your organization, and that you’re giving thought to decisions about the tools you’re using and how you’re using them.

There’s also a host of other issues we didn’t talk about, like publicity rights. I do think organizations of any size really should think about a good AI policy—one that brings in stakeholders and isn’t just something you’re imposing on everybody, but that you get people to buy into—really sitting down with those who are using it and creating a good, flexible policy you can live with and work with.

Kevin Bankston: Great. Well, thank you so much.

Lisa Oratz: Thank you for having me—it was a lot of fun.

Kevin Bankston: Hopefully I’ll see you next year!

---

Keep Your Enemies Closer? A24 and Lionsgate Ally With Google and Runway as Competition in AI Video Heats Up

It’s as if the Hollywood-AI tech prom is coming up, and everyone’s scrambling to lock down a date before the music starts.

In past issues I’ve talked about how Netflix is clearly gearing up internally for a robust AI-assisted production pipeline, between its acquisition of Ben Affleck’s startup InterPositive and its multiple (1, 2) open source AI video tool releases. I’ve also highlighted how MGM is leveraging the cloud AI services of its corporate parent Amazon, including its video production platform Project Nara, to turbo-charge its production efforts.

Now, studios that aren’t already a tech giant or owned by one are finding strategic partners to better survive the competitive challenge that is coming as AI becomes deeply integrated into the Hollywood production ecosystem.

In the few weeks since issue #4, two major strategic partnerships have been announced between traditional studios and AI labs, demonstrating the ongoing and rapid Hollywood vibe shift from treating AI as an enemy, to a subject of cautious curiosity, to now an aggressively sought-out partner. It seems Hollywood isn’t sure if it’s at war with Big Tech or wanting to get in bed with it or both. And while studios apparently have been quietly using AI for a while now—Deadline reports how one AI tech vendor claims that it has been credited in only one-fifth of the projects on which it’s been used—only now are big public alignments beginning to surface.

First, Lionsgate and Runway went from casual dating to going steady. Lionsgate was already the first major studio to partner with an AI lab when it teamed up with Runway in 2024. That partnership didn’t add up to much, but apparently the relationship was good enough that both wanted to invest more deeply in it. Just as our issue #4 was going to press, Runway announced that Lionsgate was taking a (no cash, just shares) equity stake in the AI video startup.

Then, less than two weeks later, the Wall Street Journal exclusively reported a similar deal but in the opposite direction: Google investing $75 million for equity in upstart studio A24, pairing the indie darling behind Moonlight and Everything Everywhere All at Once with Google DeepMind’s researchers in a partnership to build new filmmaking tools.

The second deal got the bigger headlines, and the bigger reaction as A24’s dedicated indie-film fanbase went apeshit with subreddit meltdowns and canceled club memberships putting A24 on the defensive. But details on the deal were sparse and it’s unclear what the partnership will actually produce for either party. $75 million is pocket change to Google, as is really the entire Hollywood production ecosystem as a potential market. $75 million is more of a big deal for A24, but it’s already worth at least $3.5 billion based on another $75 million funding round from Thrive Capital in 2024, and like with the Lionsgate-Runway deal they probably just got shares from their new partner, not cash. (FWIW, I always suspected that the now-defunct Disney-OpenAI deal was similarly cashless, with Disney equity and its license for OpenAI to use its characters being its primary investment in the AI giant, and OpenAI stock and AI credits flowing the other way to Disney.) For perspective, Google’s investment is less than a quarter of the gross box office from A24’s current horror hit Backrooms, whose director Kane Parsons is an outspoken AI critic who recently called generative AI “cultural and economic rot” and, once the deal dropped, told his Discord that the Google money was “not to be spent on Backrooms.”

However, there have been signs for a while that A24 was looking seriously at AI as a way to make itself more competitive and better serve niche audiences in a cost-effective way, and this partnership is clearly the next step in that vision. In an excellent albeit hagiographic New Yorker profile of A24 from 2025, A24 Labs head Scott Belsky speculated about use cases including “a deep conversation with an LLM to debate a character’s mindset” and “pre-visualizing characters, costumes, and particular scenes.” Another top executive mused that AI could help the studio target “microcommunities”: “You can maybe make something that’s valuable for a thousand people, because the cost is cheap enough that it doesn’t have to reach ten million people.”

The vision of an auteur and audience-focused indie like A24 using AI to serve an exponentially wider range of niche tastes is a potent one, but for now, the entire venture is framed merely as research. Nor does the deal give Google any access to A24’s library for AI training.

The Lionsgate team-up with Runway is a much more intimate partnership, with explicit plans to stand up a joint program to build new IP, starting with AI-generated short-form series mined from Lionsgate’s vast catalog that includes John Wick and The Hunger Games. This isn’t just research or PR talk or playing footsie; this is a real partnership that will soon result in real content, and (Lionsgate hopes) major production cost-savings. And although the natural angle is to think of this as legacy content looking for a life-preserver, the greater value may be for Runway as it tries to distinguish itself in a crowded field.

Certainly, Runway appears to be leading in its space, and the Lionsgate deal is only one of many big moves it’s made in the past few months. There is of course Runway’s visible community leadership through its annual bicoastal Runway AI Festivals which were held last month in LA and New York, with established directors like Ron Howard as enthusiastic speakers and with indications that the quality of the screened short films is finally reaching a real tipping point. There was also the debut at May’s “AI on the Lot” conference of Runway’s Project Luxo, an initiative using a selection of short films to demonstrate just how realistic AI-generated video has become. Runway is also physically expanding, with new offices in both London and Paris, and technically expanding, with new MCP integrations in all the major chatbots including ChatGPT and Claude and a new partnership with NVIDIA to use the chipmaker’s most advanced supercomputing platform for Runway’s state-of-the-art world models (models that actually learn how the world works, unlike large language models). These world models could be the key to Runway’s long term competitiveness since they may not just be the next step in video generation but the next foundational tech for AI computing generally.

But until then, Runway is still one of many startups trying to make itself the default video generation platform for creative professionals. For example, similarly well-regarded and well-capitalized startup Luma, which The Ankler recently characterized as “the hottest AI video startup,” is hot on its heels. And since most companies in this space are focused on owning the interface through which you use multiple frontier labs’ models in addition to or rather than trying to keep up with their own models, there’s not much of a competitive moat and therefore a surfeit of contenders for the crown (including the frontier labs themselves).

As I highlighted in my own AI on the Lot coverage, and as IndieWire also recently covered (headline: “The Model Doesn’t Matter: Inside the Race to Be the AI Production Platform Filmmakers Want to Use”), the competition amongst companies targeting this creative workflow market is vicious. There literally seems to be a new entrant every week. Just three weeks ago, Chinese model-maker MiniMax launched Hub, its own unified multimodal AI video generation platform. And then two weeks ago, some ex-Googlers launched through their startup OpenArt a new video generation platform tool called Director, which (like most of these products) allows creators to engage a range of models using a unified chat interface.

The OpenArt founders call it “vibe directing.” But the main vibe I’m getting is that there are too damn many of these companies to survive for very long, not least because this also seems like a market that the major AI labs will simply dominate themselves very soon. If you’re mostly a wrapper for the big model-makers, those same model-makers will likely come for your business eventually.

Which brings us back to the importance of strategic relationships with content producers, where it may be that the hot AI startups need the legacy studios as much or more than the studios need them. Lucky for them, there are still several big studio players who haven’t found their date for the tech prom. Disney hasn’t found a new partner since getting jilted by OpenAI, and it’s also busy suing a bunch of these companies including image and video model-maker Midjourney. WarnerMount—the unholy union of Paramount Skydance and Warner’s whose merger was just blessed by Department of Justice leadership who ignored the concerns of its own senior attorneys—also hasn’t paired off with an AI video tech player, nor has Universal or Sony Pictures (although Sony did just zig while other studios were zagging toward AI partners, instead investing $100 million in shared-reality venue owner Cosm). Meanwhile, a dozen startups are probably starting to worry that they will end up as wallflowers at this particular dance.

We’ll see soon enough who ends up hooking up, and who’s left without a partner. Until then, the one thing we can be sure of is a lot of high-school level drama.

---

Hollywood’s New Short Story Gold Rush

As highlighted a couple issues back, I’m an occasional fiction writer—I was focused on short stories for a while, even getting one sci-fi story published, but have recently been focused on developing pitches for some comic book ideas. Based on what is happening right now in Hollywood, though, if I want to write something that’ll actually get adapted into film or TV, I should probably go back to short stories. (I won’t; comics are much more fun.)

I first heard about this trend from my brother the TV creative exec: short stories are the new hotness in terms of getting optioned, he told me when I visited LA in May, both because development money for original ideas is scarce, and because an executive is more likely to read a five-to-ten page story than a 110-page script. My niece who works in a literary agent’s office further confirmed that there’s a boom in story options happening right now. According to both, the stories often aren’t even published in a magazine or an anthology; they’re just directly getting shopped by agents to buyers, or discovered on the internet.

So I went hunting for evidence of this short story gold rush, and I didn’t have to look far to find it. In fact, this now several-years-old trend was exemplified just a couple weeks ago, with two different short story-based projects being reported on in the entertainment trades.

First was the short story “Exposure,” a psychosexual short story thriller co-written by rising TV scribes Erika Vazquez and Siena Butterfield (credits include Wednesday and CSI: Vegas). The story sold to Warner Bros. in “a competitive situation” with the authors set to adapt it themselves into a feature. Notably, the story was never published; it effectively took the place of a spec script, and was certainly written in a small fraction of the time a script would have taken. Other screenwriters looking for work are probably taking note and thinking of story ideas they could quickly hand off to their agents as unpublished shorts.

Second was the announcement of casting for Seasons, a rural horror flick based on a story only ever published on Reddit. That viral story by brothers Matt and Harrison Query, originally entitled “My Wife and I Bought a Ranch,” sparked a bidding war back in 2020 before getting locked down by Netflix for low seven figures (!). The project has since moved to Amazon MGM where Lily James will star with Drew Hancock (Companion) directing from his own script; heavyweight producers include Shawn Levy’s 21 Laps, Jason Blum’s Atomic Monster, and 12:01 Films’ Scott Glassgold.

Glassgold is a key figure in this trend, which arguably kicked off with that 2020 bidding war. At the end of 2024 he founded 12:01 Films in partnership with talent agency Verve, with an explicit mandate to mine short stories for IP using his success with the Reddit-sourced Seasons as a model. At launch he already had a bevy of short story-originating productions in development, including “The Dwelling” with Michael B. Jordan at Amazon, 50-page thriller “Long Lost” (sounds more like an unpublished treatment than a short story) with Steven Spielberg attached to produce for Amblin/Universal, twenty-page horror-sci-fi “I Am Not Alone” with Jessica Chastain starring and Misha Green writing-directing at Netflix, horror-thriller novella “We Used to Live Here” with Blake Lively for Netflix, and—by the same author—“The Caretaker” with Sydney Sweeney at Universal.

Notably, none of these stories were traditionally published (although several obtained traditional book deals to expand their stories into novels as part of or after their movie deals; “We Used to Live Here” became a bestselling novel by the same name). Those that were published at all before being optioned were published on Reddit, several of them on the popular horror subreddit r/NoSleep.

Like most movies in development, most of these probably won’t ever be made, but at least one of them is currently moving forward: Sydney Sweeney’s The Caretaker movie found its director, David Bruckner (The Night House), just a couple months ago. And that’s not the only Reddit short Sweeney is attached to; last year she also acquired thriller “I Pretended To Be a Missing Girl” to star in and produce.

Nor is Reddit the only online source of short-story grist for the Hollywood content mill: Vulture in 2024 covered the “booming subgenre” of Wattpad stories being adapted to TV and film, including Amazon Prime’s most popular original film worldwide in 2023, young romance My Fault (never heard of it). (For clueless uncs like me, Wattpad is an intensely popular user-generated story platform where people exchange micropayments for the next chapter, similar to the model webtoons or microdramas.)

Meanwhile, new media ventures are popping up to try to build their own short-story-to-production pipeline. For example, media startup Run-A-Muck, co-founded by Condé Nast alums, will be publishing short stories on its culture and fashion Substack Drafting with film, TV, and podcast adaptations in mind from the outset. Its first set of writers will include Cody Behan, whose psychosexual thriller short story “The Decorator”—again, an unpublished manuscript, only circulated as a spec package in Hollywood—went to Netflix for a TV series after a bidding war in March. (“Psychosexual” is apparently a common genre marker in these deals!)

But Run-A-Muck is just the latest of these ventures. For example, Glassgold of 12:01 Films has also started 12:01 Books, a joint venture horror imprint with Simon & Schuster’s Atria Books where hot shorts can be adapted to best-selling novels in addition to being optioned for film and TV (which is exactly the path taken by The Caretaker, the imprint’s first title.)

Indeed, there’s a whole coterie of startups focused on mining this short story boom, as highlighted by Substacker and short story writer “Max Winter,” a pen name for producer Chris Goldberg. In a recent Substack note, Winter identified several of these new players, including Sweater, a “development firm for story” founded by manager Thuan Dang, which publishes a magazine of curated original genre shorts sent to producers and studio execs with an eye toward optioning, and NeoText, founded by producer John Schoenfelder, which pays established genre authors to co-develop new IP with their producing team and then publish and (hopefully) option.

The upshot is, as LA Material quoted one producer in a trend piece on the short story options market in May: “It is a boom. It is a gold rush. It’s a thing.” The market is clearly hot right now for viral online shorts or even unpublished stories as part of a spec package (at least when written by a professional screenwriter and shopped by an agent with juice). And the reasons are obvious: it’s the least work for everyone involved to generate saleable IP, both for the writer and the buyer.

Finishing and shopping a full feature spec remains a valid (if brutal) path. Specs have slowly started selling again in the past few years, too, and for one of the same reasons: funds for developing original properties are slim. But building and sharing a portfolio of strong short stories—whether on Reddit, Substack, Wattpad, or elsewhere—may offer a better chance at landing a deal, since you can get multiple shots on goal rather than betting everything on a screenplay.

At CONVERGER we like to talk about the content singularity—the way the internet and emerging tech tools are collapsing traditional boundaries between formats and platforms. This short-story boom is a clean example: narratives born in the attention economy of Reddit or on independent Substacks are now steadily flowing as feedstock into the capital-intensive pipelines of Hollywood.

That trend will likely last for the foreseeable future, as long as development remains expensive and executive time remains scarce. So if you’re hoping to break into Hollywood as a writer, consider crafting a stack of good shorts instead of that big spec. Because Hollywood is reading—and optioning—more of them than it has in years.

Well, skimming, anyway.

---

FRAGMENTS

Take This Job and Shove It Train an AI To Do It

From blue collar to white collar to everything in between, the hot new job is training an AI to take your job. On one end, you have Indian garment workers wearing head cameras to record their complex hand movements; on the other, you have Meta tracking the keystrokes of its engineers before it lays them off (and then failing to protect the private data the surveillance generates). There are unsubstantiated rumors of OpenAI paying New Yorkers to record themselves doing household chores, and the reality of other startups collecting similar data. And finally, in Hollywood, you have refugees from the TV and film industries taking gigs doing RLHF training to pay the bills. Welcome to the wondrous new AI economy!

Prestige TV Writers are Really Freaked Out By AI

Speaking of the new AI economy: Hollywood anxiety about AI is widespread, but it’s especially been showing up lately in the work of TV writers, particularly in prestige comedies. Seth Rogan’s satirical show business comedy The Studio had a subplot last spring about a movie production trying to secretly use generative AI for special effects, resulting in Ice Cube leading a chant of “Fuck AI!” at San Diego Comic-Con; the third and final season of satirical show business comedy The Comeback, which ended in May, featured Lisa Kudrow’s bumbling but lovable TV actress heroine starring in a sitcom secretly being written by an LLM; and the fifth and final season of HBO’s satirical show business comedy Hacks, which also ended in May, had CONVERGER’s favorite comedienne Hannah Einbinder delivering an impassioned rant against the forced inevitability of AI.

However, as TheWrap highlights in a story on this anxious writerly trend, it’s not just the satirical show business comedies that are worried about AI, pointing to recent TV subplots about the challenges brought on by trying to integrate AI technology into emergency rooms (The Pitt) and classrooms (Abbott Elementary). Once the doctors and teachers on TV are nervous about AI too, then the anxiety has clearly left the writers’ room and entered the culture at large.

Amazon Dumps Luca Guadagnino’s Sam Altman Biopic

Films too are beginning to address our latest AI anxieties, but one particular film seeking to critique one of our emerging AI overlords just lost its distributor: after a yearslong artistic partnership that resulted in four feature releases including Challengers and Suspiria, artificial intelligence has come between Amazon and director Luca Guadagnino. Or, more specifically, Artificial: Guadagnino’s upcoming feature biopic about OpenAI’s Sam Altman, played by Andrew Garfield, which Amazon unceremoniously dumped just as it was nearing completion.

The timing is more than a little suspicious, coming only months after Amazon announced a $50 billion investment in a strategic partnership with Altman’s AI company. A jaded Guadagnino says he is unsurprised, saying in an Italian TV interview that “these industrial policies are not new” while highlighting his concern that AI “is completely changing the face—not just of society in terms of consumption and how we relate to these tools, but the very face of the identity of a place like the United States and the entire world—with the rise of this small oligarchy that wields truly radical control.” Thankfully, Neon acquired the film last week so hopefully we can see Guadagnino’s film soon, Amazon and oligarchs be damned.

Stop! That! Train! Director Who Denied Using AI Admits to Using AI

We talked in issue #4 about the strenuous but unconvincing claims from the director of new RuPaul comedy Stop! That! Train! that accusations of generative AI use for the film were “patently not true” and that “ZERO shots [were] conceived by AI.” Well, it turns out those denials themselves were patently not true. In newly surfaced comments made to an interviewer before that denial, director Adam Shankman made clear that generative AI was used in several shots of the titular train: “Some of the [train] windows were AI,” Shankman told Xtra. “There was only one AI shot or two AI shots of the train, the rest are CG. Most of the CG is the train, then out the windows is mostly AI.” So, more than zero shots. I honestly don’t care if filmmakers use generative AI instead of CGI for their special effects; if anything, I’m supportive of it. But don’t be a coward and lie about it. Say it loud: you use AI and you’re proud!

ByteDance’s Seedance 2.5: Now Tom Cruise and Brad Pitt Can Fight in 4K for 30 Seconds

More interesting than the news about the uprezzing and longer clip length, and relevant to the discussion with Lisa Oratz above about using expressive prompts for more authorial control and thus stronger copyrightability: the new model will allow up to fifty pieces of reference in your prompt to guide your video generation, whether images, videos, or audio files.

Tribeca’s AI Selection Stinks

So, that timely AI-generated feature about the Iranian protest movement that controversially was selected for the Tribeca Film Festival? Apparently it’s garbage, receiving an absolutely scathing review from AV Club: “Dreams of Violets is essentially an AI-generated snuff film, a supercut of people who don’t exist being subjected to executions that didn’t happen,” “where there’s no care for the actual art of filmmaking in any frame of this bloody sideshow” and where “common generative AI mistakes abound.” Variety’s reviewer was more measured but ultimately found the experience “stultifying,” “a barely scripted series of anecdotes, or mere moments, with little in the way of dramatic development.” Even The Hollywood Reporter, in a piece meant to showcase the writer-director behind the movie, had to admit the film “can feel disjointed in spots,” “[n]ot exactly anthology but hardly contiguous narrative,” with “discontinuities and some of those gauzy semi-unreal slop-shots, though not as many as you’d think.”

The consensus seems to be that some of the imagery was quite impressive for something that was completely AI-generated, but the story-less piling on of contextless vignettes of state violence against undeveloped characters was the opposite of dramatically effective. Perhaps the lesson—rather than “AI-generated movies suck”--is that AI-generated movies lacking a good script that makes you care about the people in it suck, just like with human-generated movies. At this point, more than better models, AI movies need better writers.

Hollywood’s YouTube Horror Buying Spree: Dumb Money or Smart Move?

Chasing the success of YouTube-sourced hits like Backrooms, Obsession, and now The Amazing Digital Circus, Hollywood is scouring YouTube for its next hits. Is this dumb money at work, or are new deals like the recent six-figure option for YouTube horror short “Open Door” a smart play? Just in the past few days, we’ve seen two big back-to-back deals: first, on July 1st, Warner Bros. won a bidding war (so–big money!) for viral horror meme IP “Siren Head,” created by creature designer Trevor Henderson. Director Zach Cregger (Weapons) and screenwriter Brian Duffield (Whalefall) are developing the property, which centers on a tall, thin, rotting creature with–surprise–two sirens for a head. Then the very next day, Amazon-MGM won another 11-studio bidding war to acquire the rights to viral YouTube horror series The Mandela Catalogue from creator Alex Kister who will also be co-writing and directing, with United Artists and Spielberg’s Amblin Entertainment producing. Big players, big money, big properties—but will they get big box office, or even get made? We’ll see soon enough.

That’s Microseries To You, Sir

The vertical microdrama megatrend continues apace, with a nice microdrama trend roundup from TheWrap identifying five key subtrends:

• The continued expansion of AI-generated microdrama content, and startups to make or host it (TheWrap also did a separate feature on the AI microdrama space, and in another fascinating story even had its reporter demo the technology by making four episodes of her own AI-generated microdrama romance in four days for $150 in credits; meanwhile The Ankler also had a recent piece on the use of AI tools in microdrama production);

• Vertical content 800-pound gorilla TikTok’s move into microdrama production, including the new hit Issa Rae microdrama thriller Screen Time and its partnership on a microdrama writing program with the Sundance Institute;

• Major studios jumping in on the microdrama action, exemplified by Fox Entertainment’s development of its first adult animation microseries from the production company behind Bob’s Burgers, its equity stake in Holywater and its microdrama platform My Drama, and its deal with YouTuber Dhar Mann for a slate of forty microdramas; NBCUniversal’s Peacock is also going all in on verticals, with upcoming unscripted microdramas built around its Bravo reality stars and a licensing deal for more microdramas from ReelShort to help supply content for its new vertical feed;

• The expansion of microdramas from the traditional outrageous soapy romantic dramas to other genres, from horror to reality to wildlife documentary; and relatedly,

• The apparent terminology change as everyone has started saying “microseries” instead of “microdrama,” almost as if they all received the same memo from corporate.

One additional trend was spotted by Variety at Cannes Lions: how branded microdramas and product placement in microdramas are proliferating as marketers glom onto the new medium. As one ad exec put it, “P&G invented soap operas, and this is the modern-day version.”

Clearly, microseries aren’t going anywhere—or, more accurately, they’re going everywhere. I guess if I’m going to have to keep writing about these things, which I clearly will, I probably ought to watch some of them! Any recommendations?

Don’t Want to Show Your Face in Your Video Content? Then YouTube Doesn’t Want You

Following up on our earlier story about YouTube’s heavy-handed anti-AI content moderation efforts demonetizing legitimate creators, it now seems that YouTube’s censorbot algorithms have adopted yet another regrettably blunt criterion. According to The Hollywood Reporter, certain AI creators are having discovery and monetization of their content throttled because they won’t show their human faces in that content. Turns out that if you want YouTube exposure, you now literally have to expose yourself. The new crackdown on faceless creators is even leading to a new market for hired hosts to show their faces in the videos of privacy-conscious (or self-conscious) creators who don’t want to reveal their own.

What YouTube Does Want: To Train on Your Content But Not Admit It

Google’s use of YouTube content to train its models is back in the news. We had confirmation last year that Google was using some unspecified subset of YouTube content to train Gemini and its video and audio generators. Now, Google is defending the practice in court, sort of, as it is sued by a bunch of indie musicians alleging that Google trained models like its music generator Lyria using their songs. Specifically, Google is arguing that YouTube’s terms of service give it a license to use posted content to improve its products, including its models, though it won’t actually admit that it trained on these particular plaintiffs’ songs.

I can’t argue with that as a litigation strategy, and on its face I think the terms of service probably do allow Google’s conduct, but the implications are staggering. Practically every musician big or small needs to post their music to YouTube to compete in today’s attention economy. If as a part of that bargain each one of them has to agree to provide their music for Google AI training for free, that’s an unfair bargain for them and an unfair advantage for Google over every other model-maker. Even Instagram and TikTok—who are certainly also training on their users’ content based on their terms of service—can’t rival YouTube for sheer amount of data.

Has Another AI Story Won a Literary Prize?

Yes, if this X poster and his Pangram results are to be believed. It appears that the winning entry in Harper’s Bazaar’s annual short story contest got a 100% AI-generated result on Pangram. Of course Pangram and other AI detectors are far from perfect, but incidents like this have many thinking that such testing should now be standard practice for editors. However, in the category of cutting-nose-to-spite-face, some editors are apparently wary of using AI to help avoid publishing AI because they don’t want to support AI by using it.

Another example of this self-defeating logic: a comic book anthology had to be withdrawn from competition in the prestigious Eisner Awards for comic books and graphic novels after it came to light that the anthology contained a one-page AI-generated story, done in the style of a classic comic book on which the model was trained. This happened despite the Eisners not having a policy against AI-generated work—though it is now instituting one—and even though the story (reprinted here) was clearly a meta-commentary on AI, creativity, and the public domain, with a coda noting: “That last story was clearly the work of artificial intelligence… But these humans clearly haven’t learned that no AI abomination can match the narrative beauty and genius of biological life!”

When your hatred of AI is so intense as to make it impossible to use the technology to detect or critique AI content, perhaps you’ve reached a counterproductive point in your anti-AI journey?

---

FLASHES

• I don’t care what their estates say, using dead people’s voices and likenesses when they didn’t consent while living–in this case, Gene Wilder’s voice for a Willy Wonka-themed reality show on Netflix–is just ghoulish.

• Cate Blanchett becomes the face of owning your face (and voice).

• Gore Verbinski wants AI disclosures so he can decide whether he likes your movie.

• The Dude digs Suno.

• So your music helped train an AI model…

• AI Answers are a code-red memetic threat.

• No, really, AI Answers suck.

• Can we stop trying to make Tilly Norwood a thing? “Her” producer team is developing a “feature” for the virtual “actor” to “star” in, where “she” will play (so meta) a bodyless AI being in the “Tillyverse,” a “surreal digital world located somewhere up in the cloud.” Uh, cool.

• AI-generated documentary Post-Truth was (intentionally) given a film festival award.

• Are Hexed’s filmmakers centering their compositions for easy vertical clipping?

• Speaking of composition, check out this interactive tool to compare the bajillion different aspect ratios in which you can experience Christopher Nolan’s The Odyssey.

• Speaking of aspect ratios, is Instagram going horizontal?

• What’s it like to be a breakout comic book writer? This fascinating behind-the-scenes Substack from Pornsak Pichetshote (Absolute Green Arrow) will tell you.

• Netflix is really done with theaters this time (we think?)

• Digital forensics researcher Hany Farid can’t believe his eyes in the age of AI.

• Yet another AI animation startup starts up, this time from former ABC chairman Lloyd Braun.

• Yet another AI everything: the map of AI entertainment players is getting very crowded.

• The newly-introduced CREATOR ACT, creating liability for using AI to create art that imitates someone else’s style, is well-intentioned but egregiously unconstitutional garbage. You don’t protect artists by punishing artists.

• Speaking of likely-unconstitutional garbage AI legislation, New York’s got some real doozies this season: the FAIR News Act (requiring journalists to label when they use AI in their journalism) and the Stealth Crawler Act (requiring anyone crawling a news site to identify themselves and the purpose for their crawling.)

• And California too, with AI transparency legislation raising serious First Amendment questions.

• Anthropic doubles down on its hypocritical stance against model distillation; friend of CONVERGER Derek Slater talks about distillation and the future of scraping and the open web on tech podcast The Center Edge.

• Retro Media Roundup: Everything old is new again, including zines, VHS tapes, Blu-Rays and PS1 aesthetics.

• Does cheap AI content production create more creative jobs? Yes, according to this hiring trends study and this survey of creators. Jevons Paradox at work! But then again, both were published by content creation platforms with AI integrations, Invideo and Adobe, who have a particular interest in one side of the narrative. Meanwhile a new survey of visual artists by Carnegie Mellon researchers shows over half of them reporting lost income and opportunities due to image-generating AI. So ¯\_(ツ)_/¯

---

And that’s what’s converging this week! See you next time.