Interview Malte Hagener

How Is AI Changing Film Studies?

An Interview with Malte Hagener

August 13, 2026

This interview was originally conducted on January 10, 2025

How is artificial intelligence changing film studies, media production, and the humanities? In this interview, Andreas Sudmann speaks with film and media scholar Malte Hagener about the uses and limits of machine learning in research, the prospects of AI-generated cinema, the power of digital platforms, and the role of critical scholarship.

Andreas Sudmann: Malte, how does AI currently affect your work as a film scholar? Perhaps you could also say how this differs from its impact on your work as a media scholar.

Malte Hagener: To start with the question of how AI affects my work, I would say that I sometimes find it difficult to determine, perhaps contrary to public opinion, what AI really is and where it begins and ends. For example, does a Google search already entail using artificial intelligence? Does Google’s complex PageRank algorithm, which uses links and ideas from network science, already constitute artificial intelligence? I have to say I’m not quite sure. On a number of occasions, I am not even sure whether I am using AI or simply a computer application. I mention this to question the fixed notion people often have of AI. AI is when you go to ChatGPT, enter something, and get something out. Of course, there is machine learning. I would prefer to say machine learning, but I do not mind saying artificial intelligence. This is a brief preliminary remark about the point at which AI really begins.

With that said, how is AI affecting my work? At the moment, I am mainly looking into digital methods. Perhaps I should begin with the larger context. I run a large project funded by the Volkswagen Foundation together with partners in Frankfurt and Mainz, coordinated in Marburg. It is called the Digital Cinema Hub and concerns the implementation of digital methods in film studies. The project was devised before the current AI boom, which I think began in late 2022, when ChatGPT became publicly available. Of course, neural networks had already been developed for at least fifteen years, but that was when the boom truly began. The project examines digital methods in a very broad sense. These include many methods from the digital humanities, as well as fields such as computational geography, including GPS systems. What can we do with geodata in film studies? It considers different levels and applications.

Within this context, I am currently especially interested in network studies, a field between mathematics and computer science, but also cultural inquiry in the humanities in the broadest sense. I would say that systems in which you provide large datasets that are then processed in certain ways also border on the use of artificial intelligence. Again, I think that is debatable. Within that context, I am using AI. I also see people use AI methods as part of what IT specialists call a larger pipeline, where data is processed in several steps and one of those steps uses AI models. For example, within the Digital Cinema Hub project, we had a data sprint about a month ago, in December 2024. We worked with data drawn from catalogs of Super 8 films. Super 8 was a domestic format used in the 1960s and 1970s, before video became available.

You could buy films, usually reduction prints of films in circulation, but also sports, pornography, and other material. Anyway, we had these catalogs. They had been scanned, and we tried to convert them into tabular form. Doing that by hand would have taken a great deal of work, and we discovered that we could upload the PDFs to ChatGPT, which produced tables. They contained a number of mistakes, but were still easier to work with. Two or three years ago, we would have assigned this task to student assistants or partly done it ourselves. In this case, we used ChatGPT to convert the data into tabular form. It had its problems, and we learned a great deal, but I had not been aware that it could be used for this purpose. This is one example. At the moment, I see AI in the broadest sense used a great deal, although mainly for smaller, specialized tasks rather than broader analyses, because these systems do not work well for that.

You can also use it, for example, to generate questions, and then use those questions for brainstorming. In the broadest sense, I would use it for small steps in a longer chain of tasks needed to achieve something. That is where I currently see myself and others using AI.

Andreas Sudmann: You have already raised several issues, including the conceptual problems surrounding AI and the distinction between machine learning and other forms of AI. You also touched on the widespread use of large language models in the everyday work of scientists and scholars. At the same time, I sense some skepticism about their current epistemic potential. As someone with a large network in film studies, how would you assess perceptions of large language models and other AI tools within the field? Is there a general openness toward their use in film research, or does skepticism prevail?

Malte Hagener: In a way, I think it is both. There is deep skepticism, but at the same time, everyone is using these models. That is my impression. I think this is appropriate. We should use them. The main problem, to me, is that the available commercial models are black boxes. We provide the input and receive the output, but we have basically no idea what happens in between. It is somewhat different for the few open-source models, such as Stable Diffusion, the image model developed by Björn Ommer and his team. In academic research, we have an ethos that I consider appropriate and necessary: steps must be documented and reproducible. That is clear. Reproducibility is a major problem because every time you enter something into a large language model, or any model based on some form of neural network, you receive a slightly different answer. The answers are usually not completely different, but they still vary.

This question of reproducibility is a real problem. That is one reason for the skepticism. At the same time, I see and hear that a lot of people are experimenting with those things. I think that is appropriate too, because these models exist and everybody is using them. This is anecdotal, but two weeks ago I was on a train next to someone who appeared to be a student. She appeared to be writing her BA thesis. She had ChatGPT and her thesis open and was constantly moving between them. This is everyday reality for our students and, I think, for researchers as well. If you need a first draft of a text intended for internal use or for some official purpose, rather than a published research text or a grant proposal, it has become normal to have a first draft written by a large language model.

Then you work on it, of course. Previously, you might have asked someone on your team to provide a first draft, and then everyone would work on it. In that sense, I see pretty much everyone using these models and experimenting with them. I also edited a special issue of the Journal of Cultural Analytics. The issue included an article on natural language processing and sentiment analysis of film reviews. When applied directly to film reviews, the sentiment analysis did not work very well. The authors used ChatGPT as an intermediate step to rewrite the film reviews in simpler language. That approach is debatable, and the authors discussed it. But I think we have to try such approaches. We have to try these things and evaluate where they make sense and where they do not. It makes no sense to refuse their use entirely, because these models exist and everybody is using them.

It would also be wrong to assume that these models can do a great deal of work for us without critically examining the results. We need to be open toward these models while remaining very skeptical of them.

Andreas Sudmann: Let us move beyond large language models and consider AI and machine learning in film studies more broadly, particularly how they are changing the way we interact with and experience media. Media were already personalized and fragmented before the current deep-learning boom. Does AI reinforce this tendency, or are other developments more important when we examine how machine learning and artificial neural networks affect media and the media industries?

Malte Hagener: As you said, we already live in an environment where much of the available media is personalized through algorithms. To return to my first point, I would suggest that these systems have probably been using AI, some form of machine learning, or complex algorithms for at least a decade. They are too complex for us to reproduce. In that sense, we are already living in a world where media are personalized. What is being added now, I think, is that it is becoming increasingly easy to generate new content that looks human-generated. I am no expert on this, but I think there are already many AI bots, especially in the political arena. I have not been on Twitter for some time and certainly do not intend to use X, but it is probably full of AI bots. This is already everywhere. Increasingly, we encounter text, images, moving images, and sound generated by AI that are becoming so convincing in certain respects that they are difficult to distinguish from non-AI content.

I think that will push things even further. I am not sure in which direction, but this is a further step after personalization, which was an achievement of social networks and platforms such as Twitter, Facebook, Instagram, and TikTok. That is the situation as I see it. In that sense, indistinguishable AI-generated content represents another logical step. I view that with a great deal of trepidation, but there is not much I can do. As researchers, we can try to understand what is happening, address it critically in our research and teaching, and convey this understanding to students who may later enter public life. However, there are larger forces at play that are far stronger and have much greater resources at their disposal. We will see what this means for the production of films, television series, and music. In one sense, you could say that there has been a generic mode of producing music, films, and series for at least the past 150 years.

This modern mode of producing popular culture according to certain recipes is not entirely new. Of course, you still needed so-called creative people to write scripts, shoot films, and perform other tasks. AI may reduce the number of steps required. I still expect that, at least in the foreseeable future, perhaps within the next five years, AI will become a ubiquitous technology, but mostly in a modular fashion. It will be used for small tasks and steps in longer chains. We will not have an overarching system that takes over the world, or large parts of it. It will become more ubiquitous and less visible, as it already is. For example, in book production, I sometimes receive emails from people who tell me what I need to do next, and I am unsure whether the sender is a real person. The text is highly modular.

It is often in German but comes from India, so I assume it is at least automatically translated. I cannot tell whether I am communicating with people or bots. I think this ambiguity in superficial interactions will become more normal, while in other fields things will remain much as they are today.

Andreas Sudmann: We are now talking about creativity, a major topic in debates about generative AI that is closely linked to the technical and epistemic capabilities of large language models. More specifically, how do AI and machine learning affect human creativity and authorship in film and other media? You noted that the creative industries have always relied on recipes and generic procedures. The same applies to mass and individual communication, which, to quote John Durham Peters, have always involved forms of mass communication. It may therefore be inaccurate to draw a strict distinction between them. Against this background, which transformations in creativity and authorship do you consider most significant?

Malte Hagener: Well, that’s a very difficult question. If I had the answer, I do not know whether I would be sitting here. I would probably be running a company. The development is so rapid that it is difficult to say. I sometimes talk to people who work more closely with these models and ask whether anything genuinely creative emerges from them. Some say no. Others claim that they occasionally see surprising results and do not know how the system derives its output from the data and algorithm. So I think the jury is still out on that question. The question also applies to humans. Humans learn through experience. Children imitate endlessly, and that’s how they learn. In a way, you could say that learning involves imitation, followed by building on that imitation and creating something new. This is also a very human process. However, I doubt that human creativity will soon be matched by the models currently available.

These models will, of course, be used extensively. You might ask one to write a scene in which two friends discuss the love interest of one of them. It will produce a page, which people will examine and partly rewrite. I think these models are already widely used in such processes. Last summer, we already had the first song in the charts whose lyrics, I think, were written by a person, while the music and voice were generated by AI. I see extensive use in this kind of serialized commercial popular culture, and that will happen very quickly. We already see it in routine translations. People now use translation tools such as Google Translate and then revise the text, at least for basic translations that do not have literary ambitions. Literary translation is still very different. But pragmatic texts, as I would call them, are written for practical purposes, without particular aesthetic ambitions.

We will see a great deal of AI at work in this area over the next two or three years. The question remains whether this is genuine creativity and whether AI will generate results that many people consider interesting music, film, or television. In that respect, I remain doubtful. Yet I would not be entirely surprised if, within the next ten years, we saw something that genuinely impressed people. When ChatGPT was first released, I was genuinely surprised by how good it was, in terms of producing grammatically correct and coherent text. By now, we have also learned how standardized its output is in certain respects. We will see what happens with genuine creativity, which we do not fully understand even in humans. As I said, I remain doubtful that AI will reach that point anytime soon.

Andreas Sudmann: Again, I sense considerable skepticism about AI’s potential to create innovative forms of cinematic expression or aesthetics. From your perspective, is it realistic to predict that, in the not-so-distant future, we may see entirely new genres or styles of filmmaking, or is that a familiar fantasy about AI? I am also thinking of current discussions about AGI and Sam Altman’s claim that OpenAI may produce such a model this year.

Malte Hagener: I think Sora, their moving-image model, has still not been released to the broader public. That is taking much longer than initially expected or announced. This shows how difficult the task is. The more specialized the models are, the better they usually work. With good training data for specific tasks, we will see many successful applications. As for new genres, I would expect people to work in unexpected ways with the help of AI. I do not expect AI alone to generate them. For example, the Eye Filmmuseum in the Netherlands had a bot that used archival material, often films without copyright restrictions. You could give the bot a prompt, and it would generate a film. It produced essayistic or experimental films, using object-recognition tools to assemble footage. It was programmed in a particular way and required a defined pool of images. In those terms, I am certain that we will see unexpected results over the next few years.

Human intelligence will remain involved in assigning the AI specific tasks and delineating a corpus of material on which it should work. Such corpora can, of course, be far larger than humans can survey. I do not expect a machine into which you simply type “make a great film” and that then computes for a few minutes and produces an interesting film. I do not see anything like that in the foreseeable future. In combination with humans, however, there is considerable potential. Again, I return to my point that, at this moment, we mostly use AI in a modular fashion, for specific tasks within clearly delineated areas. Another example is that, ten years ago, people were predicting that cars would drive autonomously. This has proved much more complicated. It may now be possible in familiar and controlled areas, but the world is enormously complex and multilayered.

Until AI can understand and respond to conditions beyond a delineated area, I do not see it as operating on the same level as human intelligence. At the same time, of course, any AI can work with a billion images quite easily. That is something that a human is just not capable of. This gives AI a very different capacity for working with vast quantities of material. The combination of what humans do well and what machines do well is the most interesting aspect, at least to me.

Andreas Sudmann: I would like to talk a little more about films produced entirely by AI. At a conference in October, one of the speakers demonstrated how easy it has become to produce three minutes of video using AI tools alone. At present, however, AI-produced films still have a distinctive, imperfect aesthetic. They look somewhat strange, much like the portrait of Edmond de Belamy, which was central to debates about creative AI a few years ago. The discussion has since shifted to text-to-image models such as Stable Diffusion and Midjourney. Videos and films still do not meet our expectations or master the imitation game as convincingly as language models. Do you find these specific film aesthetics interesting as a subject in their own right? How might studying them help us rethink the role of AI in film studies?

Malte Hagener: In a way, we can look at everything that has happened over the past 30 years, right? Toy Story came out in 1995, which we might regard as a watershed moment. In computer animation, for example, hair, fur, surfaces, and the ways light is reflected or absorbed are produced using programs. I am uncertain whether to call these AI or conventional algorithms. This returns us to my first point. We are dealing with a long development, and all of it becomes part of the model. Producing these films will become easier. Making them meaningful will still require human input. I am sure that amusing short films generated entirely by AI will appear over the next year or two, but they will rely on prompts and prompt engineering. The results often emerge from many iterations. There is always a human in the loop, examining the result and asking for another iteration with more light here or a red dress there.

In that sense, a particular mode of filmmaking will evolve. Small teams may produce impressive-looking films, which will affect the large visual-effects and programming teams currently assembled. At the same time, I think we will increasingly see hybrid films combining material shot in studios, perhaps against blue or green screens, with material produced using AI. That is already largely a reality of film production. This will intensify. In the foreseeable future, by which I mean perhaps three to five years, I do not expect films shown at the local multiplex to be made by only a handful of people. There may be such films at festivals, and I am sure there will be examples. Still, from what I see, producing three minutes of film requires substantial computing power and extensive prompt engineering. This cannot be done in an hour on a laptop. It still requires a great deal of expertise.

It requires substantial computing power and specialists. In that sense, I see a lot of changes, yes, but again, in hybrid models where humans work very closely with AI models on many different tasks.

Andreas Sudmann: Would you also be skeptical of the idea that AI could democratize filmmaking by making it more accessible beyond the concentrated power of large studios and technology companies, allowing individuals to produce aesthetically interesting or innovative films?

Malte Hagener: Yes. Overall, yes. I am sure we will see films produced with the help of AI that become internationally visible. But there is an entire apparatus behind distribution, marketing, and placement. Film history, and now the history of the large platforms that have grown over the past twenty years, shows that controlling distribution is usually the key to controlling media. Production matters less in this respect. It will become easier to produce content, as it has over the past thirty years. Today, smartphones and consumer cameras can produce impressive results. Yet the key issue is distribution: gaining access to cinemas or becoming visible on platforms. Media history repeatedly shows that controlling platforms and distribution channels is far more important than controlling production.

Andreas Sudmann: From a film studies perspective, one would usually distinguish between distribution and exhibition. This distinction may have been even more important in the 1990s than it is today, when films can be made available on many platforms without major barriers. A YouTube video can, at least in principle, reach a large audience. By contrast, independent filmmakers seeking access to the Hollywood industry once depended heavily on admission to film festivals. Might changing infrastructures create new openings, including for AI-based filmmaking, or do established forms of gatekeeping remain largely intact?

Malte Hagener: Of course, videos by completely unknown people go viral, mainly on YouTube but also on other platforms. TikTok is another example. I would be hard-pressed to name someone who became established as a major filmmaker in this way, working in a major industry such as Hollywood, Bollywood, or the Chinese film industry after emerging from nowhere through a platform. Some people use these platforms, but they then often attend film schools. They become involved in financing arrangements with large studios and producers. In that sense, a hierarchy remains. One thing I am particularly interested in is the role of streaming platforms such as Netflix, Disney+, Amazon Prime, Apple TV+, and Paramount+. These platforms matter because they reshape what is available, and much of today’s content goes through them. In the medium term, I do not think all of them will survive. Again, distribution is crucial.

YouTube will remain, but the vast majority of videos on it will receive fewer than one hundred views. This has been discussed in research. Geert Lovink, for example, wrote a book called Zero Comments about how most material placed online never receives comments. There is a particular logic at work. The platforms capture much of the value generated by popularity. I am not as optimistic as Henry Jenkins about the democratic potential of these platforms. What we are now seeing with Elon Musk and Mark Zuckerberg accommodating Trump demonstrates that these are capitalist platforms seeking profit. They have no intrinsic interest in democratic structures. They invoke democracy when it is advantageous, but ultimately they run businesses and want to make money.

Andreas Sudmann: What do these observations mean for film studies and its contribution to understanding and critiquing the increasing use of AI in media production and consumption? What can media and film studies contribute to analysing AI’s new role, especially given its close connection to the growing power of digital platforms?

Malte Hagener: I would identify three potential areas in which film studies can contribute. The first is to remind people of the historical background. This relates to what I have already said. We know the role distribution has played historically and how important it has been. This also applies to claims about novelty, which have accompanied many technological inventions in the past. The rhetoric of novelty and revolution is itself very old. We need media and film history to understand this. Second, we need theory to contextualize these developments and critically examine the results. As I understand it, this is also part of the work you are doing in your project. This is very important. I would add a third area. I am also very curious and excited about what is possible. I have mostly emphasized skepticism, but there are also opportunities. We should experiment instead of merely sitting in our armchairs and warning people that everything will be dangerous.

Of course, we have a critical role, and I see myself in that role. We should also try these technologies and determine how we can work with them, without forgetting that the basic academic stance is one of skepticism. We do not accept everything at face value. That is particularly important. We should also collaborate with IT specialists and train models ourselves to develop alternatives, possibly open-source models that we can understand better. I am currently working with a specialist here in Marburg who studies explainable AI, and we are discussing areas in which we might collaborate. We should not overlook that this is a new field we can use productively in research and teaching.

Andreas Sudmann: Your answer leads to my final question about AI’s impact on the humanities generally and the digital humanities in particular. Given your expertise in digital methods and infrastructures, where do you see the most significant changes in how these fields understand and work with AI?

Malte Hagener: I think the humanities as a whole must take this seriously. It is a challenge, and it is not easy. But if the humanities want to survive, they have to take it up. They cannot ignore it. Nor can we claim that it is all wonderful and will solve every problem. I sometimes encounter that attitude. I am currently teaching a film analysis class in which we use a tool with AI-based plugins that analyze color, sound, and other features. It was interesting to see the students‘ reactions. I gave them a short assignment asking how they could use these tools. They initially said that AI would solve many problems and was wonderful. They then discovered that, at present, it is not very good at recognizing speech reliably, among other limitations. We encounter many practical problems. This also makes it clear how complicated these technologies become once you examine them closely.

Nevertheless, we have to work with them. We have to bring our strengths, including critical, hermeneutic, historical, and theoretical thinking, to bear on a changing world. The humanities will face serious problems if they ignore AI or approach it with naive optimism. We have to take up the challenge critically and with curiosity. The humanities will then have an important role to play in the future.