How artificial and how intelligent is artificial intelligence for film?
The roundtable at the Internationale Kurzfilmtage Winterthur is the occasion to raise questions on artificial intelligence and its application to film creation: questions about its language, customization, creativity, intellectual property, democracy, human labour, ecological impact and more, in order to plead for a wide appropriation of the technology and overcome the fear of the (human-too-human) machine.
Text: Giuseppe Di Salvatore
«Rage against the Machine?» is the title of the roundtable proposed by Gabriela Seidel-Hollaender at the Industry Lab of the Internationale Kurzfilmtage Winterthur. Rage or fear of the machine? When Gabriela proposed that I moderate the roundtable, I was particularly happy, probably also because I would have “moderated” the fear with my cautious optimism and frank curiosity towards artificial intelligence (AI). Even if I got some ideas of the basic elements of the technology during my epistemological studies at university, I am ignorant enough today to be simply thirsting for knowledge. However, there is neither a stable state of the art when it comes to artificial intelligence, nor about its application to film production and film creation, simply because the development of the tools and the results of their use are constantly changing through a practice that is definitely more empirical than scientific. Therefore, there is but one good attitude for me, that being: raising questions, sharing knowledge and practices and, once again, raising questions.
Why do we say “prompts”?
After Patrick Karpiczenko’s formidable and entertaining Keynote Speech, I was joined by three knowledgeable guests, Yun-Hua Chen, Gloria Gammer and Margaritha Windisch, who bring their experiences in the fields of, respectively: film criticism, screenplay and filmmaking, and intellectual property and copyright (here is the link to the discussion). We started speaking about the language, the technical language of artificial intelligence: why do we say “prompts”, and not “tasks” for the users’ main interaction with the machine? Is it because a Danish computer designer, Jakob Nielsen, introduced a user-friendly revolution in our relationship with the machine, and proposed not saying to the machine what to do (giving a task), but to say what we want (producing a prompt)? This new approach fits perfectly to AI machines, insofar as their neural networks are not designed to react only to specific inputs but to interpret the inputs by making them “circulate” in a broad context of information. Will the users so reduce their responsibility in giving prompts and not specific tasks to the machine though?
English language and the challenge of customization
Remaining on the topic of language, how much does the English language influence the learning potential of AI machines? Translation is one of the most used functions of AI: could it become an obstacle for new AI models that would have their training in another language, in Chinese for example? Open source “food” for the machine is largely in English, but countries without effective privacy regulations, like surveillance dictatorships, might equally have access to a huge amount of training material, which would render the AI models efficient... but is quantity (of training) the only criterion for the machine to be more powerful? Is a “know-it-all” model the only one that can cope with the demand for customization? Are custom-training software necessarily powered by all-competent models, or could we imagine models that from their birth focus on specific competences?
Is AI creative?
This question brings us back to film and the creativity of filmmaking: do AI tools tend to reproduce standard and popular imagination even if our prompts ask for more original outputs? How much and how fast can an AI interface learn from the artistic personality of each creator? How much can creators push the machine to unlearn from them in order to surprise them with original, non-standard outputs?
In the domain of the moving images, AI is already largely used for text-based practices, from fund-raising applications to screenplays, from character profiling to film criticism. For the moment, it seems to work better for form than content, better for standard formats than original ones. Is it simply a question of time? Will AI compete with film writers for artistic personality? How could we recognise the artistic signature of an AI? Who will be the signature’s owner? The AI developer? The collective body of people agreeing to make their content an open source? Or the prompt creator, the user of the AI machine?
Signatures and ownership
Margaritha Windisch tells us that, in the field of image creation, Ben Zhao and his team at the University of Chicago have launched a new tool that is able to “poison” the AI model that eventually takes protected images without permission, thereby forcing the theft to emerge. Could such tools be useful to track the path to the source, in this way allowing a better protection of the intellectual property and/or the possibility to brand signatures? And, more generally, how can a system of control of the AI models work if AI regulations are still dependent on the legislations of singular nations? Almost all free AI tools require the users to give their outputs to the machine: what about the artistic outputs? For example, a software like Promethean AI focuses on artists and is free for non-commercial use: should it therefore be legitimate for artist-users to ask to be paid for their contributions?
AI for a new film literacy – at what cost?
Text-to-Text AI tools are very effective, but also Text-to-Speech ones and increasingly even Text-to-Image tools like Midjourney, Dall-E, or Stable Diffusion. Text-to-Video tools are booming right now, from HeyGen to Runway – and Synthesia Lab is preparing a 3D video creator. Should we imagine the possibility of emancipation from text inputs and, perhaps in the near future, have an Image-to-Image AI tool, or even a Video-to-Video AI tool? This would indirectly support image and film literacy (for the pleasure of the Warburghians) and enhance film creativity. Still, we can also imagine how heavy the calculation power would be when needed for a Video-to-Video AI tool. Should we also raise the ecological question of the AI development? We can do no more than wait for studies that could measure the green impact of film production sparing locations and human labour in order to let AI machines do as much of the work as possible.
How artificial is AI?
If we consider that the training time used for a large and rich company, such as OpenAI in order to launch ChatGPT-3, has been a year and a half and consider that such neural network models requires a part (more or less a third of the model) of supervised development, one should not forget the human labour that is hidden behind the machine. Human content, human labour: how really artificial, or automatic, is AI? When we insist on the autonomy of the machine, aren’t we under the influence of a technological myth, which paradoxically should be closely analysed in terms of human labour and human choices? Shouldn’t we duly scrutinise the almost saturating presence of very large companies with their more or less hierarchical structures?
Is (or should be) AI democratic?
The wide contribution of users for the development of AI models whose access is free has created the prejudice that AI tools are democratic. Should the fact that these tools are slowly becoming less and less free make us doubt whether free access has been just a phase of AI development during which the programmers of the models needed free material to train and test the models? How transparent is the decision-making in the modelling of the AI tools? How transparent can it be if the main developers are private companies? A true democratic challenge would be constituted by the free access of tools that can allow groups of people to become not only users but makers of new AI models. In this scope, a basic technological alphabetisation should be taken into account at the level of public education, which would definitely make the absurd topic “fear of the machine” disappear. AI is far from being non-human, even if it can certainly surprise us, and its outputs can be properly creative. In order to make it fully democratic, we should start from opening our eyes concerning its human(-too-human) features. Only through a wide appropriation of the technology can we raise the otherwise obnoxious ethical questions concerning censoring or filtering. Moreover, could any limitation to AI machines truly be productive if a certain degree of unpredictability is intrinsic to the specificity of the machines themselves?
How intelligent is AI?
I know, it is quite difficult to remain focused on the film specificity of AI applications, but we are still in a period where understanding the general issues of the technology is fundamental in order to guide and imagine the development of its empirical application. For example, a last question should be raised concerning the type of intelligence that is at stake in AI: if one throws a closer look at the functioning of the neural network models and their learning capacity, reasoning and problem solving seem to describe what intelligence means in AI, and yet intelligence can’t be broken down to mean only an all-learning brain. Evaluating relevance and imagination, for instance, are two skills of intelligence that are not reducible to reasoning and solving problem. Furthermore, if we can speak of a “filmic intelligence”, we should look for alternative artificial (or non-artificial) intelligent models that are still to come – as Ben Vickers and K. Allado-McDowell’s Atlas of Anomalous AI, for example, has started to sketch out.
And, in the end, what about the intelligence of raising interesting questions? Is AI capable of raising interesting questions? If some of the questions I raised will be interesting, I will have probably proved that this text has not been written with AI…
Rage against the Machine | Industry Lab | Internationale Kurzfilmtage Winterthur 2023 | 10/11/2023
Patrick "Karpi" Karipczenko, Yun-Hua Chen, Gloria Gammer, Margaritha Windisch, Giuseppe Di Salvatore
First published: December 02, 2023