From Script to Screen: Where AI Is Already Changing Filmmaking

Few topics polarize the film industry right now as sharply as artificial intelligence. On conference stages and in promotional videos, it sounds as if the fully automated film factory is just around the corner: text in, finished film out. Look more closely, however, and 2026 tells a different story. AI has long since arrived in professional productions, but only selectively, in individual crafts, and mostly in a supporting role. The central thesis of this article: AI does not replace the entire filmmaking process; it changes individual steps and shifts decisions, responsibilities, and power.

To test that claim, it is worth walking through the production chain, from the first idea to delivery on screen.

Pre-production: Text, image, and the limits of the machine

It all starts with the material. Generative language models can spit out ideas, treatments, or dialogue variants, yet the industry hemmed in how they are used contractually before it became routine. The framework agreement negotiated by the US writers' guild in 2023 states that AI is legally not a “writer” and that AI-generated material does not count as a literary work (WGA – Artificial Intelligence). A studio therefore cannot hand a writer an AI script and pay only a rewrite fee for it; companies must also disclose when material they hand over is AI-generated.

Behind this lies a very concrete exploitation problem. In its January 2025 report, the US Copyright Office made clear that purely AI-generated material is not protectable, and that prompts alone provide “no sufficient control” to establish authorship (U.S. Copyright Office, Part 2: Copyrightability Report). Assistive use of AI, by contrast, does not affect the protection of a human work. So anyone who has a film generated without a human hand risks not being able to protect it.

A range of further assistance features cluster around the text, and they should be classified cautiously in light of the available evidence. For research and fact-checking, language models advertise a fast pre-sorting of sources and background material; tools for automatic script analysis, covering structure, character arcs, or pacing, as well as for translating treatments and dialogue, are marketed too. Solid primary evidence for a robust, standardized studio use of these specific features is not available in this research; they should therefore be understood as assistive software categories on offer, not as proven practice, especially since the disclosure and control obligations of the writers' contracts apply here as well (WGA – Artificial Intelligence).

With concept art, storyboards, and previsualization, the pattern is especially clear. Research systems such as “CineVision,” developed at the Hong Kong University of Science and Technology together with Stanford and other institutions, link the script to a real-time previsualization and offer modules for lighting, character, and costume design (CineVision, arXiv 2025). Yet the authors name the limits unsparingly themselves: the system is confined to conventional dialogue scenes, cannot handle character movement or rig animation, renders only in low resolution, and delivers 2D stills without the metadata crucial for the camera, such as focal length or camera height. There is no automatic continuity check; camera operators have to translate the boards into shot lists by hand. It works as an idea sketch, but not as a finished production template.

For casting and planning, the real turning point runs not through the technology but through contracts. In casting support, an AI automation matters less than the question of consent: the SAG-AFTRA contract ratified in 2026 permits so-called synthetic AI performers only if they offer “significant additional value” compared with a living actor or that actor's digital avatar, and it enshrines a principle that expressly favors human performances (The Hollywood Reporter, 12.05.2026). Members approved it with 91.4 percent (Variety, 05.06.2026). Digital replicas of a performer require “informed consent” and fair compensation; a strike over the issue of synthetics, however, cannot be called until 2030. In production planning too, from scheduling to budgeting, there are AI offerings for automation; but a widespread studio use documented by primary sources cannot be demonstrated in this research. Virtual production must be strictly distinguished from virtual location scouting, the AI-assisted screening or generating of possible shooting locations during planning; virtual production means the real shoot in the LED volume and is no evidence of a robust AI location search.

Virtual production: powerful, but not the same as “AI”

A common misunderstanding concerns virtual production. LED volumes, of the kind found in many big productions, are not generative AI. The Unreal Engine documentation describes “In-Camera VFX” as an interplay of LED lighting, live camera tracking, and real-time rendering, with the goal of making greenscreen compositing unnecessary and producing final images directly in the camera (Unreal Engine 5.6 Documentation). That is game-engine technology, not an image generator. The requirements are steep: a fully virtual environment can call for a volume enclosing up to 270 degrees, and without precise synchronization visible image errors loom.

For digital characters, in turn, the MetaHuman framework has established itself. According to Epic Games it allows the creation and animation of “fully rigged, photorealistic digital humans” in Unreal Engine 5, with the MetaHuman Animator generating facial animation from video and audio recordings (Epic Games – MetaHuman Documentation). Here too the rule holds: a toolbox, not the push of a button.

Post-production: the real terrain of change

AI reaches furthest where it takes on individual, clearly defined tasks. The most striking example comes from VFX. In July 2025, Netflix confirmed it had used generative AI effects in an original series for the first time: a scene with a collapsing building in the Argentine production “El Eternauta,” created by its in-house team Eyeline (The Guardian, 18.07.2025). Co-CEO Ted Sarandos said the sequence was finished “10x faster” than with classic methods. That figure is a corporate statement and has not been independently verified; it should be read as a quote, not as a reliable metric.

Robert Zemeckis's feature film “Here” is similarly instructive. The company Metaphysic de-aged and aged Tom Hanks and Robin Wright in real time on set, so that the director could see the result immediately on a second monitor (Variety, 02.11.2024). Crucial for the thesis: despite the impressive live preview, Variety reports that the models were trained on licensed material and that human compositors and animators finalized the shots afterward in post-production. VFX supervisor Kevin Baillie put it in a nutshell: “We ensured that artists were involved to maintain fidelity to the actors' performances.” “No more post work” is thus an inadmissible oversimplification.

For generative video methods, research confirms this restraint. A survey article accepted at the Transactions on Machine Learning Research does credit diffusion models with better temporal consistency than earlier approaches, but still cites “significant challenges in motion consistency, computational efficiency, and ethical considerations” (Survey of Video Diffusion Models, arXiv 2504.16081).

Tools that land on the edit desk every day carry AI within them too. Adobe sells its Firefly Video Model, which powers the “Generative Extend” function in Premiere Pro, as, in its own words, “the industry's first commercially safe AI video generation model” (Adobe Newsroom, 12.02.2025). This claim of legal safety is a vendor assertion and is moreover in a beta stage, initially at 1080p.

In voice and dubbing, the assistance principle shows most clearly. The Oscar-winning “The Brutalist” used the voice AI from Respeecher to refine its lead actors' Hungarian pronunciation, in an iterative process with a Hungarian adviser and feedback loops, according to the vendor, without a fully AI-generated voice and without any change to the English dialogue (Respeecher – The Brutalist Case Study). Commercial dubbing tools follow the same logic: the ElevenLabs documentation describes transcription, editable translations, and voice cloning, but points to sync and timing problems and notes that the Dubbing Studio now runs in “maintenance mode” (ElevenLabs – Dubbing Studio).

In sound design, the assistance reaches into the shaping of the sound itself: the ElevenLabs dubbing tool offers, for example, the generation of sound effects from a text prompt, but stays within the same maintenance-frozen assistive frame and requires manual placement. With music, the bottleneck is less quality than rights: since June 2024 the US music industry has been suing the generators Suno and Udio over mass unlicensed training on protected recordings (RIAA, 24.06.2024).

A whole series of classic post-production tasks now goes under the AI label without any serious evidence of how mature they are. For AI editing assistance, from text-based editing through automatic transcription to suggestions for rough cuts, established tools advertise their wares, yet beyond the Adobe functions mentioned there is no solid primary evidence here of robust routine operation. The same holds for AI rotoscoping (the automatic isolation of objects), AI image restoration and remastering, and the machine translation of subtitles: there are numerous software promises, but in this research no primary evidence of proven, professional studio use; they should therefore be treated as tool categories on offer, not as an established standard. With trailers and marketing materials, too, the evidence remains cautious: Netflix said it is “experimenting” with AI-generated trailers, visual marketing assets, and natural-language search, a corporate statement that documents a trial run but no assured routine operation (Economic Times, 21.07.2025).

Little seen but real, finally, is AI in delivery. Netflix describes a neural “deep downscaler” that improves image quality during encoding and was preferred by around 77 percent of test subjects in preference tests (Netflix Technology Blog). That is not a creative decision but infrastructure, and precisely for that reason a good example of how quietly AI works where it is technically and clearly measurable.

Limits and risks

Three areas of tension shape real usage in 2026. First, verifiability: efficiency figures often come from the studios themselves, like the “10x” number for “El Eternauta.” Second, rights and data: what is created purely by machine is not protectable, and training data is litigable. Third, power and transparency: the collective agreements of SAG-AFTRA, WGA, and the union IATSE govern consent, protection against displacement, and compensation; IATSE, for instance, states that no one may be forced to enter prompts in a way that displaces colleagues (IATSE, 28.06.2024). The dispute over “The Brutalist” also showed that it is not the technology that ignites the debate but the lack of disclosure.

Conclusion

The verdict is clear and at the same time unspectacular. In 2026, AI has a firm place in film production, but as a tool in individual crafts, not as a replacement for the whole. It speeds up VFX sequences, refines pronunciation, transcribes material, and optimizes delivery. Yet wherever taste, continuity, legal certainty, or a credible human performance is what counts, the human being remains the decisive authority, and the new contracts cement that. The real shift is not happening on screen but behind it: in the question of who decides, who is liable, and who gets paid.