Creative perspectives on AI vary vastly across the globe. We're at an evolutionary stage where new tools are being adopted, advancing the creative workflows in VFX, film, media, and gaming. Do we know how films are made in the age of AI? Where has it normalised into standard practice? Let's start with some definitions and a timeline, then analyse how we got here.
Definitions
The conversation around synthetic media is noisy and often imprecise. These six terms build our language.
Synthetic Media — content (image, audio, video, text) generated or significantly altered by AI rather than captured from reality.
Synthetic Data — artificially generated datasets that mirror real-world information, used to train AI models. Not for human consumption — for machine learning.
AI Replica — a synthetic likeness of a specific real person. Voice, face, or full performance. Restorative (completing existing work) or generative (creating new work the person never performed). The most visible and contested output of synthetic media.
The Realism Threshold — the point at which an unassisted human eye or ear can no longer reliably distinguish a real recording from an algorithmically generated one. In 2026, we have crossed it.
Epistemic Shifting — a legal and political strategy where real, damaging evidence is dismissed as synthetic. Also known as the Liar's Dividend.
Model Collapse — what happens when an AI model is trained on outputs from other AI models rather than authentic human data. Quality degrades progressively. The content becomes a caricature of itself.
The chain: synthetic data trains the models. Synthetic media is what those models produce. The AI replica is the most contested output. Model collapse is what happens when the system feeds on itself instead of the source.
Timeline
Seven moments that mark cultural and perceptual shifts — not software updates.
1991 — The Desire
The T-1000 wasn't built pixel-by-pixel to look like a background prop. It was the first digital character designed to mimic, stretch, and flow like a living human. The creative industry didn't just want to enhance reality — it desired to synthesise it from scratch. The tools weren't there yet. The desire was.
2001 — The Uncanny Valley Named
Final Fantasy: The Spirits Within was the first multi-million-pound attempt to build an entire cast of photorealistic synthetic humans. It never reached mainstream success across global audiences — not because the technology was flawed, but because it was too close. By mimicking humanity right up to the line, it triggered an instinctive, biologically programmed rejection. The uncanny valley moved from academic robotics papers into popular culture, and the industry spent the next two decades trying to cross it.
2017 — The Trust Problem Scales
The camera had always been capable of lying, but before this moment, deception required meticulous manual execution. When code left the academic sandbox and open-source face-swapping tools hit public forums, the barrier to manipulation vanished. The moving image and the human voice lost their default status as objective truth. This is the moment the epistemic contract was breached — not because media dishonesty was newly invented, but because a sophisticated lie became instantly democratised and infinitely scalable.
2019 — The Manufactured Icon
Lil Miquela and Noonoouri proved a cultural icon didn't need a human heartbeat. Millions of followers, luxury brand campaigns, record label deals. Synthetic media shifted from an unsettling trick into a highly profitable corporate asset. A character with no birth certificate, a dedicated management team, and zero bad days.
2022 — The Sovereign Creator
Midjourney, DALL-E, Stable Diffusion. Anyone with an internet connection could generate gallery-grade visuals in seconds. The technical friction of executing art vanished — at least for some. Making art became cheap. The eye for what compels didn't.
2025 — The Infinite Continuity
AI characters stopped morphing between cuts. The same face, wardrobe, and lighting physics held across scenes. The systemic hurdles to long-form synthetic storytelling were completely dismantled. Directors could finally maintain character permanence across multiple environments without the visual seams showing.
2026 — The Probability Paradigm
Generative video enters the primetime pipeline — not as a novelty, not in the credits footnote, but as infrastructure. The modern studio no longer builds realities brick-by-brick. It steers a probability engine. The calculation: based on everything known about light, motion, and cinematography, what should the next frame look like? It dreams the answer into pixels.
How It's Being Used
Synthetic media isn't a monolith. It is a spectrum of methods, each sitting at a different point in the production pipeline.
Restorative Synthetic
Helping an ageing actor complete a story, preserve a legacy, or sustain an iconic role. Think of De Niro and Pacino in The Irishman, a young Harrison Ford opening Indiana Jones: Dial of Destiny for 25 unbroken minutes, or James Earl Jones authorising his Darth Vader voice to continue in Obi-Wan Kenobi after his retirement. Here, the technology serves the original work.
Hallucinated Synthetic
Generating content a person never actually recorded: new choreography, new words, or new songs from artists who never performed them. Tupac at Coachella saying "What's up, Coachella!" to a crowd in 2012 — things he never said in his lifetime. Or the K-pop resurrections on South Korean television, where deceased artists sing modern songs alongside surviving bandmates in front of their tearful families. This is where the deepest legal and emotional backlash originates — not because the technology is wrong, but because the consent framework didn't exist before the output did.
Native Generative
Final pixels produced directly by foundational AI models — not post-production, not enhancement. The imagery is born synthetic. Take the production pipeline behind Anyma's ÆDEN show at Coachella, where Google's Gemini was embedded into the creative process to rapidly generate thousands of design iterations, conceptual world-building frames, and visual aesthetics directly from natural language prompts. When the distinction between "filmed" and "generated" stops being useful, foundational models shift from mere tools into core creative collaborators.
The Spectrum in Live Events
No industry has moved further or faster than live performance, because the audience is physically present and the technology has nowhere to hide.
Traditional Stage
Tupac at Coachella in 2012 used a Pepper's Ghost optical illusion to project a synthetic performance onto a physical stage. The technology was analogue; the asset was hallucinated.
Immersive Dome
ABBA Voyage built a custom arena in London to house a permanent synthetic performance — ILM's ABBAtars performing as their 1979 selves nightly, indefinitely. The Wizard of Oz at Sphere (opened August 2025) pushed this further. Google Cloud processed 1.2 petabytes of data to upscale 1939 Technicolor to 16K × 16K across 160,000 square feet of dome. Veo 2 and Imagen 3 extended the canvas so characters stayed on screen and the landscape breathed. Whilst critics described a "poreless, synthetic sheen" on Judy Garland's face, the estate sat in the audience and marvelled.
Spatial Virtual Production
The final frontier of live performance maps physical space directly into a generative digital asset. Instead of relying on rigid, hand-built CGI geometry, modern spatial workflows use real-time tracking to seamlessly bridge physical stages with generative AI engines. By integrating real-time generative models, directors can prompt camera-ready, layered 2.5D plates onto LED volumes instantaneously. With the industrial-scale adoption of 3D Gaussian Splatting, production teams can now capture real-world locations and stream them as fully navigable, volumetric neural environments. The tracking infrastructure dictates the spatial perspective, but generative synthetic media creates the world.
The Backlash
When technology moves faster than consensus, structural friction is inevitable. The expansion of synthetic media has not been met with universal acclaim, but with an organised, systemic counter-offensive spanning labour unions, legal frameworks, and the mathematics of the internet itself.
The Labour Counter-Offensive
The existential anxiety of the creative workforce has solidified into historic industrial action. Frontline unions — such as SAG-AFTRA globally and Equity in the UK — have restructured their entire collective bargaining frameworks around the concepts of digital replication, synthetic performance rights, and mandatory consent. The battle is no longer over fair wages for physical presence, but the lifetime ownership of a performer's biometric data and digital likeness.
The Epistemic Collapse (The Liar's Dividend)
The societal cost of democratised media manipulation has yielded a dark paradox. It is not just that bad actors can fabricate convincing lies; it is that authentic public figures can now plausibly deny undeniable truths. By exploiting public awareness of generative tools, individuals caught in compromising positions can simply claim the real evidence against them is "synthetic." The default authority of the captured image has dissolved, leaving a fractured informational ecosystem where objective truth is a matter of belief rather than verification.
The Technological Ouroboros (Model Collapse)
The ultimate threat to generative AI is not regulatory, but algorithmic. As the open internet becomes flooded with synthetic text, imagery, and video, foundational models are increasingly being trained on AI-generated data rather than authentic human output. This feedback loop triggers a mathematically proven phenomenon known as model collapse. Denied a continuous stream of organic human creativity, the algorithms degrade, losing their statistical nuance and regurgitating distorted, hyper-generic artefacts. The synthetic ecosystem is learning a harsh architectural truth: without a thriving foundation of authentic human data, the machine inevitably starves.
The Tactile Counter-Culture
Photorealistic synthetic media is now cheap and infinitely scalable — it has completely lost its luxury status. The avant-garde response is intentionally unpolished and fiercely tactile. Creatives are pivoting toward childlike animation, raw textures, visible physical craft, and imperfect hand-drawn lines. This isn't a regression in technique, but a deliberate deployment of friction — visible, undeniable Evidence that a human was here.
The Tools
The creative stack has three distinct layers, each serving a different class of maker.
The Clay (Stable Diffusion)
An open-source foundational model. It generates images, textures, and custom character elements. This is raw mathematics that knows what things look like, but fundamentally doesn't care how they sequence. Everything else is built from this substrate.
The Factory (ComfyUI)
A node-based visual pipeline offering infinite control — frame-by-frame masking, multi-model chaining, and custom logic. This is precisely where "AI inside the pipeline" lives. Requiring high technical expertise, it acts as the operational connective tissue wiring everything together.
The Instant Director (Seedance 2.0)
ByteDance's multimodal engine. Upload image references, motion clips, and an audio track, and it executes a single pass: calculating physics, visual consistency, and lip-synced audio simultaneously. It delivers director-level output without pipeline knowledge. The professional end of this trajectory is ComfyUI embedded directly into Unreal Engine; the consumer end is Seedance. Same direction, different audience.
The Sculptor (Luma AI)
Where other tools treat video as a sequence of flat 2D images, Luma understands depth, volume, and physics. Crane shots, parallax, and complex lighting reflections are calculated, not hallucinated. It outputs studio-grade formats like EXR and HDR that VFX artists can colour-grade exactly like raw Arri Alexa footage.
The pipeline is moving inside the tool. Claude is embedded in Blender, Unity utilises Muse, and Unreal Engine deploys the Epic Developer Assistant. ComfyUI runs natively inside game engines, whilst Adobe Firefly is stitched into Photoshop. The tools professionals already rely upon have simply become intelligent. There is no new interface — the tools themselves evolved.
Where It's Heading
The Paradigm Shift
Traditional VFX required human artists to build reality brick-by-brick using digital clay. Generative synthetic media requires human curators to steer a statistical probability engine. The modern artist is no longer an executor of pixels — they are an editor of possibilities.
Anchor on Humans (The Defence Against Model Collapse)
The studios that are navigating this shift successfully aren't removing humans — they are repositioning them. Authentic human data has become a premium commodity because of a looming, systemic threat: Model Collapse. If synthetic outputs are allowed to train the next generation of models, the structural quality degrades progressively. As the open internet fills with AI-generated content, the algorithms inevitably begin training on their own output, rendering future generations a caricature of a caricature.
To stave off this mathematical decay, the case for protecting organic human data gets stronger, not weaker. High-end VFX houses are already deploying a Two-Path Strategy: a public layer for rapid ideation, and a secure, private directed layer for final production. This private layer is anchored exclusively by human judgement and trained on closed, proprietary style libraries — a director's lighting choices, a composer's harmonic fingerprint. Synthetic data that never leaves the studio.
The Actor as Licensing Entity
Scan for two days; license your synthetic twin for entire productions. The virtual influencer proof of concept was a decade ago. Actors choosing this framework for themselves is the exact chapter we are in now.
The Provenance Response
The industry is building synthetic media and simultaneously building the tools to identify it — C2PA cryptographic watermarking and Google's SynthID operating side-by-side. The existential question is whether this provenance infrastructure can scale as fast as the generation infrastructure itself. So far, it hasn't.
Beyond the Threshold
The photon is physical reality. Light bouncing off a human face, travelling through a glass lens, hitting a camera sensor. The pixel is synthetic reality. A coordinate of colour, generated on a screen by an algorithm. In 2026, audiences can no longer tell which is which.
What remains is a more interesting question. The studios surviving this shift aren't the ones trying to out-calculate the probability engines. They're the ones anchoring their pipelines in authentic human data, strict consent frameworks, and unmistakable artistic identity. The machine scales the output. The human decides what it's for.
For decades, the gatekeeper to making high-end art was the sheer, gruelling difficulty of execution — mastering the software, rendering for days, learning the complex mechanics of lighting and anatomy. You needed immense technical skill just to get your idea out of your head and onto a screen.
Now, the machine has democratised execution. Anyone can generate a flawless, cinematic image in seconds. But because anyone can do it, the execution itself loses its premium. It becomes background noise.
What cannot be automated is the human eye and taste — the instinct to know which possibility matters, the restraint to leave things out, and the lived experience required to give an image a soul.
The Machine: Scales the output infinitely.
The Commodity: Technical execution and pixel generation.
The Premium: Curation, perspective, and intent.
Rick Rubin describes the artist not as a builder but as an antenna — a receiver of taste and instinct that no generative model can replicate. The point of view matters because no model has lived a life.
Execution is now just the baseline. Curation is the craft.
The tools have become intelligent. One question remains — what do we have left to say?