CineSight Briefs · Issue 002 · August 12, 2026
Research cutoff: August 13, 2026
For the last few years, most of the attention around AI filmmaking has gone to the same place.
Look what it generated.
A spaceship. A monster. A beautiful woman walking through neon rain. A tracking shot that would have cost an arm and a leg.
Some of it is mind-blowing.
Some of it looks like expensive screensaver material.
But we’re watching the more important change happen somewhere else.
AI is moving inside the filmmaking process.
It is organizing footage. Building rough assemblies. Searching media. Manipulating existing shots. Turning live-action footage into editable 3D scenes. Learning relationships between cuts. Tracking subjects inside cameras. Researchers are even teaching robotic camera systems to learn moves from human demonstrations.
At almost exactly the same time, audiences are beginning to push back against the flood of AI-generated crap appearing everywhere they look.
Those two things may sound contradictory.
I don’t think they are.
One is AI becoming a tool.
The other is AI becoming pollution.
And filmmakers are going to have to know the difference.
◆ ◆ ◆
What’s Actually Changing1. AI is moving into the edit
This may be the most consequential filmmaking technology story right now because it isn’t really about generating video at all.
Adobe’s Premiere AI Assistant, currently in beta, lets an editor use ordinary language to perform tasks inside a Premiere project. Adobe says it can organize media, prepare footage, add markers, work with transcripts, and assemble an initial edit while leaving the editor in control.
Adobe: Premiere AI Assistant overview
Adobe also calls this an early public beta and does not recommend using it for client work yet. Treat it as something to test, not something to trust with the only copy of a real project.
Think about what that means in practical terms.
An editor may increasingly be able to say to the machine:
Organize these interviews by subject.
Find every answer where she discusses her childhood.
Mark the slates.
Pull the useful interview clips into one sequence.
Give me a rough assembly.
That isn’t AI making a movie.
That’s AI doing some of the tedious work that happens before the editor can really start editing.
And researchers are going further.
A new experimental system called FilmGPT, published July 16, approaches editing almost like a language problem.
Large language models learn statistical relationships between words and predict what is likely to come next. FilmGPT was trained on movies to learn statistical relationships between shots.
Give it raw footage and it attempts to determine what shot should come next.
Importantly, it does not generate new footage. It selects from the footage the filmmaker actually supplied.
The researchers report that FilmGPT outperformed previous approaches in shot-sequence ordering and montage evaluations, and they demonstrate applications including automatic trimming and human-in-the-loop editing.
Read the FilmGPT research paper
Before anybody starts writing the obituary for editors, let’s get something straight.
Recognizing statistical patterns in thousands of edited movies is not the same as understanding why an editor stays on someone’s face for two uncomfortable seconds longer than expected.
FilmGPT can learn what filmmakers often did.
That does not mean it knows what you should do.
And that may become one of the most important distinctions in AI-assisted filmmaking.
Who this could affect: Editors, assistant editors, documentary filmmakers, students, postproduction teams, and eventually anyone dealing with large amounts of footage.
What filmmakers should do: Start paying attention to AI-assisted editing, not merely AI-generated video.
The interesting question is no longer whether AI can make another shot.
It is whether AI can help you understand and work with the shots you already have.
◆ ◆ ◆
2. AI is moving behind the camera, too
The same shift is happening elsewhere in production.
Autodesk Flow Studio can take live-action footage and turn it into editable CG scenes and production elements, including motion capture, camera tracking, clean plates, character passes, and 3D scene information that can continue into Maya, Blender, Unreal Engine, and other conventional workflows.
Even more interesting for education, Autodesk currently makes Flow Studio available free to eligible post-secondary students and educators. The educational version includes its Lite features plus USD export.
Autodesk Flow Studio for students and educators
That matters because it moves AI away from the idea of:
Prompt → Finished Thing
and toward:
Footage → AI processing → Editable assets → Human revision
That looks a lot more like filmmaking.
Now move one step farther toward the physical camera.
Researchers this year introduced an experimental system called IRIS, a six-axis robotic cinema arm designed to learn camera movement from human demonstrations.
Instead of programming every coordinate of a move, the system learns object-aware camera trajectories from demonstrated moves. The researchers report that their prototype costs under $1,000, supports a 1.5-kilogram payload, and achieves approximately one-millimeter repeatability.
Read the IRIS robotic-camera research
IRIS is research, not something I’m recommending anyone buy for their next short.
But the direction is worth watching.
We have gone from:
manual focus
to autofocus
to face detection
to eye detection
to subject recognition
to pose estimation
to predictive tracking.
Now researchers are asking whether the camera movement itself can be learned.
Eventually the cinematographer may not simply tell a machine where to focus.
They may tell it what kind of move they want.
And then the important question becomes the same one we keep running into everywhere else:
Who decided that was the right move?
Who this could affect: Cinematographers, camera operators, VFX artists, virtual-production teams, educators, and microbudget filmmakers who may eventually gain access to camera-control technology previously limited to expensive motion-control systems.
What filmmakers should do: Watch AI that creates control, not merely AI that creates images.
A production tool earns its place when you can give it a precise note, change your mind, and revise one thing without destroying five others.
◆ ◆ ◆
3. Meanwhile, people are getting sick of AI slop
WIRED published a piece this week documenting something we should not dismiss as another round of anti-AI complaining.
The backlash is beginning to affect company behavior.
WIRED reports that LinkedIn added a reporting option for content that “seems like AI slop,” Snapchat stopped allowing fully AI-generated videos into its discovery feed, and companies including Meta and Google have withdrawn or changed generative features following public criticism.
WIRED: The AI Slop Backlash Is Actually Having an Impact
The interesting part is why.
A lot of people are not simply saying:
“I hate artificial intelligence.”
They’re saying:
“I never asked for this shit to be stuffed into everything.”
They didn’t ask for their work to become training material.
They didn’t ask for synthetic content to overwhelm their feeds.
They didn’t ask for deepfake tools.
They didn’t ask for every search, photograph, post, advertisement, and software interface to suddenly become an AI experiment.
That’s partly a quality problem.
It’s partly a trust problem.
It’s partly a volume problem.
And underneath all three is something even more important:
consent.
For filmmakers, that creates a distinction we’re going to have to get much better at explaining.
There is a difference between:
AI helped repair dialogue in my film
and:
AI generated the film.
There is a difference between:
AI helped organize 40 hours of documentary footage
and:
AI created a fake human performance.
There is a difference between:
AI helped me communicate a production-design idea
and:
I filled the internet with 400 synthetic images because generating them cost almost nothing.
Calling all of those things “AI” tells us almost nothing.
That’s going to become a problem for filmmakers because audiences may increasingly want to know what actually happened.
And “AI was used” is a pretty shitty explanation.
A filmmaker should eventually be able to say:
We photographed the actors traditionally. AI was used for three background extensions and dialogue cleanup. No performances were synthetically generated.
That’s not an apology.
That’s production information.
Who this could affect: Anyone releasing AI-assisted work to an audience.
What filmmakers should do: Keep enough information about your process that you can explain what the technology actually did.
Not because everyone needs to love your choices.
Because you should be able to stand behind them.
◆ ◆ ◆
The Classroom Is Facing the Same ProblemFilm education has spent a lot of energy asking how to determine whether students used AI.
I’m increasingly convinced that this is the wrong question.
A student can now use AI to produce a plausible script analysis, storyboard, shot list, image, production plan, edit suggestion, or piece of footage.
That does not prove the student understands any of it.
It also does not prove the student learned nothing.
So instead of asking only:
Did AI touch this?
we should be asking:
What decision did you make?
Why?
What did you reject?
What changed after you tested it?
What did the tool get wrong?
What did you change?
Can you defend the result?
MSU Denver’s current guidance on AI and assessment makes essentially the same shift. It argues that the goal should not be creating “AI-proof” assessments, but designing assessments that still measure what students actually know and can do.
Harvard recommends approaches including reflective AI-use journals, scaffolded assignments, and oral defenses where students explain their process and use of AI.
Harvard: Academic Integrity and Teaching With(out) AI
The University of Florida’s updated 2026 guidance similarly emphasizes disclosure when required, critical evaluation of AI output, privacy, security, intellectual property, transparency, and accountability.
University of Florida AI Guidelines
That makes much more sense to me than pretending we’re going to recreate a classroom from 2019.
But there is another side to this.
Process documentation should not turn into surveillance.
Students should not have to surrender every private note, prompt, draft, creative thought, or developing idea simply to prove they did the work.
Ask for what you actually need.
A short reflection.
A selected screenshot.
A before-and-after example.
A production summary.
A brief oral defense.
Enough evidence to evaluate the learning.
Not enough to confiscate someone’s creative life.
Four questions I’d start asking students
1. Show me one important decision you made.
2. Show me one option you rejected and tell me why.
3. Show me what changed after feedback, rehearsal, production, or editing.
4. Tell me where the tool helped and where you refused to let it decide.
If the student can answer those questions intelligently, now we’re talking about filmmaking.
A detector score can’t do that.
◆ ◆ ◆
Premiere is becoming conversational
The Premiere AI Assistant beta represents something bigger than another Adobe feature.
The interface itself is changing.
Instead of hunting through menus for every operation, editors may increasingly describe what they want accomplished and let the software execute the repetitive steps.
CineSight take: Pay attention. Natural language may become another control surface inside professional software. But faster operation does not eliminate the need to understand what the software just did.
FilmGPT is learning statistical film grammar
FilmGPT’s researchers describe their goal as learning the “grammar” of film from a large collection of movies rather than relying on hand-coded editing rules.
The model predicts shot relationships and uses footage-constrained decoding to construct montages from existing material.
CineSight take: Watch closely. This is research, not a production recommendation. But AI that understands relationships between shots may ultimately matter more to editors than AI that simply generates more shots.
Robotic cinematography is getting cheaper
The experimental IRIS cinema robot demonstrates that machine learning, inexpensive fabrication, and camera robotics are beginning to collide.
The prototype is nowhere near replacing a professional motion-control system, but a sub-$1,000 research platform that learns camera trajectories from demonstration is a hell of a signal.
CineSight take: Watch, don’t buy. If intelligent motion control eventually becomes accessible to small productions, that changes what microbudget cinematographers can physically execute.
AI filmmaking education is becoming easier to put into a real classroom
Autodesk’s decision to provide Flow Studio free to eligible university students and educators is more consequential educationally than another paid AI course.
Students can work with motion capture, camera tracking, clean plates, editable 3D scenes, and USD exports that continue into established production software.
Autodesk Flow Studio Education
CineSight take: This is worth testing in filmmaking and VFX programs because students are learning a workflow, not merely generating an image.
AI video is getting a production layer
A funding announcement this week put Preview on the wire. The company says it has raised $12 million to build what it calls a production platform for AI video.
The interesting part isn’t the money.
Preview combines script breakdowns, shot lists, an infinite visual canvas, multiple image and video models, version tracking, collaboration, approvals, and a record of how each frame was made. It is trying to organize the production around the generators.
Preview: AI video production platform
CineSight take: If AI-generated production becomes real production, somebody still has to keep track of the script, shots, versions, notes, approvals, and where everything came from. Otherwise, it is just a more expensive pile of disconnected shit.
◆ ◆ ◆
Reality CheckThe phrase AI filmmaking is becoming almost useless.
It can mean:
a fully generated video,
an editor using AI to search interviews,
a VFX artist extracting a clean plate,
a cinematographer using AI subject tracking,
a filmmaker using an image model for previs,
a student asking an LLM to critique a script,
or a synthetic performer who never existed.
Those are not remotely the same thing.
If we lump all of them together, we cannot have an intelligent conversation about the benefits, risks, ethics, or creative consequences.
The better question is:
What job is the AI doing?
Then ask:
Why is it doing that job?
What material did it receive?
What happens to that material?
Can the filmmaker revise the result?
Who made the final decision?
Who owns the work?
Who consented?
Those questions survive whatever model gets released next Thursday.
◆ ◆ ◆
Worth Your TimeFor filmmakers
Before adding an AI tool to a production, ask:
5. What precise problem am I solving?
6. Is the tool actually better than the existing method?
7. Can I revise one element without rebuilding everything?
8. What happens to the material I upload?
9. Can I export everything I need if the service disappears?
10. Who owns the final judgment?
For educators
Stop designing assignments around catching AI.
Design them around exposing thinking.
If a student cannot explain why they chose the lens, changed the blocking, rejected the generated storyboard, cut away from the performance, or abandoned the AI suggestion, that’s useful evidence.
If they can explain it clearly, that’s useful evidence too.
◆ ◆ ◆
Ignore This Week“AI can edit movies now.”
No.
FilmGPT predicting shot relationships is fascinating research.
It is not Walter Murch.*
“AI is inside the camera, so cinematographers are next.”
Also no.
Autofocus didn’t eliminate cinematography.
Neither will subject recognition or intelligent motion control.
New capability changes what the cinematographer can control. It doesn’t decide what deserves to be photographed.
“The AI backlash means AI is dying.”
Not even close.
The more interesting possibility is that the market is beginning to distinguish between useful AI and unwanted AI.
That would be healthy.
“More generated content means more creative opportunity.”
Sometimes.
It also means more shit competing for the same finite amount of human attention.
When production becomes nearly unlimited, selection becomes more valuable.
Taste becomes more valuable.
Trust becomes more valuable.
And somebody still has to decide what is worth showing another human being.
◆ ◆ ◆
The CineSight TakeAI filmmaking is beginning to grow up.
The important developments are becoming less spectacular and more useful.
Media organization.
First assemblies.
Editable VFX.
Production assets.
Camera tracking.
Motion control.
Documentation.
Assessment.
Those things don’t make great demo reels.
They make workflows.
At the same time, the backlash against AI slop is telling us something from the other direction.
People do not necessarily want more content simply because machines can make more of it.
They want something worth their attention.
That puts filmmakers in an interesting position.
The machine is getting better at participating in the process.
It may eventually recognize the shot, suggest the cut, track the actor, move the camera, rebuild the background, repair the dialogue, and organize the footage.
Fine.
But somebody still has to know why the shot exists.
Somebody has to know why the cut works.
Somebody has to know when the performance is truthful.
Somebody has to decide when the technology is helping and when it needs to get the hell out of the way.
That isn’t nostalgia.
That’s the job.
◆ ◆ ◆
I set the editorial direction, selected what belonged, and made the final calls. AI tools assisted with research organization, source discovery, drafting, comparison, and fact-checking. Significant factual claims were checked against the sources linked throughout the issue.
The research process also changed this week. Future CineSight scans will specifically look beyond generative-video models and track AI entering filmmaking hardware and software, including cameras, editing systems, VFX, sound, virtual production, production software, firmware, motion control, and robotics.
That is the process I am comfortable standing behind.
◆ ◆ ◆
About CineSight BriefsCineSight looks at how AI and emerging technology are affecting films, filmmaking, filmmakers, and film education.
The point is not to report every announcement.
It is to figure out what changed, whether it matters, and what a filmmaker might actually want to do about it.
No daily tool dump.
No manufactured urgency.
No bullshit wearing an expensive suit.
*Walter Murch is an Oscar-winning film editor and sound designer whose work includes The Godfather, Apocalypse Now, and The English Patient. He is widely regarded as one of the most influential thinkers in film editing and sound.