AI has become such a broad term that it’s almost useless, especially in relation to filmmaking. When someone says “I don’t want AI used on my movie,” or “I think they used AI on that movie,” it could mean almost anything.

It might mean something we previously used other computer algorithms for—removing noise or grain from a picture, or cleaning up a background hiss on an audio track. Or it might mean something more involved, like generating images or shots from text, helping write the script, or producing an entire movie with very little human input. Those are wildly different things, and yet we reach for the same two letters to describe all of them.

Explore the interactive AI Filmmaking Spectrum →

Where the idea came from

I’ve been talking with a lot of people lately to understand how they think about these tools—what worries them, and what they’re genuinely excited about. Across all of those conversations, the same problem kept surfacing: we don’t have good language for the breadth of what “AI” actually covers in our field.

So I had an idea to build a graph that lays out that breadth—all the different kinds of tools that, under the hood, rely on the same family of machine learning algorithms. They use neural networks to accomplish things we used to do with more traditional, hand-written algorithms. And they share a common feature: they’re all trained on data.

For something simple like removing noise from an image, you feed the model noisy images paired with clean ones, and you train it to turn one into the other. Text-to-video is far more complicated, but it works out to a surprisingly similar idea: you start with what is essentially an image of pure noise and refine it, step by step, until it looks like the words in the prompt. Same underlying machinery, very different workflows for filmmakers.

And somewhere along that range, people tend to draw a line about what tools are suitable for their filmmaking.

Voting and Interactivity

What do people think about these tools?

I built a little tool that lets you rate each AI tool used in filmmaking on two axes:

  • How production-ready is it? How far along is it in its evolution—is it something you’d actually trust on a real show today, or is it still finding its footing?
  • How algorithmic vs. generative is it? Is it essentially a better (or just different) version of an algorithm you already understand, with something close to a right answer? Or is the tool making its own creative interpretations of the work?

That’s the whole chart: utility on one side, generative creativity on the other, and readiness running bottom to top.

The AI Filmmaking Spectrum: tools plotted by how algorithmic vs. generative they are, and how production-ready.

What I’ve learned sharing it

I’ve now had the chance to share the AI Filmmaking Spectrum with hundreds of people—filmmakers, technicians, engineers, artists, people working independently in film, and colleagues at The Academy. People have appreciated seeing it laid out this way.

Many have voted on the tools and added new ones, so the data collected here has become a genuinely interesting snapshot in time.

And because you can vote on a tool’s position and watch the dot animate to the new collective average, it really drives home how subjective this is. How generative or creative a tool feels, or how algorithmic it feels, depends entirely on your own experience and how you’re planning to use it. There’s no easy objective answer. This is a complicated topic with a lot of variables and a lot of squishiness.

A few notes on how it works

When I use the spectrum in a live session, I open it up for voting and adding new tools—attendees can drag the tools around and add new ones, and the chart updates live during the meeting so everyone can see the input come in. Outside of those sessions I keep voting turned off, mostly to prevent spam and abuse so I know I’m gathering this data from those who are passionate about the craft of filmmaking.

If you have ideas, contact me and let me know which tools you’d like to see added or moved on the chart.

A small meta-experiment

One last thing that felt fitting for the subject. I built this tool using LLMs to help write the code. The first working version—including the live, dynamic voting—took only an hour or two to put together. By hand, that’s the kind of thing that would have taken me a few weeks.

The tool is open source under the GPL-3.0 license. You can explore the live version here and find the source on GitHub.