Two Kinds of Creative AI
There are roughly two ways to build creative tools with AI.
The first is the vending machine. Type a prompt, get a finished image or video, done. If you don’t like it, pull the lever again. It’s impressive, sometimes astonishing, and for some uses it’s exactly right. If you need a quick illustration for a slide or a placeholder for a mockup, a vending machine is perfect.
But it leaves the person outside the work. You can request, you can reroll, you can rephrase your request, but you can’t really shape. The relationship is closer to commissioning than to making. And most people who love creative work don’t want to commission. They want to make.
The second is the collaborator. The AI works inside your process, on your material, at the level of detail you choose. It can suggest, draft, fill in, clean up, and propose alternatives. But you can take over at any moment, at any level, and the work stays yours.
That’s the kind of tool we decided to build at Charsook, and the decision shaped everything after it: the interface, the architecture, the models, and even how we measure success.
Why We Chose Collaboration
The reason is personal before it’s strategic.
I’ve been making things for a long time: graphics, games, music, animation, interactive simulations. Over those years I’ve learned that the joy of making something lives in the decisions. Which frame to cut on. Where the motion should slow down. Which color makes the scene feel late in the evening instead of early. What to leave out. Especially what to leave out.
A tool that skips the decisions also skips the joy. It hands you a result, but it takes away the experience of arriving at it. For a professional on a deadline, that might be an acceptable trade sometimes. For anyone who makes things because making things is meaningful, it’s a loss.
There’s also a practical reason. Creative work is full of intentions that are hard to put into words. A director knows the cut feels wrong before they can explain why. A designer adjusts a curve by a few pixels until it feels right. These intentions are expressed through manipulation, not description. A tool that only accepts descriptions can never capture them.
So Charsook is described as a creative intelligence platform where humans and AI make creative work together. The word “together” carries a lot of weight in that sentence. It’s not “humans ask, AI makes.” It’s not “AI makes, humans approve.” It’s a shared space where both contribute and either can lead.
What Collaboration Looks Like in Practice
In Charsook Editor, the AI can propose a cut, a transition, a rhythm, a color adjustment, or a motion. But proposals are editable objects, not finished outputs.
If the AI suggests a sequence of cuts for a music video, each cut is a real cut on the timeline. You can grab one and move it two frames later. You can delete three of them and keep the rest. You can ask for an alternative for just one section. You can undo the whole thing and start again with a different direction.
The AI also works at whatever scale you need. Sometimes you want it to take on a big chunk: “make a rough assembly of these clips matching the beat.” Sometimes you want something tiny: “smooth out this one transition.” Sometimes you don’t want it at all, and it stays quiet.
This sounds simple when written down. It was not simple to build. Generative models naturally produce finished artifacts. Making them produce structured, editable, partial contributions that respect what’s already there required rethinking a lot of the pipeline.
Authorship Lives in the Edits
One thing I noticed watching people use early versions of the editor surprised me, and it changed how I think about creative AI.
The moment someone feels the work is theirs is usually not the moment they generate something. It’s the moment they change something the AI did. When they disagree.
We have run many of these experiments, and the same scene repeats again and again. Someone accepts a suggested sequence, plays it back, and frowns. They drag one clip earlier, shorten another, swap two shots. They play it again and smile. Afterward they describe the piece as theirs without hesitation, even though most of the cuts were originally suggested by the AI. What makes it theirs isn’t the quantity of their contribution. It’s that they made the decisive choices.
Across all these experiments, the pattern holds. People don’t measure authorship by how much they produced. They measure it by whether their judgment shaped the result. Disagreement is where judgment becomes visible.
That changed how we design. We now treat every AI contribution as a draft with handles. Every suggestion should be easy to inspect, easy to partially accept, easy to reverse. If accepting a suggestion is all-or-nothing, people either accept bad work or throw away good work, and in both cases they lose the opportunity to exercise judgment.
The Hard Design Problems
Collaboration is harder to build than generation. A few problems we keep running into:
Granularity. Suggest too little and the AI is useless; the person does all the work and wonders why the AI is there. Suggest too much and it takes over; the person becomes a reviewer of someone else’s work. The right amount depends on the person, the project, and the moment. A beginner might welcome a complete draft. An experienced editor might want only a gentle hint. The same person might want different levels on different days.
We’ve experimented with explicit controls, like a slider for how much the AI should do, and with implicit signals, like how often someone modifies suggestions. Neither is perfect. I suspect the best answer will be something like a good human assistant: someone who learns your preferences over time and adjusts.
Explanation. “Why did it do that?” deserves an answer. But it needs to be a short answer. Not a paragraph; a hint. “Cut on the downbeat.” “Matched the color of the previous shot.” “Shortened to keep the pace consistent.” If the explanation takes longer to read than the decision takes to evaluate, nobody will read it.
Memory of intent. If someone rejects a style three times, the AI should notice. Respecting a “no” is part of being a good collaborator, human or not. Early versions would cheerfully suggest the same rejected idea again and again, which made people feel unheard. Teaching the system to remember and respect preferences made a bigger difference to satisfaction than any improvement in suggestion quality.
Undo as a promise. In a collaborative tool, undo is not just a convenience. It’s a promise that nothing the AI does is permanent unless you want it to be. That promise is what makes people willing to let the AI try things. Without reliable undo, every suggestion feels risky. With it, suggestions feel like play.
Not interrupting. Creative work requires flow. An AI that constantly offers ideas, however good, breaks concentration. We’ve learned to make the AI quieter by default, available when called, and careful about when it volunteers anything.
The Ethics of Creative AI
Building creative tools with AI also means thinking about questions that go beyond product design.
Whose style is the AI drawing on? Generative models learn from enormous amounts of human work. When a model suggests a cut or a color grade, it’s echoing decisions made by many editors and colorists who never agreed to teach it. I don’t think there are easy answers here, but I do think creative tools should be honest about what they are and should help people develop their own voice rather than flattening everyone into the same average style.
What happens to skill? If the AI can do the basic work, will beginners ever learn it? I worry about this, especially as someone who has spent years coaching students and engineers. My hope is that a well-designed collaborator can be a teacher too: explaining its suggestions, showing alternatives, and helping people understand why something works, not just doing it for them.
Who gets credit? In professional contexts, this matters for careers and livelihoods. I believe the answer should follow the pattern I observed: credit belongs to whoever made the decisive judgments. But the norms are still forming, and they’ll be shaped by tools like ours.
What I Believe Now
I don’t think AI will replace human creativity. I think it will replace a lot of tedious work around creativity, and that’s a gift. Syncing clips, cleaning up audio, finding the right frame among thousands, matching colors across shots, rendering variations. These tasks consume enormous amounts of creative people’s time and give back very little joy.
The risk is building tools so eager to help that they take the creative act away along with the tedium. That’s the line we try to walk every day: take the drudgery, leave the decisions.
I also believe creativity is not a fixed resource that AI either replaces or doesn’t. It’s a capacity that grows with use. If tools make people more creative by giving them more room to experiment, more people will develop creative confidence. If tools make people passive consumers of generated content, that capacity will weaken. The design of the tool determines which direction it pushes.
The best compliment we’ve heard about Charsook so far is simple: “it feels like my work, just faster.” That’s the target. Not “the AI made something amazing.” Not “I barely had to do anything.” Just: it’s mine, and I got to spend my time on the parts that matter.
If we can keep earning that sentence as the AI gets more capable, I’ll feel we’ve built the right thing.