The Setup
Priya Venkataraman makes documentary shorts. Not the kind with Netflix budgets and field producers. The kind where she is the director, the sound person, the editor, and the person who stays up until 2am arguing with herself about whether a cut is emotionally honest. She has been doing this since 2019 out of a converted spare bedroom in Portland, releasing work through Vimeo, a modest Substack, and a growing Patreon that as of early 2025 was pulling in just under $3,400 a month from 280 subscribers.
Her documentaries run between 18 and 35 minutes. She covers labor stories mostly... garment workers in Los Angeles, warehouse contractors outside Memphis, a three-part series on what happened to a small-town newspaper when its parent company sold it for parts. The work is good. It wins things occasionally. More importantly, people finish watching it, which is its own metric in an era where the average YouTube retention rate on videos over 10 minutes sits around 40 percent.
In January 2025, Priya started integrating AI into her workflow. Not tentatively. Aggressively. She had read the think pieces, watched the tutorials, and made a decision: she was going to let the tools do what the tools could do so she could do what the tools could not.
What she did not anticipate was how quickly she would need to figure out exactly where that line was.
The Problem
The first six weeks were genuinely good. She used AI transcription to process interview footage that used to take her three hours per hour of tape. She used automated clip-finding to surface moments she had tagged but buried. She used a scene-detection tool to rough-cut B-roll sequences before she ever touched the timeline. These were real time savings. The math worked like this: a 25-minute documentary that used to take her roughly 140 hours from shoot to export was now taking closer to 95. That is a 32 percent reduction in production time, which for a solo creator is not a minor efficiency... it is the difference between four releases a year and six.
Then she hit the problem.
She was finishing a piece about a union organizing drive at a fulfillment center in Sparks, Nevada. The central subject, a woman named Cora, had given her 11 hours of interview footage across three sessions. Priya had been using an AI tool to surface the strongest moments based on engagement signals and transcript coherence. The tool kept surfacing the same four clips. They were technically excellent: clear, articulate, emotionally legible. Cora was persuasive in them. She made her points well.
But Priya knew something the tool did not. The best moment in the whole 11 hours was a 90-second sequence in session two where Cora said almost nothing. She started to answer a question, stopped, looked out the window, and then said quietly, "I just want to be able to tell my daughter I tried." The tool did not flag it. By transcript coherence and engagement signal metrics, it was low-value. It had a long pause. Cora did not complete her initial sentence. From a data standpoint, it looked like a stumble.
Priya had almost missed it herself, buried in 11 hours of tape. But she had found it. She knew it was the film.
That is when she realized what had almost happened. She had been training herself to trust the tool's selections. Slowly, across six weeks, she had started treating AI recommendations as a first pass that was probably right rather than a first pass that was probably useful. The distinction is not semantic. It is the whole game.
What They Did
Priya did something that felt counterintuitive at first. She drew up a written taxonomy of her workflow. Every task she did in a production cycle, she categorized into one of two columns. The first column she labeled "execution tasks" ... things where the correct answer exists and the variable is how long it takes to find it. Transcription. File organization. Color reference matching. Rough assembly from pre-selected clips. Subtitle generation. The second column she labeled "judgment tasks" ... things where the correct answer does not exist until she decides what the film is trying to say. Clip selection with emotional weight. Structural sequencing. The specific moment a scene ends. Whether to use silence or music in a transition. Interview order. What to cut entirely.
The rule she made for herself was simple: AI touches column one. She touches column two. Not AI-assisted. Not AI-suggested-and-she-approves. She touches it, full stop, with no algorithmic input shaping her attention before she gets there.
She also changed how she used the AI clip-finding tools. Instead of asking them to surface the best moments, she started using them to surface the moments she had not yet watched carefully. She used the tool's low-engagement flags as a map to buried material, not as a recommendation engine. She was, in effect, using the algorithm's blind spots as a research prompt.
This took discipline. The tools are designed to surface recommendations. Using them as negative-space detectors required actively resisting the framing the interface wanted her to adopt.
What Happened
The Sparks documentary came out in March 2025. It is the best thing she has made. The 90-second Cora window sequence became the emotional center of the piece, the moment that got screenshotted and passed around in labor organizing circles on Signal. A journalist at The Intercept linked to it. Patreon went from 280 to 410 subscribers in six weeks. Revenue went from $3,400 to just under $5,000 a month.
But the more significant thing was what happened to her output rhythm. With the taxonomy in place, she released three pieces in the four months following. Before, her average was about one every eight weeks. The execution tasks were no longer bleeding into her judgment time. She was not tired in the same way at the end of a production cycle because she was not spending cognitive energy on tasks that did not require it.
What most people miss about creative burnout is that it is not always about making too much. Often it is about spending finite creative attention on tasks that don't deserve it, so when you get to the task that does, you are already depleted. Priya had been using her best hours to do work a machine could do, and then making her hardest creative calls on the remainder.
Flipping that order changed everything.
What I Learned
There are four things worth extracting from what Priya figured out, because I think they apply to almost every solo creator wrestling with these tools right now.
First: the tool's confidence is a design choice, not an accuracy signal. AI recommendation interfaces present selections with a kind of quiet authority. There is no uncertainty language. There is no "this is a guess based on incomplete data." The framing is always curatorial, not probabilistic. That is a UX decision made by an engineering team, and it shapes how you relate to the output. If you do not consciously push back on it, you will start treating recommendations as judgments.
Second: what an algorithm flags as low-value is often where the actual art lives. The math works like this... engagement signal models are trained on what held attention across millions of data points. They find the mean. The mean is not where distinctive work lives. The silence before Cora said what she said is exactly the kind of material that will never train well. It is slow, it is incomplete, it is one specific woman in one specific room. It is also exactly why Priya's film is better than the one the algorithm would have assembled.
Third: the taxonomy approach only works if you maintain it with bureaucratic stubbornness. The temptation to let the tool "just suggest" before you look is constant. The interface encourages it. Resisting it is not a technical problem... it is a habit problem, and habits erode under time pressure. Priya told me she still catches herself defaulting to the recommendation view when she is behind on a deadline. The solution she found was logging when she broke the rule and why. That log became a forcing function. She has broken it four times in six months. She considers that a success.
Fourth: Apple's Creator Studio Pro framing matters more than the product itself. When a company with Apple's cultural weight releases a tool and explicitly positions AI as a tedium-remover rather than a decision-maker... that is not just a product description. It is a permission structure for how creators are supposed to relate to AI. It says: the machine does the grunt work, you do the thinking. That framing is worth defending. Not because Apple is trustworthy, but because the frame is correct. Stewart Brand spent decades arguing that access to tools does not change who does the thinking. It just changes what the thinking can afford to spend its time on. That is still the right way to hold this.
Priya figured out the distinction under pressure, with real work on the line. Most people figure it out after they have already made something that felt hollow and could not explain why. If you can draw the column before you hit that moment, you are ahead.