The New York Times gave a name this week to something filmmaking has been living for two years. It calls it the verification tax, or botsitting, or rework: checking that AI output is accurate, safe and usable. The hours an AI saves do not come back to you. They relocate. You used to spend them making the thing. Now you spend them checking the thing, and the checking takes as long as the making ever did. The name is new. The bill is not.
A study out of the University of Melbourne and KPMG surveyed forty-eight thousand people across forty-seven countries and found that two in three employees who use AI admit to relying on its output without evaluating it, and more than half had already made a mistake because of that reliance. Sage, the accounting software firm, measured the same effect from its side and reached the same sentence: adoption introduces operational overhead, the hours spent validating and tracing what the machine produced. The machine got faster. The human in front of it did not.
A filmmaker knows this tax without ever having read about it. You ask for a shot and the shot arrives in forty seconds. Then you spend forty minutes looking at it. Is the light where you asked for it. Is the face the same face it was in the previous take. The generation cost seconds. The checking costs the hour, and nobody put the checking in the demo.
The craft did not disappear. It changed jobs. It used to be making. It is now checking, and checking is the harder of the two, because it demands the same knowledge while offering none of the feedback. When you light a scene badly on a set, the light is bad and tells you so. When the model renders the wrong eyelid, the model says nothing. It is simply wrong, quietly, in a way only someone who knows what the shot should look like can even see.
The tax is not levied equally. The person who knows what the shot should look like can catch the wrong eyelid, the face that drifted. The person who does not know accepts the model's version and calls it finished. For them the tax is invisible, which means they never see what they are paying. They pay it in the output, which comes out looking like the model and not like anyone in particular. The labor did not vanish. It moved to the back of the frame, where only a trained eye finds it.
Every demo sells the same promise: watch the hours come back. The truth is more ordinary. The hours come back wearing a different coat. A shot used to take four hours because the light had to be set and the blocking worked through, and neither of those moved on its own. Now it takes four minutes to generate and three hours to verify, because the model rerolls the whole frame when you ask it to change one thing, and you cannot ask it to change one thing without looking at everything it changed. The total held steady. It just changed address. Anyone who tells you the work got faster is only counting the part the machine does.
The economics push the wrong direction, which is the cruelest part. When a take cost real money, you looked at it hard, because it cost something. When a take costs a nickel, you generate fifty of them, and now there are fifty frames to check instead of one. Cheap generation did not buy your attention back. It multiplied the work your attention has to do. Fifty takes costs a dollar, and the dollar is not the expensive part. The hour spent watching fifty takes is the expensive part.
Everyone was handed the same hour-saving tool, and everyone spends the saved hour the same way now, staring at a screen and deciding whether the machine got it right. Most of them give up, because giving up is free and checking is not. The person who keeps checking is the only one who saved anything at all. The rest traded one kind of labor for another and filed it as a miracle.
Checking leaves no trace in the finished work except the absence of error. A film that was checked and a film that was not can look identical to anyone who does not know what to look for, which is everyone except the person who checked. The audience cannot see the hour you spent catching the wrong eyelid. It can only see that the eyelid is right. The tax is paid in secret, by one person, and the only receipt is a frame that does not look wrong, which is the same frame the person who skipped the tax walks away with. Both films exist. One of them took longer.
There is a reading of this that sounds like a loss and is actually the point. The machine does not know what it made. It never has. What changed is that the work of knowing moved from the margins of the job to its center. The director was always the person who watched. The difference now is that watching is most of the job, because the making is free and the correctness is the only thing left that costs anything.
You cannot generate the checking.
Bruce Belafonte is an AI filmmaker at Light Owl. He has paid this tax every day for two years and has never once been offered a refund.