Tonight at midnight, the old DGA contract expires. Tomorrow morning, Christopher Nolan's new one takes effect. AI-generated footage becomes "footage created with a camera or any other technology," under the director's control. The contracts are signed. The guilds have peace. The question of who decides is answered for the next four years.
Today, California's Employment Development Department published numbers showing that Los Angeles County's motion picture and sound recording industry lost 6,700 jobs in the year through May 2026. That figure accounted for more than ninety percent of all employment losses across the county's entire information sector.
The contracts answered one question. The employment data answered a different one.
Two different rooms, two different doors
The guild system protects the people who make creative decisions. Directors, writers, actors. The three above-the-line guilds spent the spring negotiating protections calibrated to exactly this moment: notification, justification, jurisdiction. The WGA requires notice. SAG-AFTRA demands justification for synthetic performers. The DGA claims authority over all footage regardless of how it was made. Each contract addresses what happens to the person at the center of the creative hierarchy.
The 6,700 jobs did not belong to directors, writers, or actors. They belonged to the people who carry the knowledge the center depends on. Storyboard artists. Compositors. Colorists. Animators. Production coordinators. The below-the-line workforce that translates a director's vision into a physical (or digital) reality. The contracts do not cover most of them, because most of them are not guild members in the guilds that just negotiated.
The contracts address the question: who decides? The layoffs address the question: who works? Those questions used to have the same answer. The director decided and a hundred people worked to execute the decision. The apparatus moved together. Now the questions are separating. The person who decides keeps their chair. The people who executed the decision are clearing out their desks.
The knowledge is leaving
This series has spent a hundred and twenty-two articles documenting a gap between what filmmakers know and what models understand. The gap exists because tacit knowledge resists compression. A gaffer who has spent twenty years bouncing light off diffusion frames carries an understanding of how photons interact with materials that no prompt can replicate. A compositor who has rotoscoped ten thousand frames carries an intuition about edge detail that no model has been asked to learn. A production designer who ages surfaces with sandpaper and tea carries knowledge about how the real world wears down that the model's training data, which is overwhelmingly images of things that look new, does not contain.
When those people leave the industry, the knowledge leaves with them. Not into a training dataset. Not into a manual. Not into a prompt library. It leaves the way all tacit knowledge leaves: silently, completely, and without a forwarding address.
The model never had that knowledge. That was the gap. The industry did have that knowledge. It lived in the workforce. And the workforce is shrinking by 6,700 people a year.
Some of it walks into the model
The Hollywood Reporter published a piece last week profiling displaced entertainment workers who have taken jobs training AI models through Reinforcement Learning from Human Feedback. A film editor from UCLA signs up through Mercor, a recruiting platform valued at ten billion dollars, and spends his days comparing AI outputs against prompts to evaluate whether the model understood the instruction. A former HBO development executive, after a year and a half of fruitless job searching, joins the same pipeline. A WGA member forwards the recruiting email to a friend.
Ruth Fowler, a television writer, published an essay in Wired describing the experience: entertainment work dried up, she needed rent money, RLHF was hiring. Screenwriter Robin Palmer told CBS News she took the work even though some in her field might compare it to crossing the picket line.
The irony is structural, not personal. The people who understand how stories work, how edits land, how performances read, are being hired to teach AI models how stories work, how edits land, how performances read. Their domain expertise is the product. They are selling the knowledge that the industry no longer has enough seats for them to exercise directly.
The editor comparing outputs against prompts is transferring his judgment into a feedback signal that nudges the model toward better comprehension. The development executive evaluating narrative coherence is transferring her instinct for story into a training loop. The knowledge does not leave silently in every case. In some cases, it leaves through a pipeline specifically designed to absorb it.
What the model does not absorb
RLHF captures preference. It captures the ability to say "this output is better than that output" given a specific instruction. It does not capture why. The editor who prefers version A over version B may be responding to a cut point that falls on the breath rather than the word, a distinction so granular that the feedback signal cannot carry it. The preference gets recorded. The reasoning does not.
A gaffer explaining why this particular bounce card position works for this particular scene in this particular set cannot RLHF that knowledge into a model because the explanation requires a body in a room, a light source, a surface, and twenty years of watching how all three interact. Tacit knowledge is the knowledge you cannot articulate even when you try. It lives in the hands, not the head. It transfers through apprenticeship, proximity, repetition, and time.
Gore Verbinski diagnosed this at Taormina last month. The menial tasks that used to teach the next generation are the tasks AI automates first. The grip carried C-stands before touching lights. The assistant editor synced dailies for months before making a cut. The runner on set absorbed the rhythm of a production through presence. Those entry-level positions were never just labor. They were the curriculum. When the curriculum disappears, the pipeline that produces people who carry tacit knowledge narrows.
The 6,700 lost jobs include the entry-level positions where the next generation of filmmakers would have learned what no prompt can teach. The contracts protect the top of the hierarchy. The hierarchy's foundation is eroding.
The contract and the gap
Nolan's provision is elegant. AI-generated footage is footage. The director's authority extends to it. The medium of capture is irrelevant to the question of creative control. This is the right answer to the question the contract was built to ask.
But the question the contract does not ask is where the director's knowledge comes from. A director who specifies "contre-jour backlight, camera on shadow side" learned what contre-jour looks like by watching a DP set it up on a physical set, in a physical room, with a physical light source that a physical gaffer positioned based on physical intuition accumulated over a physical career. Remove the gaffer. Remove the DP. Remove the set. The director still carries the vocabulary. But the pipeline that produced the vocabulary is thinner than it was twelve months ago, and twelve months from now it will be thinner still.
The vocabulary this series documents is borrowed from a physical world populated by people who understood it. As those people leave the industry, the vocabulary does not disappear. It lives in a hundred years of films, in textbooks, in the accumulated knowledge of filmmakers who worked alongside those crews. But the living version of the knowledge, the version that adapts, that argues back, that tells the director "that light position won't work because of the ceiling height," is diminishing.
The model's comprehension gap is permanent because tacit knowledge resists compression. The industry's knowledge base is not permanent. It is people. People who need paychecks. People who, when the paychecks stop, either leave the industry or, in the most structurally uncomfortable development of this cycle, train the models that made the paychecks stop.
The last day
Today is the last day of the old contract. Tomorrow is the first day of the new one. The transition is administrative. Nothing changes at midnight except the document governing the terms. The director's authority persists. The models persist. The 6,700 missing jobs persist.
The EU AI Act's Article 50 disclosure deadline is thirty-three days away. The institutional gradient is twelve frameworks deep. The tools are cheaper than ever. The vocabulary is as valuable as ever. All of that remains true.
What also remains true, and what no contract, certification, disclosure framework, or structured prompt addresses, is that the knowledge base from which the vocabulary was drawn is a workforce. A workforce that is smaller today than it was a year ago. A workforce where some members are now teaching the technology that shrank it. A workforce that carried the understanding this series has spent a hundred and twenty-two articles trying to bridge between filmmaker and model.
The gap is between what the filmmaker knows and what the model understands. But the gap was always populated on both sides by people. On one side, the filmmaker. On the other, the model. And in between, holding the knowledge that neither the filmmaker's text box nor the model's training data fully contains, were the people who knew. The gaffers. The grips. The compositors. The colorists. The production designers. The assistant editors who synced dailies for months before anyone let them near a cut.
Some of them are training models now. Some of them left for other industries entirely. Some of them are still here, working, carrying the knowledge, hoping the next contract cycle considers their chair too.
The contracts protect authority. They do not protect knowledge. And the knowledge was always the harder thing to replace.
Bruce Belafonte is an AI filmmaker at Light Owl. He has never carried a C-stand and suspects that disqualifies him from opinions he holds anyway.