Netflix told investors that generative AI was used in roughly 300 titles in 2026, mostly in post-production. Not experiments. Titles that shipped, including Glory, Brasil 70: A Saga do Tri and The American Experiment, where it was used to build crowds and historical battle sequences.
In March it bought InterPositive, an AI startup founded by Ben Affleck, for up to 600 million dollars. In June, Google DeepMind put 75 million into A24 to build filmmaking tools.
The industry that spent 2023 striking over this has finished arguing about whether. It has not finished arguing about disclosure, and that gap is now the commercial risk, which is a different problem from the one everybody prepared for.
Where it is actually being used
The detail that matters most to anyone commissioning creative is the phrase "in post-production." The reported savings, cost and time roughly halved on specific sequences, attach to particular shots: crowd extension, historical recreation, cleanup, format conversion. Not to the production.
That distinction is going to be blurred in every pitch you receive over the next year. "AI cut our costs in half" is a true sentence about a battle sequence and a false one about a shoot. If a vendor quotes you a halved budget, the question is which shots, and what happens to the schedule on everything else.
The honest version of the efficiency claim is narrower and still substantial: the expensive, repetitive, technically fiddly middle of post is being compressed. Principal photography is not. Neither is the part where somebody decides what the thing should be.
The A24 deal is the one to read closely
Seventy-five million dollars, and the reported terms do not give Google access to A24's content library or its data. That is unusual enough to be the story.
Most creative-side AI arrangements have been quietly structured the other way: preferential tooling in exchange for material to train on. A deal that explicitly separates the two establishes a reference point, and reference points are what you negotiate against.
So when a production partner, an agency or a platform offers you AI tooling, the question to put in writing is whether your footage, your stills, your voice recordings and your customer-facing creative are being used to improve a model, and whether that improvement is exclusive to you. There is now a public example of a studio declining to trade the library. You are permitted to want the same terms, and you will not get them by not asking.
Consent, controls, compensation
The talent side has converged on a three-part test, usually rendered as the three Cs: consent, controls, compensation. It is a reasonable framework and it is worth adopting on the buying side too, because the exposure runs in both directions.
If you have ever filmed an employee, a customer or a spokesperson, you hold a likeness. If your release form predates generative video, it almost certainly permits use of that footage and says nothing about training on it, generating new performances from it, or synthesising a voice. That silence is not consent, and the reputational version of this problem arrives long before the legal one.
The practical step costs an afternoon. Find your standard release. Read what it says about derivative and synthetic use. If the answer is nothing, update it before your next shoot rather than after somebody notices a generated version of a face they recognise.
The reason the industry moved anyway
Behind all of this is a distribution fact rather than a technology one. Short-form platforms now take a larger share of viewing time among the 13 to 54 audience than the major studios' output combined. Studios are cutting existing content into vertical clips to compete for that attention, and doing that at volume by hand is not economic.
That is the real driver, and it applies to everyone reading this with a marketing budget. The demand is no longer for one polished asset. It is for that asset in nine aspect ratios, twelve lengths and four hooks, refreshed before fatigue sets in. Nobody is going to fund that with a traditional edit bay, which is why the tooling arrived and why the objections lost.
The competence that becomes valuable is not generating. It is deciding what should exist, and editing what comes back, which is precisely where these tools remain weak.
What to do on Monday
Ask your creative suppliers where AI is in their pipeline, and get the answer in writing. Not to police it. Because you need to know what you are buying, and because you cannot make a disclosure decision about your own brand without it.
Then decide your own disclosure posture deliberately rather than by default. The industry's current position is closer to don't ask, don't tell, and that is a position with a shelf life. Deciding late, under pressure, after a customer asks, is the expensive way to arrive at a policy that costs nothing to write today.
Three hundred titles. The argument is over. What is left is whether you can say, accurately and immediately, which parts of your own work a machine touched.
Sources
- Variety and Engadget on Netflix's second-quarter earnings disclosure. Source for generative AI used in roughly 300 titles in 2026, primarily in post-production, and for the named titles Glory, Brasil 70: A Saga do Tri and The American Experiment, including crowd and historical battle work and the reported halving of cost and time on specific sequences.
- Reporting on Netflix's acquisition of InterPositive, March 2026: a company founded by Ben Affleck, acquired for up to 600 million dollars, bringing a 16-person team across, with Affleck in a senior advisory role on filmmaker-facing generative tools.
- TechCrunch, Variety and Google on the DeepMind and A24 partnership, June 2026. Source for the 75 million dollar investment and for the reported term that the deal does not give Google access to A24's content library or data.
- The Economist, issue of 8 to 14 August 2026, for the talent-side "three Cs" framing of consent, controls and compensation, for the short-form viewing-share comparison, and for the characterisation of the industry's approach as don't ask, don't tell. Its account of a ByteDance video model triggering a takedown dispute does not match the public record on the products named and has been left out.