Netflix used GenAI in roughly 300 titles this year

Netflix's Q2 2026 shareholder letter states that generative AI workflows "have been used in roughly 300 of our titles" in 2026. It is the hardest number any company has published about AI in production, it came from a document with legal consequences attached, and the interesting part is not the 300. It is where the letter says the work happened: post-production.

The letter names three titles and what the technology did in each. Glory, from India. Brasil 70: A Saga do Tri, from Brazil. The American Experiment, from the US. The applications listed are enhanced crowds, historical battle sequences and worldbuilding establishing shots. Then a sentence that is more candid than most studio communications manage: "In some cases, productions would have had to leave out key shots and sequences in the absence of GenAI technology." On the earnings call, Ted Sarandos described Netflix as "primarily builders, not buyers." The reporting on the letter is here.

“In some cases, productions would have had to leave out key shots and sequences in the absence of GenAI technology.”

Read the three use cases again

Crowds. Battles. Establishing shots. Those three have one thing in common and it has nothing to do with AI. They are the places where a production needs scale it cannot afford, in frames the audience looks at for two seconds.

Nobody watches an establishing shot. They register it, place themselves, and wait for the scene. A crowd is read as a quantity, not as faces. A historical battle is read as chaos and mass. These are the shots where a budget meeting ends with someone saying the sequence will have to go, and in Netflix's own words that is exactly what used to happen.

Now notice what the letter does not claim. It does not say generative AI performed a scene, carried a story beat, held a close-up, or delivered a line. It does not say the work replaced principal photography. "Concentrated in post-production" is a precise phrase from a company with lawyers. Post is where you extend, populate, clean, replace and composite around something that was already shot.

The number is a floor, not a ceiling

Three hundred titles across a slate of that size sounds like transformation and reads more like adoption of a normal post-production tool. One generated crowd extension in one sequence qualifies a title for that count. So does worldbuilding across an entire series. The letter does not break it down, and we have not seen a breakdown published anywhere, so the figure tells you that the tooling is in general use and tells you nothing about intensity.

Which is still the most valuable fact available, because it settles an argument. The question of whether generative AI is used in premium, globally distributed, union-staffed, legally scrutinised television is closed. It is. In post. For scale.

What happens when it carries the story instead

There is a useful counterweight from the advertising side. On 18 May 2026 Ipsos and Syracuse published a study comparing human-made ads against AI-generated versions of the same brief. The human versions scored 14% higher on short-term sales impact and 17% higher on brand equity. Only 25% of viewers were even somewhat confident they were watching AI, which means audiences were largely not detecting it. They were responding worse to it anyway.

The breakdown is the part to keep. AI performed best on product-driven, direct briefs. It performed worst on storytelling. Ipsos titled the piece around the finding that AI ads are good enough, and that this is the problem.

Put the two sources side by side and they say one thing twice, from entertainment and from advertising. Generative AI is strongest where the work is mechanical and weakest where the work is meaning. Netflix found that boundary by shipping 300 titles and concentrating the work in post. Ipsos found it by measuring what audiences did afterwards.

Builders, not buyers

Sarandos's line is worth taking literally, because it describes a cost structure. Netflix is building its own workflows rather than licensing someone else's product. That is a choice available to a company with that engineering capacity and it is not available to most brands or most agencies. What is available is the discipline underneath it: knowing precisely which shot in which sequence the technology is being used for, and being able to say so in a shareholder letter.

Most AI briefs we see are the opposite. They name a tool rather than a shot. The useful version of the question is not "can we use AI on this film". It is "which frames in this film are crowd, scale or environment, and which frames are a person the audience has to believe".

How this maps onto a brand film

Run the Netflix split across a typical commercial and the division is usually obvious once someone draws it.

• Mechanical. Crowd and set extension. Sky replacement. Environments that exist for one second of screen time. Versioning and resizing. Product-driven direct-response cutdowns, where the Ipsos data says AI holds up best.
• Meaning. The hero performance. The joke. The turn in the story. Anything where a viewer's judgment of a brand is formed by a human face doing something specific.

The honest commercial reading of the Ipsos number is that a fully generated brand film can be cheaper and still cost you 17% of brand equity against the same idea made by people. For a direct product film that trade may be worth making. For a brand film it rarely is, and a studio that does not say so is not being straight with you.

Netflix published a number, named three titles, said which part of the process it applied to, and admitted some shots would not have existed otherwise. That is the standard of disclosure to hold a production partner to. Ask them which shots.