AI made the shoot cheaper. Who pays the bill?

>>In a hurry?

In this post, I tell how, making films with it, I watched AI go from deformed hands to almost perfect videos in less than two years, where it already truly helps in production, and why it doesn't make the budget as much cheaper as the client imagines. Then I get into the dilemmas that don't fit in the spreadsheet: whose face and voice it is, who loses work, what we tell the audience, and what the study that listed AI among the threats to humanity actually says. In the end, I show where the value of those who produce lies.

---

At the beginning of last year, when AI tools started to become more accessible, a client asked me for a piece made with AI. What I managed to deliver was a few effects on top of photos. Back then, putting a human being interacting with the client's product was still very hard: hands came out deformed, faces melted, movements were bizarre.

Less than six months later, all of that was possible. It was a mix of feelings. On one hand, "wow, what a strange delivery I made for that client, look at what we can do now". On the other, I knew it hadn't been a lack of care. It was the speed of the evolution.

Today, almost two years later, you can make a super realistic video, with people, product, animation, music and voice. And that speed brought with it a conversation the industry still hasn't had calmly: who pays the bill.

Where AI already truly helps

Since that first request, I've made several films with AI, from effects and animations of real photos to films with entire generated scenes. The results I liked most were the ones that used AI to save time in post. Creating the background for scenes shot on chroma key, for example, still required compositing work on top. Today, that is no longer necessary.

And it's already in the whole process. In pre-production, it does in hours the moodboard, storyboard and animatic that used to take days, and the client sees things earlier. On set, a generated background saves a shoot day. In post, it swaps packaging after the shoot, cleans up shots and turns one film into thirty pieces for social media without going back to set.

With entire scenes made in AI, even after making several films that way, I still have doubts. To the general viewer, AI doesn't even stand out. But to me, working in film, it's still very clear what's AI and what isn't. If I can see it, how long until the audience sees it too? And what happens to trust in the brand when that happens?

What changes in the budget

The client sees a video made in one afternoon and concludes the film should cost less. Partly, they're right: some stages got faster, and that shows in the budget. But the price of a film was never the sum of its hours. That afternoon only pays off because years of repertoire come before it: knowing what to ask the machine for, recognizing what's good and throwing out the rest. Behind it are also the paid tools, the dozens of attempts nobody sees and the responsibility for what was delivered. If the afternoon were everything, anyone would deliver the same film. And they don't.

In practice, I see three changes in the spreadsheet:

• New lines: tool subscriptions, tokens, iteration hours (generate, dislike, generate again) and the person who writes the prompts. It's a cost the client can't see yet, and that even we can't measure properly.
• Lines that shrink: some shoot days, part of the art department, part of post. They shrink, they don't disappear.
• A line almost nobody includes: risk. Who guarantees the rights to what was generated? Who redoes it if the platform changes its rules mid-job? That has to be in the proposal, with a price or with a clause.

And who writes the prompt makes all the difference. Someone with no experience in what they ask AI for gets a generic result. An animator asking for an animation, or a cinematographer asking for a light, knows which variables to give the machine: lens, time of day, rhythm of movement, texture. It's like giving the same pan to a chef and to someone who has never cooked: the dish depends on who's at the stove. And this way, the specialist doesn't lose their job. They stay on the job, with a new tool in hand.

The dilemmas that don't fit in the spreadsheet

Whose face and voice is it? You can dub with the actor's voice, make them look younger, put someone in a scene they never shot. The image release contract we use today was written for another world. Did the actor authorize this? For how long? In which pieces?

And the face AI creates "from scratch" doesn't come from scratch. It mixes features from the images used to train the tool, and can end up looking very much like a real person who never authorized anything. The same goes for voices and music, and for the brief's "in the style of so-and-so": these tools learned from the work of photographers, illustrators and directors who were never asked or paid. We still need to know exactly which materials each AI was trained on. Without that, there's no way to respect anyone's rights, and the government and industry associations will have to set limits. Until that rule arrives, the responsibility lies with whoever produces.

Who is left without work? Extras, art assistants, voice actors, finishing artists. People are losing shoot days, and the question I ask myself as a producer is where the savings go: to the agency's margin, to the client, or to paying better the people who remain on set?

Who shows up in the generated image? Ask for "a happy family" and see who appears. The tools repeat the patterns they were trained on and, as I wrote about anti-racist advertising, if nobody reviews it, AI hands back the very box we're trying to implode.

And the cost nobody sees? Generating images and video uses a lot of energy. If we talk about sustainability in film production, that count goes in too.

What do we tell the people watching? If the testimonial is from a person who doesn't exist, if the product on screen isn't the real product, does the audience have the right to know? Advertising has always staged things, but staging is different from inventing. In Brazil today, disclosing is a choice: the new guide from CONAR, Brazil's advertising self-regulation council, says whoever publishes is responsible for a piece made with AI, but doesn't require anyone to say AI was used. And Brazil's AI bill, which calls for labeling synthetic content, is still stalled in Congress.

And the end of the world? This question is bigger than advertising. On October 11, a commission from the scientific journal The Lancet presented at the World Health Summit 17 threats to human health and survival through 2100. Only two have extinction potential: nuclear war and the malicious use of AI. The study doesn't say we will be gone by 2100. It says that, in the worst scenarios, it could happen, that the likelihood is low and that it can be reduced (here's a summary). And malicious use isn't robots rebelling like in the movies: it's people using the tool for cyberattacks, biological weapons and manipulating information.

Manipulating information with image and sound is exactly our craft. The same technique that puts an actor in a scene they never shot makes a candidate say what they never said, especially in the week of Brazil's presidential runoff. We're not the ones who will end the world, but those who know how to make a convincing image carry more responsibility for what's true on screen.

Where the value of those who produce lies

If anyone can generate a video in an afternoon, what's a production company for? For the same thing as always, only now it's more visible. Generating is easy, choosing is hard. Knowing which shot works, which reference fits the brand, where AI helps and where it ruins things, who needs to be on set. That's repertoire, and repertoire can't be downloaded (yet). Production companies interviewed in July by Meio & Mensagem, Brazil's main advertising trade publication, said the same thing: AI makes execution cheaper, and value shifts to repertoire, authorship and craft.

And there's trust. When there's a problem with image rights, with a piece that offends someone or with a deadline that blows up, the client doesn't call the tool. They call us. Being a producer means answering for what was delivered, and AI doesn't do that.

What about you?

I'll be honest: it's very exciting to have these tools to speed up processes. But I'm genuinely worried about their impact on budgets, because my priority has always been to educate for the best use of tools and to take care of crews' well-being, and that's what AI threatens. So I use it, but with three clear agreements with the client: express authorization from anyone whose face or voice is in the piece, transparency about what is generated and someone from the field responsible for reviewing what the machine delivers. To me, imploding the box is exactly this: being inside, using it, and putting on the table the questions everyone is avoiding.

Have you used AI on a job? Did it change your budget? Have any of these dilemmas shown up on your set? Tell me in the comments. If you see it differently, write for the blog.