How Matt Lenski used AI to champion humans in The Last Commercial

The era of ‘AI filmmaking’ is well and truly upon us and, depending on your viewpoint, is either the beginning of a brave new world or the end of conventional cinema as we know it.

Imagining the more creatively apocalyptic version, but viewing it very much through a comedic lens, Arts & Sciences director Matt Lenski’s new short The Last Commercial is a deliciously wry look at the human quirks generative digital creation struggles to mimic.

The twist: the film itself combines live-action and gen AI processes, with real actors’ performances driving AI-generated characters in a world built from photoreal storyboards. 

shots sat down with Lenski to find out how he kept performance at the heart of the process, where the technology needed a firm directorial hand and why ‘Beep Boop Cinema’ is no laughing matter.

What sparked the idea for The Last Commercial? Was there something in your own experience that made you want to explore the subject?

I made a film about AI replacing commercial production, using AI, then spent months directing actors – humans – and deciding whether a make-believe algae cube was reading properly on camera.

It bubbled up from a feeling I think a lot of people are having: holy shit, look what this can do, followed almost immediately by holy shit, if it can do all this faster and cheaper, what happens to us? And just to make it stranger, a lot of AI still kind of sucks… but for how long?

So I started thinking about the last human commercial. Not because I think there will literally be one (spoiler: capitalism survives), but if you want to satirise existential dread, a commercial set is basically already dressed for it.

I’ve been working with AI on films and commercials, so this wasn’t about discovering the tools. It was about building an AI process that keeps people at the centre, around what I actually care about: performance, humour, dialogue, collaboration, timing.

There’s an obvious irony in using AI to tell this story. Was that part of the idea from the start?

That contradiction was basically the whole reason to make it.

I wanted to use the thing everyone is anxious about to make a film about that anxiety. In the original brief, I asked if we were carving a headstone with the Grim Reaper’s scythe.

A lot of the language came from choices I’d make on any shoot: performance first, handheld camera, anamorphic lenses, little reactions, awkward pauses.

Credit to Mal and Marc at Arts & Intelligence for getting behind it; there’s something great about an AI-forward production company funding a filmmaker to ask what happens if this goes too far.

The film isn’t a prediction. It’s closer to: maybe we don’t?

The production process is part of the argument. Could we use AI while still building the film around actors, performance, direction, editing, and human choice? That was there from day one.

What were your starting points for the film’s look and tone? Which references helped you bring its world into focus?

I thought of it as a really mundane Children of Men. Or The Office if robots had somehow ended up in management.

For the story to connect, it needed to feel familiar, even in a scene with a robotic brand director giving notes. So a lot of the language came from choices I’d make on any shoot: performance first, handheld camera, anamorphic lenses, little reactions, awkward pauses. Black and white, because silliness gets funnier when you photograph it like history.

No Beep Boop Cinema. That’s my name for the hyper-polished AI look: sweeping camera moves, perfect compositions, dialogue that sounds like an audiobook narrator. AI desperately wanted Beep Boop Cinema. A lot of the work was me and Jose from Banquet steering it back towards something human.

Above: The AI augmented imagery [left] and its performance-led source [right].

Can you tell us a bit about the ‘transformation’ of the footage? What was shot, what was generated, and how were they brought together?

Simply: the performances are human; the imagery is AI-generated. Think Andy Serkis in a mocap suit pretending to be an ape, except the suit is running on a laptop. (RIP to the one that died mid-production. All gave some, some gave all.)

The Banquet team and I built the world in photoreal storyboards first, then I shot real actors performing the scenes. In post, those performances drove the generated characters, sometimes transferring directly, sometimes informing a new shot, like a great cover song, with a lot of compositing and human labour throughout.

I could explain every trick, but half will change in two weeks. Part of my job is staying close to the tools and knowing how to use them and when not to use them at all. The goal was always the same: build the world, then have humans act it out, whatever digital costume they’re wearing.

One shot in the film is completely live action. Did you catch it? It’s the moment the actor finally nails the take. The moment the human performance breaks through, we leave it alone.

What were you looking for in the casting? How much could the actors understand or see of the world they would eventually inhabit? How did you approach working with SAG-AFTRA?

Casting was everything (as always). This is a comedy. You need actors who can mess with the rhythm, surprise each other, make a pause weird, make a line better by stepping on it.

Part of my job is staying close to the tools and knowing how to use them and when not to use them at all.

That calibre of actor tends to be union, so we got the film SAG-AFTRA certified. Not quick. At times it felt like we were writing the blueprint for every SAG AI contract to come. Quite possible. But if you’re making a film about keeping humans in the process, you should probably start by hiring them properly.

Ed Herbstman, who co-wrote the film and plays the actor, is an improv icon. God, he’s good. The actors had our storyboards and animatic, so they could see the world; then we put them on a green-screen stage and asked them to behave like it was all there. Mostly they just did what actors do: breathe life into it.

The film uses humour to approach a subject that carries real anxiety for people in this industry. How did you find the tone, and who – or what – did you want the audience to be laughing at?

Satire is useful because you can smuggle an uncomfortable conversation in by wearing a stupid hat.

The fear around AI is real, and I feel it too; a surprising amount of this film was written during sleepless nights at 2 a.m., apparently when my existential career panic gets productive.

Advertising is absurd, and it’s healthy to laugh at yourself. The funniest stuff is rooted in something painfully true: we wrestle with huge questions about creativity, jobs, authorship, and the future, then spend 30 minutes deciding which way the product should face. Very big anxiety, very small problems.

And underneath it, the film is about choice. The technology is not making these decisions by itself. We are. Sometimes thoughtfully, sometimes nobly, sometimes like complete dumbasses.

Comedy depends on precise timing. How much control did this process give you over that? What role did the edit and sound play in making the jokes land?

Comedy is a dictatorship of frames. I don’t care how miraculous the image is; if the pause is two frames too long, the joke is dead. So we built the performance edit first, from actors. You don’t get that timing or interplay from strictly AI characters.

From there, the edit becomes its own creative session. The take doesn’t have to be the take, but the changes still come from the performance. In the final days of the edit at Mackcut, we could hold a reaction, extend a beat, add a robot to the background, or maybe the CD should sniff a piece of cheese. Yes! Generating towards the cut, whatever it needed. I love this.

Comedy is a dictatorship of frames.

Sound mattered just as much: Eric at Heard City rebuilt the world around the images; Margy at Found Objects gave it a score that made it feel like a film.

It took a lot of smart humans to make the thing work.

Every director has wished for one more take. Now sometimes you can go get it, using the actor’s performance as the thing you build from. That is where the line starts to blur. What counts as the performance? As creative? Let’s talk about it.

What was the biggest gap between what you thought the technology could do and what it actually delivered? How did you solve the problem that caused you the most trouble?

The AI demos make it look like you push a button and go make lunch. In reality, AI is dumb as shit with excellent presentation skills. It can hand you something beautiful and have no idea why it’s very wrong.

I’m proud of the human-driven performances in the film, but the tools still have a way to go. A direct line to Luma Labs helped. Anne Marie and the team made themselves available, but there’s no single right way to do any of this. You learn by doing, and by talking about it.

The best work came from people around a table, trying things, arguing, laughing, discussing lunch, solving the problem. Very old-fashioned, really.