The First Fully AI-Generated Commercials: Case File

"Fully AI-generated commercial" is a phrase that almost never survives contact with the production notes. Nearly every milestone spot that carried the label used a human-composed score, human-directed editing, corrective visual effects, or a specialist studio spending weeks on character consistency. That does not make the claims fraudulent. It makes them a useful lens on what generative AI could and could not do at each moment. This is the verified timeline of the ads that moved the line, what each one actually automated, and an honest read on the craft. If you are trying to figure out where the technology really is, the gap between the "fully AI" headline and the credits is where the answer lives.

TL;DR

• "Fully AI" almost always has an asterisk. Score, edit, VFX cleanup, or character-consistency work is usually still human. Read the credits, not the press release.
• The timeline runs Adobe Firefly (Oct 2023) → Toys "R" Us / Sora (mid-2024) → Coca-Cola holiday (late 2024, again 2025) → Kalshi / Veo 3 (June 2025) → Svedka / Super Bowl (Feb 2026).
• The register matters more than the model. Chaotic or synthetic subjects (robots, absurd characters) survived AI's fingerprints. Warm, nostalgic, human-centered spots got mocked.
• The honest craft read: by 2026, AI can carry a 30-second spot end to end if the concept flatters the tooling, but "hands-off" is still a myth at broadcast quality.

What "fully AI" actually means

A commercial has more parts than the footage. Script and concept, visuals, voiceover, music, sound design, editing, color, and finishing. A genuinely "fully AI" ad would generate all of them with no human in any seat, which has essentially never shipped at broadcast quality. What ships instead is a spectrum:

• AI visuals, human everything else. The most common milestone. Footage is generated; script, edit, music, and cleanup are human.
• AI visuals plus AI voice, human music and edit. More recent, as native-audio video models arrived.
• AI concept-to-cut with human finishing. The current frontier: models generate most assets, a human still directs the sequence and finishes it.

Keeping this spectrum in mind is what stops you from either overclaiming or dismissing the technology. Nearly every "first" below sits at a specific, honest point on it.

The timeline

October 2023: Adobe Firefly, the first Firefly-powered TV ad

Adobe premiered what it billed as the first TV commercial powered by its Firefly generative AI, running during sports primetime to promote Photoshop's AI features. This was AI-assisted, not AI-generated: Firefly produced creative elements inside an otherwise conventional spot. It matters as the moment a major company put generative AI on broadcast at all.

Mid-2024: Toys "R" Us, the Sora brand film

The one most people mean by "first." Toys "R" Us, through agency Native Foreign, made a brand film largely with OpenAI's Sora, telling the origin story of founder Charles Lazarus and mascot Geoffrey the Giraffe, and screened it at Cannes. The credits are the lesson: the film included corrective visual effects and an original human-composed score by Aaron Marsh of the band Copeland. So the milestone "first fully AI commercial" was, precisely, AI visuals with human finishing and human music. It drew heavy backlash for uncanny footage, an early demonstration that the technology's weak spot was exactly the warm, human, nostalgic register the brand was reaching for.

Late 2024 and again 2025: Coca-Cola, the cautionary tale

Coca-Cola released AI-generated holiday ads two years running, produced with multiple AI studios (Secret Level, Silverside, and Wild Card among them) using several generative models. Both years drew significant backlash; reporting on the 2025 round noted only about 37% positive social sentiment. The technical achievement was real and the reception was not. Coca-Cola is the clearest evidence that "can we generate it" and "should we generate it" are different questions, and that iconic, emotionally-loaded brand moments are the riskiest place to visibly automate.

June 2025: Kalshi, the primetime cost breakthrough

The prediction-market platform Kalshi aired a 30-second spot during Game 3 of the NBA Finals, made almost entirely with Google's Veo 3. Filmmaker PJ Accetturo produced it in two days using 300 to 400 Veo generations trimmed to 15 clips, for roughly $2,000. It went viral, over three million views on X. Kalshi is the milestone that mattered most economically, because it proved a broadcast-quality-looking spot could be made for the price of a nice camera and run in primetime. It also nailed the register lesson Coca-Cola failed: chaotic, absurd characters where Veo's occasional weirdness read as intentional comedy.

February 2026: Svedka, the Super Bowl claim

Svedka aired what it called the first "primarily" AI-generated national Super Bowl spot, a robot-party ad built over about four months with AI studio Silverside. The four-month timeline is the honest detail: even in 2026, a broadcast-grade AI spot with consistent characters was a multi-month specialist effort, not a prompt. See the full Super Bowl LX breakdown for the wider context of that night.

The honest craft read

Three conclusions after watching the line move for two and a half years.

The tooling caught up to the concept faster than to the credits. By 2026 a model can plausibly carry a full 30-second spot's visuals. What it still does not do is remove the human from music, edit, and finishing at a quality bar audiences accept. The Toys "R" Us score and the Svedka four-month build are not footnotes; they are where the remaining craft lives.

Register beats resolution. The single strongest predictor of whether an AI ad landed was not the model or the resolution. It was whether the subject flattered the technology. Robots, absurdist characters, and stylized chaos survived. Warm nostalgia and human faces at rest did not. If you are choosing what to generate, choose a register where synthetic texture is a feature. Knowing that boundary saves you from shipping the uncanny version and getting mocked for it.

Character consistency is the real frontier. Both Svedka's four months and the general difficulty of keeping a face or figure identical across shots point at the same unsolved-until-recently problem. This is why character consistency and identity locking are the features that actually gate professional work, far more than raw resolution.

What this means for brand teams

Do not chase the "fully AI" headline. It is a press label, not a production strategy, and it invites exactly the scrutiny that got Coca-Cola boycotted. Chase the parts of the spectrum that pay off: generate what flatters the tooling, keep humans on the score, the edit, and the finish, and pick a register that makes the synthetic texture intentional. The rights and disclosure picture matters here too, because "fully AI" claims draw attention to exactly the questions you would rather answer proactively.

The practical move is to build the spot on a canvas where you can generate across models, lock your character early, and hand a finished sequence to a human editor. That is the workflow behind every milestone that actually worked. Start on 8frame to run it, and see the best AI-made ads of 2026 for the campaigns that got the balance right.