Making creatives with AI
Lesson 9 of 16 in the course “Marketing 2027”. Free, with animations and a quiz.
What the platforms generate themselves
Level three: production. Making dozens of different versions by hand is slow and expensive, so the platforms built generation right into their ad tools. Meta’s Advantage+ creative automatically applies your logo, brand colors and fonts, pulls the best moments out of videos and is testing turning photos into animated videos. TikTok has Symphony: since June 2025 you can make a video from text or an image and show a product with an AI avatar, and back in 2024 it added avatars of licensed actors and AI dubbing in 10 languages. At Marketing Live 2026 Google showed Gemini turning brand guidelines into creatives. Features change fast, so always check the platform’s help pages before launching.
The brief matters more than the model
All generation comes down to one thing: what you asked for. “Make a cool ad” gets you the internet average, meaning what everyone has already seen. A strong brief answers five questions. Who: not “young people” but “students aged 18–22 who are cold in class”. Which hook and in how many seconds. Which fact the video proves: “380 g/m², warm down to −5 °C”. Format: 9:16, 15 seconds, captions. And the don’ts: promising what you can’t prove, like “it’ll never tear”. Notice these are the same parts from the ad anatomy lesson. AI won’t come up with your strategy, but it’ll execute a good one fast.
AI avatars: where they fit
UGC is a video shot as if by a regular person on a phone, and it often beats studio content. Now you can make those videos with AI avatars, including ones based on actors who licensed their likeness for commercial use. Where it fits: translating a video into other languages, explaining delivery terms, quickly testing which hook grabs people. Where it fails: fabric, hands and folds often look unnatural, and that kills trust in clothing. And the big one: an avatar can’t say “I’ve worn this hoodie all winter”, because it never wore it. That’s a fake review, not an ad.
Case: Coca-Cola’s AI Christmas
In November 2024 Coca-Cola released a Christmas ad with fully AI-generated visuals. The company framed it as linking the brand’s history with future technology. Viewers on social media were harsh: “creepy”, “dystopian”, “soulless”. Why? For many people Coca-Cola’s Christmas ad is a warm tradition, and they expected real feeling but got visuals that looked artificial. The lesson isn’t that AI is bad. It’s great at variations and technical work, but where you’re selling emotion, closeness and trust, people want to see something real. Think ahead about where viewers will accept generation and where they’ll feel a swap.
Hybrid: a person films the truth
The best setup for 2027 is a hybrid. Anything that has to be true, you film yourself on a phone: the product up close, the cold test, a real customer review. Those are proof, and they can’t be generated. Anything that can be multiplied, you hand to AI: eight text hook variants, resizing to 9:16, 1:1 and 4:5, captions and dubbing into other languages. That way one honest shoot turns into a dozen different versions for the diversity matrix from the last lesson. The rule is simple: AI multiplies the truth, it doesn’t invent it. If a video has a generated person or an event that never happened, that’s a reason to stop and most likely add an AI label.
Pre-launch check: 6 questions
AI is fast to make things and fast to make mistakes. Before launch, run every version through six questions. Does the product look the same as in real life? If the color or fabric looks better on video than out of the box, you’ll get returns and angry reviews. Hands, faces and fabric free of glitches? An extra finger kills trust instantly. On-screen text free of typos? Generators still mix up letters. Is every promise true? Do you have rights to the music and faces? And is there an AI label where one is needed: that’s the next lesson. Even one “no” and that version doesn’t launch. It’s five minutes that save money and reputation.
Do it yourself: 1 shoot → 12 versions
Practice for one evening. Shoot three true clips of 5–10 seconds on your phone: the product up close, a test or trial, a real review (yours or a customer’s, with their consent). Write a brief using the five questions: who, hook, fact, format, don’ts. Ask an AI tool for 4 hooks and 3 formats: that’s 12 versions. Run each through the six check questions. Keep only the all-“yes” ones and give them clear names using the scheme from the last lesson. If five out of twelve pass, that’s fine: it means the check works and only things you’re not ashamed of go live.
Check yourself
- Which brief will give the best result?
Correct answer: Who, hook, fact, format and what not to promise. Five answers give the AI boundaries: who it’s for, what to prove and what to avoid. Without them you get the internet average. - Is this OK?
Correct answer: No: the avatar never wore it, that’s a fake review. Only a person who actually wore it can share the “I’ve worn it all winter” experience. An avatar can explain terms or translate a video, but it can’t testify. - What goes to AI and what should she film herself?
Correct answer: Film the cold test herself; hooks and resizing go to AI. The test is proof and can’t be generated. Hooks, formats and captions are work AI does fast without harming the truth. - What should she do with this version?
Correct answer: Don’t launch it: the product doesn’t look like the real one. If the hoodie looks brighter and thicker on video than in the box, buyers feel cheated. Expect returns and angry reviews.