Coca-Cola Ran It Again. The second time was the decision.
The first AI Christmas ad was called an eyesore and AI slop. A year later the company made another one, bigger, and said so on purpose. What it weighed is the same thing sitting in your meeting right now — and the evidence everyone reaches for says something neither side expects.
You know the meeting. Someone senior has seen a demo, and it was genuinely impressive, and the question arrives already answered: why are we not doing this. You have an objection and it is a real one. The humanity is the thing people are paying for. That is what separates us from the cheap version of us on the same shelf.
Then it is your turn to speak, and what you have is a conviction. What the other side has is a cost line. The meeting is going to be decided by whoever brings a number, and neither of you has one.
So the argument becomes an argument about taste, which is an argument you lose slowly.
Video clips produced and refined by a team of five people in 30 days, for Coca-Cola's 2025 AI-generated Christmas advertising, according to the company's own behind-the-scenes video.
Hold that against what a thirty-second broadcast spot normally costs in people, days and insurance, and the pull becomes obvious. This is not a story about executives who do not understand craft. It is a story about a cost structure that changed by an order of magnitude while the quality argument was still being had.
The first one ran in 2024. Coca-Cola described it in its own release as a global milestone, the first entirely generative-AI-created film on broadcast media, set to the music of the company's 1955 ad. The reception was not warm. The Verge called it an eyesore. Social media settled on AI slop.
A year later the company made another one.
A first attempt is a bet. A second attempt, after the first was publicly dismantled, is a position — and the people who ran it said so plainly rather than hiding behind the launch.
We need to keep moving forward and pushing the envelope. The genie is out of the bottle.
Pratik Thakar, Coca-Cola's head of generative AI, on the 2025 advertising, as reported November 2025
In the same reporting he addressed the previous year directly: people criticized the craftsmanship, and this version was ten times better. And then the sentence that actually contains the strategy: the company cannot keep everyone one hundred percent happy, and proceeds where most consumers respond well.
Read that as a trade rather than an opinion, because that is what it is. Coca-Cola weighed a production economics gain of roughly an order of magnitude against a loud, visible and repeated authenticity complaint, and took the gain. Twice.
Here is the part worth sitting with: on the evidence available, that may well be the right call for Coca-Cola. Nobody pays a premium for a Coke because a person drew the truck. The brand's price is held up by availability, memory and a hundred years of distribution. The authenticity complaint is real, it is loud, and it is landing on an asset the company was not charging for.
Which is exactly why the case does not transfer to the brand in your meeting, and exactly why quoting it as precedent is the mistake both sides are about to make.
When this argument reaches for evidence, it reaches for the trust literature, and the trust literature does not say what either side assumes.
The most directly relevant recent experiment — 617 participants, a two-by-two design, published in Behavioral Sciences in 2026 — tested what happens when content carries an AI disclosure label. The label measurably reduced perceived authenticity. It had no direct effect on trust or on intention to act at all.
That finding is bad news for both people in the room.
It is bad news for the objection, because the straightforward version of it — people will not trust us — is not what was measured. If you walk in saying AI will cost you trust, the first person to read the paper wins the meeting back.
It is bad news for the enthusiasm, because the damage did not disappear; it moved. In the same model, lower perceived authenticity predicted higher perceived risk, and higher perceived risk predicted lower trust and lower adoption. The effect is real and it is indirect. It runs through authenticity, which means it shows up in the places authenticity is doing commercial work — and for a premium brand, the place authenticity is doing commercial work is the price.
The wider literature points the same way without settling it. A 2025 systematic review covering 97 peer-reviewed articles on virtual influencers finds trust is built by anthropomorphic traits, content quality and identity cohesion, and undermined by the uncanny valley, perceived inauthenticity, and this: virtual influencers are generally perceived as carrying greater social-psychological distance than human ones. That distance costs a mass brand very little. It is close to the entire proposition of a premium one.
Calvin Klein and Lil Miquela
May 2019Calvin Klein's #MYCALVINS campaign paired Bella Hadid with Lil Miquela, the virtual influencer, in an advertisement that included a kiss. Criticism followed quickly and the company apologized.
It is cited constantly as the cautionary tale about virtual influencers and brand trust. It is not one. The criticism, and the apology, concerned accusations of queerbaiting — the use of a same-sex kiss between two figures presented as women to sell product. The virtual influencer was in the frame; it was not the complaint.
The most-quoted evidence in this argument is evidence of something else. Anyone bringing it to your meeting has not read past the headline, and that includes anyone bringing it on your side.
Registered before writing, because the thin parts matter more than usual here.
Labels are not influencers. The 617-person experiment tested disclosure labels on content, not a brand replacing a human spokesperson with a synthetic one. The mechanism is likely to carry. The size of the effect is not established for the harder case.
The review reports no effect sizes. The 97-article synthesis is directional. It tells you which way the forces point and not how hard they push, which is precisely the number this decision wants.
Backlash is not sales. Coca-Cola's two rounds of criticism are visible, measurable and almost entirely on social media. What happened to volume is not public. A company that runs the same play twice has seen something the audience has not, and reading loud complaints as commercial damage is the error that makes marketers sound unserious in finance meetings.
The two most on-point papers are behind paywalls. A 2024 study in the Journal of Consumer Behaviour on virtual influencer disclosure and brand trust, and a 2025 study in Psychology & Marketing on virtual influencers and luxury perception specifically, both sit directly on this question and neither is quoted here, because I could not read them. If this decision is worth real money to you, those two are worth the access fee.
- Your price sits meaningfully above a functional equivalent — and you cannot say in one sentence what the gap buys. The gap is the thing being spent here, and nobody in the meeting is pricing it.
- Someone has proposed AI creative as a cost saving — which is the correct argument for it. Treat it as a cost decision with a brand cost attached, rather than a brand decision with a cost attached.
- Your category competes on availability and memory rather than meaning — in which case Coca-Cola's trade is probably yours too, and the objection in the room is taste rather than economics.
- You are being asked to prove a negative — that something will hurt. You will not win that. The answerable question is what the human part is currently worth, and that one has a number.
Run one disclosed test in a channel you own, and measure willingness to pay rather than engagement.
Take one offer at full price. Build two creative sets for it — one made with AI and labeled as such, one made the way you make things now. Run them against matched audiences in a channel you control, for four weeks, and judge them on conversion at full price. Not likes, not comments, not sentiment; the literature already tells you sentiment will be worse, and sentiment is not the asset. Set the threshold before you start: decide now what gap in full-price conversion would make you say the human part is paying for itself, and what gap would make you admit it is not.
- What it costs
- Four weeks of one channel's budget, two creative builds instead of one, and the genuine possibility that the AI set performs well enough to settle the argument against you.
- How you'll know
- You can state, in one sentence and with a number behind it, what your audience pays for the human part. Whichever way it lands, the next version of this meeting takes ten minutes instead of a quarter.
Coca-Cola did not win an argument about authenticity. It declined to have one, because it had already worked out that authenticity was not the line item under threat.
The brand in your meeting may not have that luxury, and the difference between the two situations is a number nobody in the room is holding. Until somebody measures it, the loudest person wins — and the loudest person is usually the one with the cost line.
Related from Sound Decisions: A Third of the Shelf, Two Percent of the Growth · Anyone Can Look Like You Now · The Memory Isn't Yours
One conversation. What your audience is paying for the human part, and whether you can still charge it.
A read on where your price premium actually comes from, so the AI question becomes a cost decision with a known number attached instead of an argument about taste.
Get a free positioning review Or read the frameworks →