Reflections on Coca-Cola’s AI-assisted Christmas ad reboot
A lot has been written about Coca-Cola’s decision to make extensive use of generative AI in the creative for the 2024 remakes of its iconic 1995 ‘Holidays Are Coming’ Christmas ad.
Common themes have been:
Quality of the visuals, with a particular focus on AI anomalies (implausible hands, wheels not rotating, objects out of proportion).
Impact on jobs, with an assumption that an AI-free production would have resulted in more work for more people.
Use of AI models which have been trained on unlicensed material.
Impact on the environment, with an assumption about the relative impact of AI-generated video vs live action.
Speed and cost of the production process, with an assumption that cost-saving was the primary driver for making such extensive use of AI.
Negative viewer reaction (“ads that sparked mockery and disgust from social media users”, “people are appalled”).
A few reflections:
1.) Industry professionals are not well placed to gauge the reaction of regular viewers to AI-assisted creative
It’s not hard for those with a critical eye to find shortcomings in the visuals for this ad. AI-generated imagery and video have improved dramatically in the last couple of years but the models’ lack of understanding of anatomy and real world physics continues to result in anomalies. These anomalies are particularly jarring when it comes to rendering people, whose features and gestures we are hard-wired to scrutinise. We instantly know if something is off.
Whilst casual viewers may still notice - and be bothered by - anatomical howlers, they’re generally not scrutinising ad creative to the same degree as media professionals.
This was borne out in System1’s creative testing of the ad, prompting the company’s SVP, Andrew Tindall, to declare “I was wrong…This ad, that gave me that visceral reaction, just secured effectiveness gold…No-one had any sort of negative reaction to it”.
As the quality (and adherence to the laws of anatomy and physics) of AI-generated imagery and video continues to improve, its use in creative is going to be increasingly difficult for regular viewers - and, in time, seasoned professionals - to detect.
I’m not sure I would have spotted the AI-generated anomalies in Vodafone’s latest ad (below) had I not been primed to look for them.
2.) Knowledge of AI use triggers a set of negative assumptions for many people
As Tindall acknowledges in an opinion piece for The Drum, had System1 informed viewers the ads were mostly comprised of AI-generated visuals, they would likely have garnered much less positive reactions.
Many people associate AI with a threat to jobs, particularly in creative domains. Video which features AI-generated human likenesses are particularly triggering (in a way that animations featuring computer generated characters aren’t), as it suggests the displacement of human actors.
Many also have an awareness of the majority of generative AI models being trained on unlicensed material.
This tweet post on X by Alex Hirsch, creator of Gravity Falls, combines these concerns:
“FUN FACT: @CocaCola is “red” because it’s made from the blood of out-of-work artists! #HolidayFactz”.
And lots of us are not looking forward to the inevitable tsunami of ‘AI slop’ that generative AI will precipitate.
Almost every high-profile instance of generative AI being used in creative over the last 2 years has prompted strong reactions along these lines.
3.) The quality bar is higher for high-profile brands telling human stories
The first two takeouts from my recent round-up of AI in Video Advertising Creative, were:
The quality bar for video is high and viewers will likely be unforgiving of any AI-generated anomalies (unless you’re parodying).
Avoid falling between two stools. Creative should either be obviously AI-generated and you’re making a comedic/educational/stylistic virtue of that or so good that the use of AI isn’t apparent.
Whilst the viewers System1 were assessing may not have picked up on the AI-generated elements within the ad, the creative from a brand as high-profile as Coca-Cola is always going to be heavily scrutinised.
Attempting to tell a human story (versus a passport or potato) raises the quality bar higher. Remaking one of the most beloved ads of all time pushes the bar sky high and Coca-Cola hasn’t cleared that bar here.
Having ‘Always the Real Thing’ as your slogan also really doesn’t help.
4.) Creating engaging creative using AI requires a large number of decisions
Last week Jason Zada, Founder of Secret Level (one of three studios who produced the ads), posted a ‘making of’ video on LinkedIn “showcasing how human creativity and advanced technology came together” to create the ad. It references 17 artists creating 18,000 images. That’s a lot of decisions.
This chimes with my experience of using generative AI tools. I can explore more ideas more quickly but it requires more iterations and no less creative decision making.
“Using AI we could iterate faster than ever but while AI streamlined the process, it was really our team’s artistic decisions that helped shape every frame and scene” (Davide Bianca, Co-Director, Secret Level)
5.) When it comes to creative output, ‘AI-assisted’ may be a more helpful label than ‘AI-generated’
I’m not sure ‘AI-generated’ is the most helpful label in this context. It risks giving the impression that someone typed ‘remake the 1995 Holidays Are Coming Coke ad’ into an AI video generator and then uploaded the resulting output to YouTube.
As I’ve written previously, use of AI in creative work is not a binary, all-or-nothing equation. Some creative endeavours will continue to make zero use of AI, but many will use it to a greater or lesser degree.
The producers of the Secret Level ad “orchestrated original music performed by a real choir and cast real people as the likeness for our characters” but decided not to hire real trucks, fly a production team somewhere snowy, ship in a tonne of artificial snow or rely solely on classic CGI (note: the original 1995 ad included effects by Industrial Light and Magic).
Whilst elements of the ads may have been AI-generated, the campaign as a whole was AI-assisted.
6.) The environmental impact of AI use in creative isn’t a straightforward calculation
“I feel like I’m watching the death of art and our planet unfold in front of my eyes“ (@MoistyJc on X)
I’m very concerned about the environmental impact of training and running AI models but accurate information on this is hard to come by.
Without any expectation of accuracy, I asked ChatGPT to have a crack at estimating the relative environmental impact of filming the original 1995 ad versus the 2024 AI-assisted remake.
It estimated that the original “likely produced several tonnes of CO₂, depending on scale (average film productions can emit 100-200 tonnes CO₂ for large commercials)”
It estimated the remake produced “0.1-0.5 tonnes CO₂ for 18,000 images”.
Of course, those estimates may or may not have any basis in fact (ChatGPT’s response was citation-free) and the remake estimate only takes account of the image generation.
A bigger challenge is how to apportion the significant CO₂ output of the original training of the AI models used. Those models existed prior to Coke’s use of them, but all AI use is, in aggregate, encouraging the training of future models.
7.) Some of the world’s biggest brands are generating creative using image and video models which have been trained on unlicensed data
Coca-Cola haven’t disclosed the names of the AI models used to generate the imagery in the ads, However, Leonardo, Luma, Runway, Kling, Stable Diffusion and DALL-E have all been mentioned in interviews with representatives from the three studios who made the ads (Secret Level, Silverside AI, and Wild Card).
The visuals in Motorola’s recent AI-assisted campaign were reportedly generated using a range of models, including Midjourney, Krea and Luma.
Adobe Firefly is currently the only video model which claims to be trained exclusively on licensed images, although many Adobe Stock contributors are unhappy they weren’t consulted on images they’d previously submitted being used for AI training and images generated by other AI models have been found within Adobe Stock.
It’s striking that whilst many agencies and production houses are steering clear of using AI models trained on unlicensed data in consumer-facing work, some of the biggest brands in the world have decided it’s worth the legal and talent relationship risk to make wholesale use of these models in commercial work.
8.) Coca-Cola were open about the use of AI, but kept the in-ad messaging low-key
Coca-Cola UK put out a ‘Holidays Are Coming!’ press release on 5th November, leading with the real-world Truck Tour and mentioning “an AI-generated reinterpretation of the classic advert” in a ‘campaign will also include’ section lower down the release.
The posts about how the ads were made came from the production studios rather than Coke.
The ads themselves includes a subtle ‘*Created by Real Magic AI’ message in a small font for a few seconds at the start of the ad (I think ‘with AI’ would have been better than ‘by AI’).
Interestingly, the Vodafone ad includes no AI disclaimer.
9.) I don’t believe Coca-Cola did this to save money
Coca-Cola has made $28bn dollars profit from $46bn revenue in the past year. Whilst these ads may well have been cheaper to create than an AI-free ad, I don’t believe reduced production costs was the primary driver of the decision to make extensive use of AI.
Secret Level’s ‘making of’ LinkedIn video states that the content was “all created in under 3 weeks”. However, I don’t believe speed was the primary driver (I think Coke have known Christmas was coming for a while).
The video also talks up the ability to create personalised/localised versions of the campaign. I think this was also secondary.
I believe their primary driver was to garner attention, drive engagement and put Coca-Cola front of mind for consumers ahead of the holiday season. I’d suggest these ads have achieved that. Viewers unaware the ads were made with AI appear to love them. Those who are angry AI was used have been busy talking about them, amplifying their online impact. Whilst some have threatened a boycott, I would wager this campaign will be a significant net positive for sales of Coke over the festive period (of course, AI-free ads may have been much more impactful).
You might not love the ads (I don’t). You might feel the extent of AI use in these ads is too much, too soon (I do). You might feel that creatives should be consulted and remunerated for content used for AI model training (I do). You might be worried about the impact of AI model training/inference on the environment (I am). However, blanket opposition to the use of AI in advertising creative and always assuming the worst in terms of creative/human/environmental impact feels reductive and unhelpful.
I would love to see greater focus on what different models were trained on, how contributors to AI model training can be consulted and remunerated, the environmental impact of both training and inference and where it makes sense to use AI in creative and where it doesn’t (depicting a burning oil rig = yes, conveying human emotions = maybe not).
It will be interesting to see whether the Vodafone ad (released a few days ago) prompts an equally strong reaction or whether the smaller number of obvious anomalies and the greenfield canvas (vs reimagining a beloved classic) tempers the outrage and prompts a more nuanced conversation.