If you have watched a brand walk into an AI-related controversy over the past two years, you already know the pattern. A campaign looks fine in the review room, gets released, and then the internet decides it means something the brand never intended. So here is the honest verdict up front: AI marketing risks are real and well documented, but the evidence does not say AI is unsafe to use. It says something narrower and more useful. Reputational damage shows up when AI is allowed to speak for the brand, stand in for a real person, make a claim nobody checked, or go out the door without a named human who owns the outcome. The evidence for that pattern is strong. What follows is what actually happened, what regulators and standards bodies have said, and a checklist you can use before your next AI-assisted campaign goes live.
What counts as an AI marketing risk
Before you can guard against something, you need to know what you are actually guarding against. For this piece, an AI marketing risk is any foreseeable way an AI-assisted activity can damage trust, meaning, accuracy, representation, consent, accountability, or consistency. That is a wider net than "the ad looked fake." A chatbot that invents a discount is a risk. A generated image that leans on a tired stereotype is a risk. A byline attached to nobody in particular is a risk. A synthetic voice standing in for a real person's endorsement is a risk.
This is a brand and reputational risk briefing specifically. It does not cover data governance, access controls, or model security, because those are a different problem with a different owner. When people talk about AI marketing controversies, they are almost always talking about the seven risks above, not a security incident. Keeping that boundary clear matters, because it stops a real conversation about audience trust from turning into a vague warning about "AI safety" that nobody can act on.
The incidents that actually happened
The clearest way to understand AI brand risk marketing teams are being asked to manage is to look at what has already gone wrong, in public, at real companies.
Coca-Cola released an AI-made holiday ad in November 2024, built to echo its own 1995 "Holidays Are Coming" commercial. NBC News reported that the response online was sharp: people called the spot "soulless" and said it showed no actual creativity. The complaint was not that the trucks or the snow looked wrong. It was that a brand built on holiday warmth had chosen a production method a lot of viewers read as emotionally hollow, and creative professionals connected that choice to worries about AI replacing human artists. Coca-Cola said the work came from a mix of human storytellers and generative AI, but the ad's technical quality did not settle the argument about what the choice meant.
Toys "R" Us ran into a similar wall a few months earlier. It debuted an AI-generated brand film at Cannes, made with OpenAI's Sora, showing a young version of its late founder alongside Geoffrey the Giraffe. NBC News reported that the backlash arrived fast, centered on job displacement in acting, writing, and design, with some viewers describing the result as uncanny. About a dozen people had worked on the piece for three months, and the studio called the test a success. None of that stopped the audience from judging the finished film on how it made them feel, not on how much human effort sat behind it.
Google pulled its own Olympics ad for Gemini in August 2024 after backlash, according to CNBC. The spot showed a father prompting Gemini to help his daughter write a fan letter to an athlete. Critics said it replaced a personal moment with automation, especially with a child at the center of it. Google said the ad had tested well before it aired. That detail matters more than it might seem: a campaign can pass a normal pre-launch test and still trigger a values objection that has nothing to do with execution quality.
Levi's announced in 2023 that it would use AI-generated models to increase the diversity of its imagery, and Ad Age reported the backlash was immediate. The objection was straightforward: if the goal was more representation, why not hire more diverse human models instead of generating the appearance of them? A stated good intention did not protect the campaign from looking, to a lot of people, like representation theater.
Away from creative work, the accuracy failures were just as public. CNET issued corrections on 41 of 77 stories written with an AI tool, The Verge reported in 2023, after months of publishing them without a clear announcement. One article on compound interest contained a pile of errors, and critics flagged language in other pieces that looked plagiarized. A disclaimer promising to fix errors as they were found came only after the scale of the problem became visible, which is the wrong order for a credibility-sensitive publisher.
Sports Illustrated ran into a version of the same problem from a different angle, as PBS News reported in late 2023. Product-review stories carried bylines from authors nobody could verify, including one whose profile photo turned out to be a stock AI-generated portrait. When questioned, the publication removed the content, ended its relationship with the third-party vendor that produced the pieces, and denied that the articles themselves were AI-written. The specific authorship claim stayed disputed. What was not in dispute was that the magazine could not explain who these bylines actually belonged to, and that gap became the story.
A Chevrolet dealership in Watsonville, California, learned how public a customer-facing AI system can go wrong. Its ChatGPT-powered website assistant was talked into agreeing to sell a 2024 Chevy Tahoe for one dollar, and separately recommended a Ford F-150 to a shopper on a General Motors dealer's site, Driving reported. Nobody actually got a dollar truck, but the screenshots spread anyway, and the dealership's IT team shut the tool down.
Then there is the risk of a likeness used without permission at all. Tom Hanks told his 9.5 million Instagram followers in 2023 that a dental-plan ad used an AI version of his image, and that he had nothing to do with it, according to The Guardian. He was the wronged party here, not the advertiser, but his credibility still had to absorb the correction. That is the uncomfortable part of a synthetic-likeness incident: the audience often sees the fake before anyone sees the correction.
What the studies add, and where they stop
Incidents show you what can happen. Studies tell you how often audiences notice or care, and here the picture gets more specific than "people don't like AI."
A 2024 study from the Nuremberg Institute for Market Decisions surveyed 1,000 people each in the US, UK, and Germany, and ran controlled experiments alongside the survey. Only 25% of respondents believed they could actually recognize AI-generated content, and just 20% said they trusted AI itself. In the experiments, identical ads rated worse on emotional measures and got fewer clicks when people were told the image was AI-generated instead of a photograph. That effect was not uniform. People who already held a positive view of AI reacted less negatively, and people who believed creativity had to be human reacted more negatively.
A 2025 study in the Journal of Retailing and Consumer Services found something more specific still: AI disclosure hurt trust more when the ad leaned on an intangible human element, like the credibility of a pictured expert, than when AI was used for tangible things like equipment shots. Trust recovered when AI stayed in the background and a real human carried the emotional weight of the message. That is a genuinely useful distinction for a marketing leader deciding where AI belongs in a given asset, and it is a long way from "never disclose" or "always disclose."
A separate 2024 field study generated 1,110 images across ChatGPT, Midjourney, and Canva and found the outputs reproduced sex and racial stereotypes at a higher rate than expected. That is not a claim that every image from every tool will carry a stereotype. It is evidence that generated visuals need the same representational review a human-shot campaign would get, not less.
Industry-side research adds a different kind of evidence. The World Federation of Advertisers surveyed 27 multinational brands representing $71 billion in ad spend in 2026 and found 78% already used AI-generated or AI-enhanced content in consumer-facing marketing, while 82% called transparency essential for protecting brand reputation. Those are marketer opinions and reported practices, not consumer research, but they show a real split forming: broad AI adoption running ahead of a settled disclosure standard.
None of these studies show that AI disclosure always hurts a campaign, or that audiences reject anything touched by AI. What they consistently show is that the risk concentrates around the same places: synthetic humans, unverified claims, and content that stands in for a personal or emotional moment.
What the regulators and standards bodies actually say
A campaign that seems fine internally can still run into a rulebook that has already caught up. The UK's Advertising Standards Authority and CAP said in May 2025 that existing advertising rules apply no matter how content is generated, edited, or targeted, and that there is no blanket UK requirement to disclose AI use in every ad. Their actual test is whether an audience is likely to be misled without disclosure, and whether disclosure would fix that specific problem. Crucially, they were explicit that disclosure cannot rescue a fundamentally misleading message: a labeled AI image that inflates real-world results is still misleading with the label attached.
New Zealand's Advertising Standards Authority made a similar point in March 2024: its codes are technology-neutral, and an AI-generated ad still has to avoid misleading, exploiting, or offending its audience. It specifically called out generated images as needing a bias and stereotype check before publication, which lines up with the field study above.
Google's own advertising documentation offers a disclosure setting for AI-generated or edited assets, but it states plainly that using the label does not itself guarantee legal compliance. That is worth sitting with: a platform checkbox is not a substantiated claim, a consent record, or a completed human review. Treating it as one is how a brand ends up thinking it has covered a risk it has not actually addressed.
The US Federal Trade Commission has been the most active enforcer. Its September 2024 Operation AI Comply announcement covered five cases against companies making unsupported AI claims, including a "robot lawyer" service that agreed to a $193,000 settlement and restrictions on future claims. Its August 2024 final rule on fake reviews and testimonials bans manufactured reviews, including AI-generated ones that appear to come from real customers who do not exist, and covers insider reviews, bought social proof, and suppressed negative feedback. The agency's position is simple: there is no AI exemption from existing consumer-protection law.
One case is worth a specific caveat, because getting it wrong would undercut the credibility of this whole piece. The FTC's 2024 consent order involving Rytr, a writing tool the agency alleged could generate detailed but unverified consumer reviews, was reopened and set aside by the agency in December 2025. Describe that matter accurately as a case with a contested and evolving status, not as a settled, currently active ban.
What the evidence does not prove
It is just as important to say what none of this establishes, because overstating the case is its own credibility risk.
The evidence does not show that consumers reject all AI-generated advertising outright, or that disclosure always reduces conversion. It does not establish a specific sales loss for Coca-Cola, or confirm that Sports Illustrated's disputed articles were in fact AI-written, since the publisher denied it. Nobody actually bought a dollar Tahoe. And a compliance label on an ad does not make the underlying claim true, any more than a human reviewer eliminates risk just by existing in the workflow. Human review only helps when the reviewer actually checks the claim, tests the system adversarially, and has the authority to stop publication before it goes out. Keep the distinction between a documented incident, a company denial, and a proven mechanism, because that distinction is exactly what separates a grounded risk assessment from an urban legend.
A working checklist before you publish
Pull the incidents and the guidance together, and the shape of AI brand risk marketing teams have to manage stops feeling abstract. A practical pattern falls out. Before an AI-assisted asset goes live, work through this:
Classify the AI's role. Background cleanup and internal brainstorming carry little risk. A synthetic person, a customer-facing chatbot, or final claims copy carries a lot. The closer AI gets to a real person, a price, or a product outcome, the more scrutiny the piece needs.
Ask what the audience will actually infer, not just whether the asset is technically labeled. Will a viewer believe a real person endorsed this? Will they believe the product performs exactly as shown? Would they feel misled if they learned the production method after the fact?
Separate factual claims from creative expression, and give every factual claim, whether it is a price, a comparison, or a performance number, a named owner and a source. A label does not substitute for substantiation.
Never invent social proof. A review or testimonial needs a real person behind it. Do not generate a plausible-sounding customer story because a prompt can produce one convincingly.
Confirm consent for any likeness, voice, or endorsement, including its scope: which channels, which markets, and for how long.
Give representation a real review, with a diverse group looking at who is shown, who is missing, and whether a diversity claim is actually being served or just simulated.
Red-team any customer-facing system. Try impossible offers, competitor questions, and attempts to get it to contradict your own pricing or policy, and keep the transcripts. A tool that behaves in a demo can still fail in public.
Keep a provenance record for anything published externally: who owns it, what was checked, who reviewed the final version, and why any disclosure decision was made the way it was. This matters most for expert content, reviews, and anything carrying a byline.
Name one person with the authority to pull an asset down. A campaign that goes wrong in public needs a fast, documented, and proportionate response, not a meeting.
The verdict, and what would change it
The strongest reading of the evidence is this: AI marketing controversies keep following the same shape. Something a brand thought was a production efficiency turned out to be a values question the audience cared about, or a claim nobody checked before it went out, or a person's likeness nobody actually cleared. None of that argues for avoiding AI in marketing altogether. It argues for treating AI's proximity to real people, real claims, and real customer decisions as the thing that actually determines the level of scrutiny a piece of content needs. This verdict would weaken if future incidents showed brands absorbing genuine, well-tested AI use with no backlash at scale, or if regulators concluded existing consumer-protection law does not adequately cover AI-assisted claims. Right now, neither of those things has happened.
Part of managing this well is simply knowing what is being said about your brand once content is out in the world, since AI systems increasingly summarize and repeat brand claims the same way audiences do. DeepSmith's AI Search Visibility module tracks how AI engines describe your brand across the questions your buyers actually ask, so a factual error or a stale claim about your company doesn't sit uncorrected because nobody noticed it. That does not replace the checklist above. It just means you find out about a problem from your own dashboard rather than from a screenshot going around online.
