Brand Safety and Generative AI: Protecting Reputation
Brand safety used to mean where your ads ran. With generative AI the risk comes from inside: the incidents so far, what audiences think, and my guardrails.
Brand safety used to be an adjacency problem: keeping your ads off pages you would be embarrassed to fund. Generative AI added a second, worse problem, because now the embarrassing content can come from inside the building - published under your logo, spoken by your chatbot, or starring in your own ad. The adjacency version had vendors and blocklists. The new version has your workflow or nothing.
I think of this as the reputational layer of the governance work I keep writing about. The hallucination guardrails protect you from publishing false things. Brand safety is the wider net: protecting yourself from publishing things that are merely wrong for you - off-voice, tone-deaf, obviously synthetic, or beneath the standard your audience associates with your name. This post covers where that risk actually lives, what the documented incidents teach, what the audience research says, and the approval system I would run. (The seat that owns all of this day to day has its own playbook now.)
The three surfaces where AI can hurt a brand
Mapping the exposure is the first step, and I find it splits cleanly into three surfaces with different risk profiles.
Published content - the blog posts, product copy and social output AI now drafts at volume. The risk here is rarely one catastrophic error. It is erosion: content that reads as machine-made, at a volume no editor is really reading, until the audience quietly recalibrates what your name is worth.
Live surfaces - chatbots and assistants that speak for the brand in real time with no editor between the model and the customer. This is the highest-severity surface, because the failure is instant, public and attributable, and a tribunal has already ruled on whose problem it is.
The creative itself - AI-generated imagery, video and ad concepts. The newest surface, and the one where the risk is least about accuracy: creative can be flawless and still misjudge how audiences feel about AI making it.
What the incidents so far actually teach
A handful of cases have defined this territory, and each one maps to a surface.
Sports Illustrated and the authors who did not exist. In November 2023, Futurism reported that Sports Illustrated had been publishing product reviews under fake author names with AI-generated headshots - one bought from a marketplace that described the face as a “neutral white young-adult male with short brown hair and blue eyes.” After Futurism asked questions, the fake authors vanished from the site, and the publisher blamed a contractor. The lesson I take is about provenance, not tools: the scandal was never “they used AI to help write reviews.” It was the fabricated humans wrapped around the content - a deception the audience was meant not to notice. A famous brand spent a century of credibility on a content-farm arrangement, and the discount was steep.
Air Canada and the chatbot’s promise. The Moffatt v. Air Canada ruling established that a company is responsible for what its chatbot tells customers - the airline’s argument that the bot was a separate entity got the reception it deserved. I covered the accuracy mechanics in the hallucination post. The brand safety reading is different: the reputational cost dwarfed the CA$812 award, because the story every outlet ran was “airline argues its own chatbot is not its problem.” The incident response was the amplifier. Own your AI’s mistakes fast, or the argument becomes the headline.
Google’s “Dear Sydney” ad. In August 2024, Google pulled a Gemini ad from its Olympics rotation that showed a father using AI to help his daughter write a fan letter to a track star. Nothing in the ad was false, and by Google’s own statement it “tested well before airing.” Audiences hated it anyway - the reading was that AI had been inserted into a moment that belonged to a child and a parent. This is the case I show people who think brand safety is a fact-checking problem. The third surface fails on sentiment, and sentiment does not show up in a compliance review.
What audiences actually think of AI-made marketing
The pattern in the Dear Sydney backlash is measurable, and I would want every approval workflow calibrated by it rather than by the team’s own enthusiasm.
NIQ’s consumer research on AI-generated ads, which included EEG measurement of brain activity for about 150 of its 2,000+ participants, found AI-generated ads described as more “annoying,” “boring” and “confusing” than traditional ads - and, more striking, that AI-generated ads elicited weaker memory activation even when viewers rated them as high quality. The conscious opinion and the measured response both came back negative.
YouGov’s 17-market survey from early 2024 found only around a third of consumers fairly or very comfortable with AI-created brand ambassadors, and about two in five comfortable with AI-generated or edited product images. Comfort was lowest in several of the biggest Western ad markets.
My read is not “never use AI in creative” - the numbers will keep moving, and plenty of AI-assisted work passes unnoticed precisely because it is good. My read is that visibly synthetic work starts from a trust deficit, so the more emotionally loaded the placement, the more human the work should be, and the more certainly you should know your disclosure position before someone asks.
The guardrail system I would run
Four pieces, sized so a real team will actually keep doing them.
A use map, written down. One page: where AI is allowed freely (internal drafts, research synthesis, variant generation), where it is allowed with review (anything published), and where it is not allowed (final creative for tentpole moments, crisis communications, anything trading on personal authenticity - the Dear Sydney category). The map’s value is that the argument happens once, calmly, instead of per-asset under deadline.
Risk-tiered approval, not uniform approval. Uniform review of everything is how review dies. My tiers: routine volume content gets the standard editor pass plus the verification gate for claims. Anything customer-facing and paid gets a named approver. Anything on the use map’s sensitive list gets a second set of eyes from outside the team that made it, because the team that made it tested well before airing too.
Voice and claims fences at the source. Most off-brand output is preventable upstream: voice rules and real examples in every standing prompt, banned claims and banned phrases in the shared system prompt, and the never-invent-a-number rule from the hallucination guardrails. Fences do not replace review. They cut the failure rate review has to catch.
An incident habit. Decide now who gets told, who decides on the correction and how fast it ships when AI-touched content goes wrong in public. Air Canada is the cautionary shape: the lasting damage came from the defense, not the error. A brand that corrects fast and plainly usually gets a one-day story. A brand that argues gets a precedent named after it.
FAQ
What does brand safety mean for generative AI? Traditionally it meant controlling ad adjacency - where your ads appear. With generative AI it extends to what your own tools produce under your name: published content, chatbot answers and creative. The controls shift from vendor blocklists to your own use policy, approval workflow and prompt-level fences - the one-page structure that holds all three is the marketing governance framework.
Is AI-generated content itself a brand safety risk? Used with review, mostly no - audiences react to quality and honesty, not to tooling they cannot see. The documented damage clusters around three behaviors: deceiving the audience about provenance, letting unsupervised systems speak for the brand, and putting visibly synthetic work into emotionally loaded placements. Avoid those three and most of the risk goes with them.
Should we disclose that content or ads used AI? It depends on the surface, the jurisdiction and the platform, and I keep a separate guide on exactly that question. The brand safety rule underneath it: never build on the assumption that the AI’s involvement stays secret. Sports Illustrated made that bet.
Who should own AI brand safety on a marketing team? The same person who owns brand standards, with the use map giving them jurisdiction over AI specifically. What matters is that it is a named person rather than a shared mood, and that the wider governance framing gives them backing when they say no to something the calendar wants.
This post is part of the ethics and governance guides, where the accuracy layer is the hallucination guardrail system and the transparency layer is the AI disclosure guide. The regulatory floor is in the FTC rules and the EU AI Act checklist, and the wider practice starts at the complete guide to AI marketing.