Bias Toxicity, AI, and Contextual Advertising
Inviting AI to play a leading role in the match game of contextual advertising raises a number of red flags, and one of the most concerning is the built-in bias in the way that LLMs are trained. As Digital Culture Group’s Crystal Foote points out, “A lot of adtech companies just aren’t that diverse,” and that can leave significant blind spots and “stereotypical insights” that find their way into AI models that manifest themselves in ad decisions. Foote, Sargeway’s Sarge Sargent, and Dataxis’ Ophelie Boucaud discuss what Sargent calls “bias toxicity” and its implications for AI-powered scene analysis, contextual advertising, and audience representation in this clip from Streaming Media Connect 2026.
Understanding Cultural Context
Boucaud suggests that one reason that AI bias might distort perceptions of cultural relevance and lead to unfortunate ad placement is that the sheer amount of data an AI-powered system can absorb reduces or drastically limits the effectiveness of human oversight.
"Using AI in the pipeline is expanding the scale of the amount of data that we're able to work with," posits Boucaud. "We still need humans in the loop to make sure that we understand the nuances, that we understand the context, especially the cultural resonance of everything." Turning to Foote, she asks, "How do you make sure that the AI is not pushing you into the wrong direction, that it's not hallucinating? Where do you put humans in the loop?"
Foote argues that the problem with often starts earlier with the cultural blind spots of the humans who train the LLM and the composition of the teams who build the systems. "I think that it has to start from just building out the AI platform and making sure the framework looks at all audiences," she says. "A lot of ad tech companies are not diverse. Let's just call a spade a spade." If the company that develops the LLM is not diverse, "there's going to be bias, and what we're thinking of as hallucinations is actually just inaccurate data that are the outputs from this AI because there wasn't any thought process going into building it out to make sure that it's a fair representation of all the audiences, in the US or globally. I think that also there needs to be more governance over AI and there needs to be more QA'ing. Our engineering team [does] QA every day. We also make sure we QA against other models too to see what the outputs would be. I think that's extremely important, especially dealing with big brands."
Foote goes on to explain the consequences of these failures of representation when inappropriate or insensitive ads are served, whether they stem from cultural myopia at the training stage or insufficient oversight. The brands that companies like DCG work with "don't want to get it wrong because that is tied to transactions that are tied to their bottom line revenue. And so we take this very seriously and we make sure that [the LLM] doesn't have any type of stereotypical audience insights built in. Even if it's not intentional, we want to make sure that it represents all audiences."
"That's a very important thing when it comes to understanding the cultural context of where the audiences sit," Boucaud agrees.
Counteracting Bias Toxicity
So, what countermeasures can ad tech companies take to avoid culturally insensitive ad decisions, or misfires that cause AI to "hallucinate a scene or a sentiment that really doesn't fit the vibe?" Turning to Sargent, Boucaud asks, "Do you have examples of how a human has to step into recalibrate?"
"It's exactly what Crystal said," Sargent affirms. "You have to have a human in the loop to dissolve some of these hallucinations." The root of the problem, he says, is "bias toxicity. If you put The Wiz and The Wizard of Oz through an AI, you'll likely get the same output from a scene analysis. There'll be some variances because it depends on the LLM. If it's in tune to some of the differences in, for example, just the music," he says, it might recognize that "The Wiz is more rock, more blues, more Black, if you will, whereas The Wizard of Oz is more ballad-y, more traditional. So your LLM has to understand the differences because at the program level, they both describe the same thing. It's a musical about a girl who wants to get back home. But that's just the baseline. If your LLM doesn't go beyond that, then you're not going to grab the entire sentiment of the overall program itself or at the different scene levels."
He goes on to say that an LLM properly trained on the actual content will "go deeper" and recognize "the significant differences between a Black Dorothy an a white Dorothy and the intended audiences." This, in particular, is where bias toxicity comes in. "African-American vernacular is deemed more toxic or has a higher toxicity level than what would be your standard white American English. So if that's built into your LLM, then you're going to get hallucinations. You're going to just get bad results in your scene analysis."
Join us November 9–11, 2026 for more thought leadership, actionable insights, and lively debate at Streaming Media Connect 2026! Registration is open!
Related Articles
The companies covered here have substantially leveraged AI or incorporated it into their currently shipping tech. The representatives I spoke to discuss some very interesting approaches for what essentially is image interpretation, including detecting if content has been tampered with using AI, local AI processing, and automated clip generation such as vertical formatting.
07 Sep 2026
AI's potential to transform CTV and OTT ad preparation, ad serving, ad placement, contextualization, video analysis, performance, and reporting is fairly well understood and rarely understated. But how are streaming platforms, content owners, and advertisers actually operationalizing AI today? Revry's Alia J. Daniels, Digital Culture Group's Crystal Foote, Dataxis' Ophelie Boucaud, and Sargeway's Sarge Sargent discuss how publishers are leveraging AI's advantages and also addressing some of its limitations in this clip from Streaming Media Connect 2026.
01 Sep 2026
CTV contextual advertising, like many things in life, is all about making good decisions, and making informed decisions based on a wealth of data means leveraging the right tools—often AI-driven—to gather and distill and interpret that data. Sometimes developing sound contextual media plans involves working with in-house tech and other times it means working with third-party tools, as Team Whistle (a DAZN company) president Joe Caporoso and Intersection CEO Chris Grosso explain in this discussion with SVTA subject matter expert Bhavesh Upadhyaya in this clip from Streaming Media Connect 2026.
09 Mar 2026
"Contextual advertising is a really hot topic in TV advertising right now," declared TVREV's Alan Wolk at Streaming Media Connect 2024. But what exactly is it, and how is its growing presence in CTV and streaming changing the TV experience and particularly the way advertisers buy media? Wolk, Estrella MediaCo's Christina Chung, and Mad Leo Consulting's C.J. Leonard explore these issues in this clip from their panel at Streaming Media Connect.
04 Oct 2024