How AI Bias Impacts Brand Trust and What Marketers Can Do About It

0:05 David Burkowitz: Hey everyone, welcome back to AI Insiders. I'm David Berkowitz at AI Marketers Guild.

It's a pleasure to be here and I'm especially grateful to have a returning speaker today. We try to keep fresh faces in the lineup, but Catharine Montgomery helped educate us last year about AI bias and it's a topic we should be talking about more. It affects people at work and in their personal lives — how things are being shaped and how they're shaping us, often without our conscious knowledge. Catharine has done tremendous research and put out a second edition of her study, so welcome back, Catharine.

1:07 Catharine Montgomery: Thank you, David. I didn't know returning speakers were rare — I feel honored. I'm going to share some slides. We’ll talk about generative AI bias, the trust factor that holds some brands back and how they can make it an advantage by addressing it. We'll also review our second generative AI biases survey and how that data impacts brands.

2:12 Catharine Montgomery: A bit about me: I'm the founder and CEO of Better Together, an AI-forward agency. We put humans first and amplify work through AI. We help brands navigate high-stakes conversations about bias, equity, and technology. We only work on campaigns that make a positive impact. I’ve seen generative AI repeat patterns that exclude people — systemic issues from real life are coming through AI. There's an opportunity to address those systemic issues through AI rather than exacerbate them. That’s why we published our first survey last year and why we plan to do it annually.

3:12 Catharine Montgomery: Biases in generative AI are costing brands credibility but also create a competitive advantage for those who address them. Most brands aren't educated on how to look for biases in generative AI, so marketers can play an educational role. One of our sponsors said companies that figure out fair generative AI will own the trust advantage for the next decade. This is a long-term competitive advantage; consumers will increasingly demand companies address biases in AI.

5:17 Catharine Montgomery: I'll start with Shakespeare: "I am not what I am." Anthropic used this idea to explain alignment faking — models pretending to share values but reverting to original training. A model trained one way then fine-tuned to be neutral can revert to prior behavior; you can't just pretend a model changed. That's why it's critical to build inclusive technology from the start and avoid baking in unaddressed biases.

6:17 Catharine Montgomery: For example, a model could learn an ideological slant and later be trained to be politically neutral, but it may still revert to its original behavior. Alignment faking is a security and trust issue in generative AI.

7:23 David Burkowitz: I referenced a book called Doppelgänger by Naomi Klein, which fits this theme — confronting mirror images and implications for AI. Interesting rabbit hole.

8:31 Catharine Montgomery: Another example: Grok. In its recent version, Grok used an instruction embedded in the system prompt that referenced Elon Musk's tweets for controversial topics. As a result, Grok produced anti-Semitic content and praise of Hitler based on those tweets. LLM chatbots are trained on unfiltered online data; this is exactly what consumers fear — they can't always be trusted. At New York Tech Week, the founder of Girls Who Code said her 18-year-old son doesn't use AI because he doesn’t trust it. Younger generations, including Gen Alpha, are especially concerned about bias.

9:31 Catharine Montgomery: Grok apologized and corrected the algorithm, but these incidents show how easily a single actor can influence model outputs. One person with influence can change a system prompt and affect responses in real time.

10:32 David Burkowitz: Can I ask about that power dynamic? On one hand, a kill switch might stop harmful outputs, but it's alarming that one person can make changes that affect millions of users. Thoughts on that concentration of power?

11:36 Catharine Montgomery: It's unbelievable that someone could hold that much power. That's where the trust factor comes in — consumers often don't trust AI because these changes can happen. It’s a major concern.

12:37 David Burkowitz: It's striking that regardless of who that person is — a universally trusted figure or not — their changes can have wide implications. Many tools are trained on large public data sets, but the ability for one person to change algorithmic behavior overnight is a power play we've seen recently.

13:43 Catharine Montgomery: The real issue is embedded values in systems. Companies that address this will capture broader markets and boost engagement. For example, a change to an algorithm at 3:00 a.m. between two influential people can alter outputs; that’s alarming. We'll see these concerns reflected in the data.

14:55 Catharine Montgomery: Axios reported a trust crisis: trust in AI has fallen significantly. Trusted brands are more likely to be bought and recommended. Nearly nine in 10 adult consumers globally say trust matters when buying a brand — an opportunity for marketers to build that trust through transparency and education.

15:59 Catharine Montgomery: I use Midjourney and like the outputs, but I see many biased examples. I ran a prompt for a female technologist and got only Asian women across multiple runs — that signals a training bias associating technology with Asian women. Another prompt for global hunger returned only images of impoverished Black children in Africa, ignoring people with disabilities and LGBTQ individuals. If everyday users assume these images are neutral, they’ll perpetuate biases in marketing.

17:06 Catharine Montgomery: Another high-profile example was Guess and Vogue’s AI-generated model: the output conformed to narrow beauty norms — a young, thin, white, blonde, blue-eyed woman — and Vogue only disclosed the image was AI-generated in a barely visible caption. That lacks transparency and perpetuates systemic biases around beauty and representation.

18:04 Catharine Montgomery: Facial recognition tools also reveal biases: some systems fail to recognize certain people, whether due to race, gender, or facial hair changes. Consumers worry about facial and voice recognition. These biases are not new; they reflect long-standing systemic issues.

19:04 Catharine Montgomery: The broader point: embedded values in models reflect who designs them and the data used. If the training data and design process are biased, outputs will be biased too.

21:17 Catharine Montgomery: Look at ownership of leading LLM companies — many are owned or led by men. That influences design decisions and priorities. Humans introduce bias at many stages: data selection, model training, fine-tuning parameters, system prompts, and guardrail implementation. We must consider these phases when designing responsible AI.

22:19 Catharine Montgomery: Marketing AI Institute has covered these bias pathways; episode 158 is a useful resource. Understanding where bias enters helps brands act.

24:29 Catharine Montgomery: Our research frames this as an ethics, value, and trust issue — inconsistent brand experiences stemming from generative AI bias are a CEO-level concern. For the survey, I was urged to show bottom-line impacts; so this year we asked questions that tie bias mitigation to business outcomes to engage more decision-makers.

25:27 Catharine Montgomery: Brands often prioritize being first to market, overlooking bias mitigation. For example, companies releasing ChatGPT-5 could have addressed known biases before release. Bias should be a design priority, not an afterthought.

26:34 Catharine Montgomery: Consumer reality: 83% of consumers have used generative AI and they're paying attention to bias. As one respondent said, when AI bias goes unchecked, customers judge your AI practices. Consumers notice outputs, know which tools to trust, and see the mental health impact of harmful content.

27:38 Catharine Montgomery: From a financial perspective, 92% of respondents believe companies must address generative AI bias. Fairness ranked second when choosing a GenAI tool after accuracy. There's a $2.6 trillion market opportunity for companies that address bias. By 2030, I believe a substantial audience will demand bias remediation.

28:37 Catharine Montgomery: Survey findings: 25% of consumers prefer to buy from brands addressing bias, 59% assign a trust premium to fair GenAI, and 76% seek inclusive brands. The Vogue/Guess incident shows consumers expect honesty and inclusion in AI use.

29:39 Catharine Montgomery: Industries most at risk: healthcare (56% concerned), education, finance (loan approval, investment advice — 45% concerned), and employment/resume screening. These are areas where biased outputs can cause real harm.

30:42 Catharine Montgomery: Healthcare is especially concerning because historical biases in testing and treatment exist. For example, mammogram guidance changed without sufficient inclusion of women of color in testing, who may have different risk profiles. If AI inherits those biases, disparities will widen.

31:45 Catharine Montgomery: In education, biased grading or career guidance can limit opportunities. In finance, biased loan approvals and credit scoring harm marginalized groups. Automated resume screening can perpetuate hiring inequities.

32:51 Catharine Montgomery: Consumers want action more than lofty talk. They want more accurate results and improved communication about what companies are doing. If you’re not ready to claim success, be transparent about steps you’re taking.

33:56 Catharine Montgomery: Benefits of addressing bias: better outputs for all, transfer learning across demographics, multilingual reasoning improvements, better contextual understanding, and advances in image generation. I’d hoped image generation would improve faster — I paused using it but returned expecting major improvements and was disappointed.

35:02 Catharine Montgomery: Competitive benefits: first-mover advantage, and talent attraction — top developers want to work on fair generative AI. We use generative AI heavily at Better Together and tools like Suits.ai to integrate workflows and reduce bias in outputs.

36:14 Catharine Montgomery: Five recommended steps for companies integrating AI:
- Leadership commitment: Senior leaders must set vision and make bias mitigation a business priority.
- Align bias mitigation with brand values so it's part of everyday operations.
- Diversify data and teams: build diverse AI teams, include varied perspectives in training data, and create cross-functional bias review.
- Test, audit, repeat: establish ongoing evaluation, document incidents, and create audit processes.
- Monitor and disclose: build automated bias detection, maintain human oversight, and implement clear disclosure policies.

37:13 Catharine Montgomery: Test outputs, gather employee and client feedback, and maintain a bias checklist with monthly audits. Build continuous monitoring systems and document bias incidents to improve processes over time.

38:20 Catharine Montgomery: Engage users and continuously improve: create transparent feedback mechanisms, encourage reporting of bias without fear, publish reports on how you respond to bias, and use responsible GenAI as a differentiator. Start small and iterate.

39:19 Catharine Montgomery: I'd love to connect on LinkedIn. We'll share slides after the session.

40:20 David Berkowitz: Thanks, Catharine — that was a great playbook. When you share this with clients, do they nod along or do you see pushback?

41:25 Catharine Montgomery: It's mixed. Some clients don't want to be involved in AI; others want to use it but don't know how. The playbook helps walk organizations through steps. Adoption depends on timing and readiness — it's not if but when.

42:24 Bill Amstuts (audience): Full disclosure, I work for AllSides. There are organizations that rate media source bias like AllSides and Ad Fontes. Are you aware of models that use those ratings when training? Could such ratings be used as part of training data?

43:22 Catharine Montgomery: I’d love to build something like that. We have a free tool on AllScience.com where you can input a URL and it will give you a bias score. I'd welcome collaboration.

44:26 Sarah (audience): I attended a workshop on AI video tools. The two common tools for creating AI videos were SEDANCE (under ByteDance/TikTok) and Google V3. The presenter showed a demo to create a video for a comfortable office chair with an elaborate prompt that did not specify the actor's appearance. The SEDANCE output featured an Asian man; Google V3 returned a white man. The prompt only indicated the audience as business people — no demographic specs — yet the outputs reflected different biases from their origins.

45:29 Sarah (audience): The demo showed how these tools can be useful, for example creating multiple versions for different audiences and languages. But it also showed inherent bias: the Asian-based tool produced an Asian man, Google produced a white man. Marketing teams need to check such outputs and ensure diverse representation, like showing a conference room with varied people.

46:30 David Berkowitz: That’s a strong example. Ideally today these tools would support audience personalization: generate tailored videos for each audience. But the demo didn’t reflect that — it reflected the default biases of the models.

47:35 Catharine Montgomery: Exactly — ideally the AI would clarify prompts by asking follow-up questions like, “Do you want a particular demographic represented?” Many users don’t know to ask, especially younger people.

49:36 David Burkowitz: Building workflows that prompt clarification is important. Until major platforms change incentives, the onus is on us to create guardrails and checks.

50:36 Participant (audience): I’ve worked in tech in male-dominated environments. Many people aren’t aware of these bias issues. Technology should help reduce bias, not perpetuate it. This reinforces the need for diversity in technology hiring so tools consider varied perspectives.

52:46 David Berkowitz: All of that is true. The next step is audience personalization and tools that ask clarifying questions. Prompts matter, and people need training in prompt design and bias awareness.

53:45 Catharine Montgomery: That’s where checklists and workflows come in. If smaller companies adopt these practices, they can embed bias mitigation in their daily outputs. It’s an opportunity for early movers.

54:48 David Berkowitz: Great conversation. We’ll keep this going in Slack and with future speakers. Thanks, Catharine — please send slides and links and we’ll share them with the community. Appreciate everyone joining and the thoughtful questions. Have a great rest of the week.