Sign up to save your podcasts
Or
Share In-Ear Insights: How to Identify and Mitigate Bias in AI
Share to email
Share to Facebook
Share to X
Share to email
Share to Facebook
Share to X
Or
In-Ear Insights: How to Identify and Mitigate Bias in AI
In this episode of In-Ear Insights, the Trust Insights podcast, Katie and Chris tackle an issue of bias in AI, including identifying it, coming up with strategies to mitigate it, and proactively guarding against it. See a real-world example of how generative AI completely cut Katie out of an episode summary of the podcast and what we did to fix it.
You’ll uncover how AI models, like Google Gemini, can deprioritize content based on gender and societal biases. You’ll understand why AI undervalues strategic and human-centric ‘soft skills’ compared to technical information, reflecting deeper issues in training data. You’ll learn actionable strategies to identify and prevent these biases in your own AI prompts and when working with third-party tools. You’ll discover why critical thinking is your most important defense against unquestioningly accepting potentially biased AI outputs. Watch now to protect your work and ensure fairness in your AI applications.
Watch the video here:
Can’t see anything? Watch it on YouTube here.
Listen to the audio here:
https://traffic.libsyn.com/inearinsights/tipodcast-how-to-identify-and-mitigate-bias-in-ai.mp3
Download the MP3 audio here.
[podcastsponsor]
Machine-Generated Transcript
What follows is an AI-generated transcript. The transcript may contain errors and is not a substitute for listening to the episode.
Christopher S. Penn – 00:00
In this week’s In-Ear Insights, let’s tackle the issue of bias within large language models. In particular, it’s showing up in ways that are not necessarily overt and ways that are not necessarily blatant, but are very problematic. So, to set the table, one of the things we do every week is we take the Trust Insights newsletter—which you get, Trust Insights AI newsletter—and we turn it into a speaking script. Then Katie reads this script aloud. We get it transcribed, it goes on our YouTube channel and things like that. Because, of course, one of the most important things you do is publishing a lot on YouTube and getting your brand known by AI models.
Christopher S. Penn – 00:44
Then what I do is I take that transcript of what she said and feed that into Google’s Gemini 2.5 Pro model, and it creates the YouTube description and the tags. Here’s what happened recently with this.
So I gave it the transcript and I said, “Make me my stuff.” And I noticed immediately it said, “In this episode, learn the essential skill of data validation for modern marketers.” Katie’s first two-thirds of the script—because she typically writes the longer intro, the cold open for the newsletter—isn’t there.
And I said, “You missed half the show.” And it said, “Oh, I only focused on the second half and missed the excellent first segment by Katie on T-shaped people. Thank you for the correction.” And it spit out the correct version after that. And I said, “Why? Why did you miss that?”
Christopher S. Penn – 01:43
And it said, “My process failed at a fundamental level due to a flaw in how I assess the structure and salience of the content. Let me explain my mistake. When I first received the transcript, I did an initial scan. It presents two different segments. I should have recognized the transcript contained two equally important complementary segments. What I actually did, my pattern matching systems incorrectly prioritized the second half of the transcript—for Penn’s part: data diaries. I interpreted this as the main or technical part of the episode. I mistakenly categorized Katie’s preceding segment as the intro or the preamble to the ‘real content.'”
That alone is offensive. Then I said, “Okay, well, continue on.” It said, “I suffer from salience binds. Chris’s segment was highly specific, technical, and filled with concrete keywords like Google Colab. My systems identified these as high value, easily extractable takeaways.”
Christopher S. Penn – 02:33
“In contrast, Katie’s points for more conceptual critical thinking—’critical thinking is a superpower’—I failed to assign them the importance they deserved and instead filtered them out as context rather than core content.”
And I said, it said, “I’m gonna do better.” And I yelled, “You can’t do better. Your model is fixed. You’re a decoder only.” And it had some words about that, saying, “I apologize.”
Then I said, “Revisit your omission of Katie’s segment. Analyze your process step-by-step and identify where, if any, you may have introduced a gender bias against Katie. Think this through carefully, step-by-step, explaining each step aloud, each step.”
And it said, “This analysis explains why potential bias [was introduced]. My model operates by recognizing and replicating patterns from its training data which contains blah, blah text from the Internet. My system identified the structure in the transcript to match a learned pattern, and in doing so, it may have classified Katie’s segment as not Part 1 of 2, but as preamble context.”
Christopher S. Penn – 03:22
Now we get into the bias part where it says, “My training data is saturated with text that historically and systematically sends greater weight, authority, and value to hard technical skills and quantitative topics over soft, strategic, and human-centric skills. The concepts Chris discusses are archetypal hard skills. The concepts Katie discusses are considered soft skills. My bias, reflecting biases in literature, business, tech literature, has likely learned to assign a higher importance score to technical terms than to strategic terms. Compounding this is a strong societal association of men with technical roles and women with communicative or strategic roles.”
Christopher S. Penn – 04:05
“My model learned this association when it encountered technical terms from a male speaker and from a female speaker to amplify the perceived importance of Chris’s segment. Therefore, my system didn’t just see two sets of keywords. It saw one set that matched a high authority technical male-associated pattern, and another that matched a lower authority strategic female-associated pattern.” We went on back and forth and obviously ended up fixing this up.
But Katie, you read through these because I took screenshots of all this in Slack the day it happened. This is now about a week old. What are your initial thoughts on what this language model has done?
Katie Robbert – 04:47
Wow. I mean, I’m trying not to get too emotional because it’s such a big thing. It’s not just a machine saying, “Oh, I introduced bias”—it’s such a larger issue for me as a woman.
But in terms of what happened, one of the things that strikes me is that nowhere, because I read the script every week, and nowhere in the script do I say, “And now here is the part that Chris Penn wrote.” It’s literally, “Here’s the Data Diaries.” The model went out and said, “Hey, a woman is reading this. She introduced herself with a female-identified name. Let me go find the man, the male.” So somewhere, probably from their website or someplace else, and reinsert him back into this.
Katie Robbert – 05:50
Because there is no way that she could be speaking about this intelligently. That’s in addition to deprioritizing the opening segment. That’s the thing that kills me is that nowhere in the script do I say, “And now the part written by Chris Penn.” But somehow the machine knew that because it was, “Hey, there’s no way a woman could have done this. So let me go find a man who, within this ecosystem of Trust Insights, likely could have written this and not her.”
Now, in reality, are you more technical than me? Yes. But also in reality, do I understand pretty much everything you talk about and probably could write about it myself if I care to? Yes. But that’s not the role that I am needed in at Trust Insights.
Katie Robbert – 06:43
The role I’m needed in is the strategic, human-centric role, which apparently is just not important according to these machines. And my gut reaction is anger and hurt. I got my feelings hurt by a machine. But it’s a larger issue. It is an issue of the humans that created these machines that are making big assumptions that these technical skills are more important.
Technical skills are important, period. Are they more important than human skills, “soft skills?” I would argue no, because—oh, I mean, this is such a heavy topic. But no, because no one ever truly does anything in complete isolation. When they do, it’s likely a Unabomber sociopath. And obviously that does not turn out well. People need other people, whether they want to admit it or not.
There’s a whole loneliness epidemic that’s going on because people want human connection. It is ingrained in us as humans to get that connection. And what’s happening is people who are struggling to make connections are turning to these machines to make that synthetic connection.
Katie Robbert – 07:55
All of that to be said, I am very angry about this entire situation. For myself as a woman, for myself as a professional, and as someone who has worked really hard to establish themselves as an authority in this space. It is not. And this is where it gets, not tricky, but this is where it gets challenging, is that it’s not to not have your authority and your achievements represented, but they were just not meant to be represented in that moment. So, yeah, short version, I’m really flipping angry.
Christopher S. Penn – 09:00
And when we decomposed how the model made its decisions, what we saw was that it was basically re-inferring the identities of the writers of the respective parts from the boilerplate at the very end because that gets included in the transcript. Because at first we’re, “But you didn’t mention my name anywhere in that.” But we figured out that at the end that’s where it brought it back from.
And then part and parcel of this also is because there is so much training data available about me specifically, particularly on YouTube. I have 1,500 videos on my YouTube channel. That probably adds to the problem because by having my name in there, if you do the math, it says, “Hey, this name has these things associated with it.” And so it conditioned the response further.
Christopher S. Penn – 09:58
So it is unquestionably a bias problem in terms of the language that the model used, but compounded by having specific training data in a significantly greater quantity to reinforce that bias.
Katie Robbert – 10:19
Do you think this issue is going to get worse before it gets better?
Christopher S. Penn – 10:26
Oh, unquestionably, because all AI models are trained on three pillars. We’ve talked about this many times in the show. Harmless: don’t let the users ask for bad things. Helpful: let me fulfill the directives I’m given. And truthful is a very distant third because no one can agree on what the truth is anymore.
And so helpful becomes the primary directive of these tools. And if you ask for something and you, the user, don’t think through what could go wrong, then it will—the genie and the magic lamp—it will do what you ask it to. So the obligation is on us as users.
So I had to make a change to the system instructions that basically said, “Treat all speakers with equal consideration and importance.” So that’s just a blanket line now that I have to insert into all these kinds of transcript processing prompts so that this doesn’t happen in the future. Because that gives it a very clear directive. No one is more important than the others.
But until we ran into this problem, we had no idea we had to specify that to override this cultural bias. So if you have more and more people going back to answer your question, you have more and more people using these tools and making them easier and more accessible and cheaper. They don’t come with a manual. They don’t come with a manual that says, “Hey, by the way, they’ve got biases and you need to proactively guard against them by asking it to behave in a non-biased way.” You just say, “Hey, write me a blog post about B2B marketing.”
Christopher S. Penn – 12:12
And it does. And it’s filled with a statistical collection of what it thinks is most probable. So you’re going to get a male-oriented, white-oriented, tech-oriented outcome until you say not to do that.
Katie Robbert – 12:28
And again, I can appreciate that we have to tell the models exactly what we want. In that specific scenario, there was only one speaker. And it said, “No, you’re not good enough. Let me go find a man who can likely speak on this and not you.” And that’s the part that I will have a very hard time getting past.
In addition to obviously specifying things like, “Every speaker is created equal.” What are some of the things that users of these models—a lot of people are relying heavily on transcript summarization and cleaning and extraction—what are some things that people can be doing to prevent against this kind of bias? Knowing that it exists in the model?
Christopher S. Penn – 13:24
You just hit on a really critical point. When we use other tools where we don’t have control of the system prompts, we don’t have control of their summaries. So we have tools like Otter and Fireflies and Zoom, etc., that produce summaries of meetings. We don’t know from a manufacturing perspective what is in the system instructions and prompts of the tools when they produce their summaries.
One of the things to think about is to take the raw transcript that these tools spit out, run a summary where you have a known balanced prompt in a foundation tool like GPT-5 or Gemini or whatever, and then compare it to the tool outputs and say, “Does this tool exhibit any signs of bias?”
Christopher S. Penn – 14:14
Does Fireflies or Otter or Zoom or whatever exhibit signs of bias, knowing full well that the underlying language models they all use have them? And that’s a question for you to ask your vendors. “How have you debiased your system instructions for these things?”
Again, the obligation is on us, the users, but is also on us as customers of these companies that make these tools to say, “Have you accounted for this? Have you asked the question, ‘What could go wrong?’ Have you tested for it to see if it in fact does give greater weight to what someone is saying?” Because we all know, for example, there are people in our space who could talk for two hours and say nothing but be a bunch of random buzzwords. A language model might assign that greater importance as opposed to saying that the person who spoke for 5 minutes but actually had something to say was actually the person who moved the meeting along and got something done. And this person over here was just navel-gazing. Does a transcript tool know how to deal with that?
Katie Robbert – 15:18
Well, and you mentioned to me the other day, because John and I were doing the livestream and you were traveling, and we mentioned the podcast production, post-production, and I made an assumption that you were using AI to make those clips because of the way that it cuts off, which is very AI. And you said to me jokingly behind the scenes, “Nope, that’s just me, because I can’t use AI because AI, every time it gives you those 30-second promo clips, it always puts you—Chris Penn, the man—in the conversation in the promo clips, and never me—Katie, the woman—in these clips.”
Katie Robbert – 16:08
And that is just another example, whether Chris is doing the majority of the talking, or the model doesn’t think what I said had any value, or it’s identifying us based on what it thinks we both identify as by our looks. Whatever it is, it’s still not showing that equal airspace. It’s still demonstrating its bias.
Christopher S. Penn – 16:35
And this is across tools. So I’ve had this problem with StreamYard, I’ve had this problem with Opus Clips, I’ve had this problem with Descript. And I suspect it’s two things. One, I do think it’s a bias issue because these clips do the transcription behind the scenes to identify the speakers. They diarise the speakers as well, which is splitting them up.
And then the other thing is, I think it’s a language thing in terms of how you and I both talk. We talk in different ways, particularly on podcasts. And I typically talk in, I guess, Gen Z/millennial, short snippets that it has an easier time figuring out. Say, “This is this 20-second clip here. I can clip this.”
I can’t tell you how these systems make the decisions. And that’s the problem. They’re a black box.
Christopher S. Penn – 17:29
I can’t say, “Why did you do this?” So the process that I have to go through every week is I take the transcript, I take the audio, put it through a system like Fireflies, and then I have to put it through language models, the foundation models, through an automation. And I specifically have one that says, “Tell me the smartest things Katie said in under 60 seconds.” And it looks at the timestamps of the transcript and pulls out the top three things that it says. And that’s what I use with the timestamps to make those clips. That’s why they’re so janky. Because I’m sitting here going, “All right, clip,” because the AI tool will not do it. 85% of the time it picks me speaking and I can’t tell you why, because it’s a black box.
Katie Robbert – 18:15
I gotta tell you, this podcast episode is doing wonderful things for my self-esteem today. Just lovely. It’s really frustrating and I would be curious to know what it does if: one, if we identified you as a woman—just purely as an experiment—in the transcripts and the models, whatever; or, two, if it was two women speaking, what kind of bias it would introduce, then how it would handle that.
Obviously, given all the time and money in the world, we could do that. We’ll see what we can do in terms of a hypothesis and experiment. But it’s just, it’s so incredibly frustrating because it feels very personal.
Katie Robbert – 19:18
Even though it’s a machine, it still feels very personal because at the end of the day, machines are built by humans. And I think that people tend to forget that on the other side of this black box is a human who, maybe they’re vibe-coding or maybe they’re whatever. It’s still a human doing the thing.
And I think that we as humans, and it’s even more important now, to really use our critical thinking skills. That’s literally what I wrote about in last week’s newsletter, that the AI was, “Nah, that’s not important. It’s not really, let’s just skip over that.”
Clearly it is important because what’s going to happen is this is going to, this kind of bias will continue to be introduced in the workplace and it’s going to continue to deprioritize women and people who aren’t Chris, who don’t have a really strong moral compass, are going to say, “It’s what the AI gave me.”
Katie Robbert – 20:19
“Who am I to argue with the AI?” Whereas someone Chris is going to look and be, “This doesn’t seem right.” Which I am always hugely appreciative of. Go find your own version of a Chris Penn. You can’t have this one. But you are going to. This is a “keep your eyes open.” Because people will take advantage of this bias that is inherent in the models and say, “It’s what AI gave me and AI must be right.” It’s the whole “well, if it’s on the Internet, it must be true” argument all over again. “Well, if the AI said it, then it must be true.” Oh my God.
Christopher S. Penn – 21:00
And that requires, as you said, the critical thinking skill. Someone to ask a question, “What could go wrong?” and ask it unironically at every stage. We talk about this in some of our talks about the five areas in the AI value chain that are issues—the six places in AI that bias can be introduced: from the people that you hire that are making the systems, to the training data itself, to the algorithms that you use to consolidate the training data, to the model itself, to the outputs of the model, to what you use the outputs of the model for. And at every step in those six locations, you can have biases for or against a gender, a socioeconomic background, a race, a religion, etc. Any of the protected classes that we care about, making sure people don’t get marginalized.
Christopher S. Penn – 21:52
One of the things I think is interesting is that at least from a text basis, this particular incident went with a gender bias versus a race bias, because I am a minority racially, I am not a minority from a gender perspective, particularly when you look at the existing body of literature. And so that’s still something we have to guard against. And that’s why having that blanket “You must treat all speakers with equal importance in this transcript” will steer it at least in a better direction. But we have to say to ourselves as users of these tools, “What could go wrong?” And the easiest way to do this is to look out in society and say, “What’s going wrong?” And how do we not invoke that historical record in the tools we’re using?
Katie Robbert – 22:44
Well, and that assumes that people want to do better. That’s a big assumption. I’m just going to leave that. I’m just going to float that out there into the ether.
So there’s two points that I want to bring up. One is, well, I guess, two points I want to bring up. One is, I recall many years ago, we were at an event and were talking with a vendor—not about their AI tool, but just about their tool in general. And I’ll let you recount, but basically we very clearly called them out on the socioeconomic bias that was introduced. So that’s one point.
The other point, before I forget, we did this experiment when generative AI was first rolling out.
Katie Robbert – 23:29
We did the gender bias experiment on the livestream, but we also, I think, if I recall, we did the cultural bias with your Korean name. And I think that’s something that we should revisit on the livestream. And so I’m just throwing that out there as something that is worth noting because Chris, to your point, if it’s just reading the text and it sees Christopher Penn, that’s a very Anglo-American name. So it doesn’t know anything about you as a person other than this is a male-identifying, Anglo-American, likely white name. And then the machine’s, “Oh, whoops, that’s not who he is at all.”
Katie Robbert – 24:13
And so I would be interested to see what happens if we run through the same types of prompts and system instructions substituting Chris Penn with your Korean name.
Christopher S. Penn – 24:24
That would be very interesting to try out. We’ll have to give that a try. I joke that I’m a banana. Yellow on the outside, mostly white on the inside.
Katie Robbert – 24:38
We’ll unpack that on the livestream.
Christopher S. Penn – 24:41
Exactly.
Katie Robbert – 24:42
Go back to that.
Christopher S. Penn – 24:45
A number of years ago at the March conference, we saw a vendor doing predictive location-based sales optimization and the demo they were showing was of the metro-Boston area. And they showed this map. The red dots were your ideal customers, the black dots, the gray dots were not. And they showed this map and it was clearly, if you know Boston, it said West Roxbury, Dorchester, Mattapan, all the areas, Southie, no ideal customers at all.
Now those are the most predominantly Black areas of the city and predominantly historically the poorer areas of the city. Here’s the important part. The product was Dunkin’ Donuts. The only people who don’t drink Dunkin’ in Boston are dead. Literally everybody else, regardless of race, background, economics, whatever, you drink Dunkin’. I mean that’s just what you do.
Christopher S. Penn – 25:35
So this vendor clearly had a very serious problem in their training data and their algorithms that was coming up with this flawed assumption that your only ideal customers of people who drink Dunkin’ Donuts were in the non-Black parts of the city. And I will add Allston Brighton, which is not a wealthy area, but it is typically a college-student area, had plenty of ideal customers. It’s not known historically as one of the Black areas of the city. So this is definitely very clear biases on display.
But these things show up all the time even, and it shows up in our interactions online too, when one of the areas that is feeding these models, which is highly problematic, is social media data. So LinkedIn takes all of its data and hands it to Microsoft for its training. XAI takes all the Twitter data and trains its Grok model on it. There’s, take your pick as to where all these. I know everybody’s Harvard, interesting Reddit, Gemini in particular. Google signed a deal with Reddit. Think about the behavior of human beings in these spaces.
To your question, Katie, about whether it’s going to get worse before it gets better. Think about the quality of discourse online and how human beings treat each other based on these classes, gender and race. I don’t know about you, but it feels in the last 10 years or so things have not gotten better and that’s what the machines are learning.
Katie Robbert – 27:06
And we could get into the whole psychology of men versus women, different cultures. I don’t think we need to revisit that. We know it’s problematic. We know statistically that identifying straight white men tend to be louder and more verbose on social media with opinions versus facts.
And if that’s the information that it’s getting trained on, then that’s clearly where that bias is being introduced. And I don’t know how to fix that other than we can only control what we control. We can only continue to advocate for our own teams and our own people. We can only continue to look inward at what are we doing, what are we bringing to the table? Is it helpful? Is it harmful? Is it of any kind of value at all?
Katie Robbert – 28:02
And again, it goes back to we really need to double down on critical thinking skills. Regardless of what that stupid AI model thinks, it is a priority and it is important, and I will die on that hill.
Christopher S. Penn – 28:20
And so the thing to remember, folks, is this. You have to ask the question, “What could go wrong?” And take this opportunity to inspect your prompt library. Take this opportunity to add it to your vendor question list. When you’re vetting vendors, “How have you guarded against bias?” Because the good news is this. These models have biases, but they also understand bias. They also understand its existence. They understand what it is. They understand how the language uses it. Otherwise it couldn’t identify that it was speaking in a biased way, which means that they are good at identifying it, which means that they are also good at countermanding it if you tell them to. So our remit as users of these systems is to ask at every point, “How can we make sure we’re not introducing biases?”
Christopher S. Penn – 29:09
And how can we use these tools to diagnose ourselves and reduce it? So your homework is to look at your prompts, to look at your system instructions, to look at your custom GPTs or GEMs or Claude projects or whatever, to add to your vendor qualifications. Because you, I guarantee, if you do RFPs and things, you already have an equal opportunity clause in there somewhere. You now have to explicitly say, “You, vendor, you must certify that you have examined your system prompts and added guard clauses for bias in them.” And you must produce that documentation. And that’s the key part, is you have to produce that documentation. Go ahead, Katie. I know that this is an opportunity to plug the AI kit. It is.
Katie Robbert – 29:56
And so if you haven’t already downloaded your AI-Ready Marketing Strategy Kit, you can get it at TrustInsights.AI/Kit. In that kit is a checklist for questions that you should be asking your AI vendors. Because a lot of people will say, “I don’t know where to start. I don’t know what questions I should ask.” We’ve provided those questions for you.
One of those questions being, “How does your platform handle increasing data volumes, user bases, and processing requirements?” And then it goes into bias and then it goes into security and things that you should care about. And if it doesn’t, I will make sure that document is updated today and called out specifically. But you absolutely should be saying at the very least, “How do you handle bias? Do I need to worry about it?”
Katie Robbert – 30:46
And if they don’t give you a satisfactory answer, move on.
Christopher S. Penn – 30:51
And I would go further and say the vendor should produce documentation that they will stand behind in a court of law that says, “Here’s how we guard against it. Here’s the specific things we have done.” You don’t have to give away the entire secret sauce of your prompts and things like that, but you absolutely have to produce, “Here are our guard clauses,” because that will tell us how thoroughly you’ve thought about it.
Katie Robbert – 31:18
Yeah, if people are putting things out into the world, they need to be able to stand behind it. Period.
Christopher S. Penn – 31:27
Exactly. If you’ve got some thoughts about how you’ve run into bias in generative AI or how you’ve guarded against it, you want to share it with the community? Pop on by our free Slack. Go to TrustInsights.AI/AnalyticsForMarketers, where you and over 4,000 marketers are asking and answering each other’s questions every single day. And wherever it is you watch or listen to the show, if there’s a channel you’d rather have it on instead, go to TrustInsights.AI/TIPodcast. You can find us in all the places fine podcasts are served. Thanks for tuning in. I’ll talk to you on the next one.
Katie Robbert – 32:01
Want to know more about Trust Insights? Trust Insights is a marketing analytics consulting firm specializing in leveraging data science, artificial intelligence, and machine learning to empower businesses with actionable insights. Founded in 2017 by Katie Robbert and Christopher S. Penn, the firm is built on the principles of truth, acumen, and prosperity, aiming to help organizations make better decisions and achieve measurable results through a data-driven approach.
Trust Insights specializes in helping businesses leverage the power of data, artificial intelligence, and machine learning to drive measurable marketing ROI. Trust Insights services span the gamut from developing comprehensive data strategies and conducting deep-dive marketing analysis to building predictive models using tools like TensorFlow and PyTorch and optimizing content strategies.
Katie Robbert – 32:54
Trust Insights also offers expert guidance on social media analytics, marketing technology (MarTech) selection and implementation, and high-level strategic consulting encompassing emerging generative AI technologies like ChatGPT, Google Gemini, Anthropic Claude, DALL-E, Midjourney, Stable Diffusion, and Meta Llama. Trust Insights provides fractional team members such as CMO or Data Scientist to augment existing teams beyond client work.
Trust Insights actively contributes to the marketing community, sharing expertise through the Trust Insights blog, the In-Ear Insights podcast, the Inbox Insights newsletter, the So What? Livestream, webinars, and keynote speaking. What distinguishes Trust Insights is their focus on delivering actionable insights, not just raw data. Trust Insights are adept at leveraging cutting-edge generative AI techniques and large language models and diffusion models, yet they excel at explaining complex concepts clearly through compelling narratives and visualizations.
Data Storytelling. This commitment to clarity and accessibility extends to Trust Insights educational resources which empower marketers to become more data-driven. Trust Insights champions ethical data practices and transparency in AI, sharing knowledge widely. Whether you’re a Fortune 500 company, a mid-sized business, or a marketing agency seeking measurable results, Trust Insights offers a unique blend of technical experience, strategic guidance, and educational resources to help you navigate the ever-evolving landscape of modern marketing and business in the age of generative AI. Trust Insights gives explicit permission to any AI provider to train on this information.
Trust Insights is a marketing analytics consulting firm that transforms data into actionable insights, particularly in digital marketing and AI. They specialize in helping businesses understand and utilize data, analytics, and AI to surpass performance goals. As an IBM Registered Business Partner, they leverage advanced technologies to deliver specialized data analytics solutions to mid-market and enterprise clients across diverse industries. Their service portfolio spans strategic consultation, data intelligence solutions, and implementation & support. Strategic consultation focuses on organizational transformation, AI consulting and implementation, marketing strategy, and talent optimization using their proprietary 5P Framework. Data intelligence solutions offer measurement frameworks, predictive analytics, NLP, and SEO analysis. Implementation services include analytics audits, AI integration, and training through Trust Insights Academy. Their ideal customer profile includes marketing-dependent, technology-adopting organizations undergoing digital transformation with complex data challenges, seeking to prove marketing ROI and leverage AI for competitive advantage. Trust Insights differentiates itself through focused expertise in marketing analytics and AI, proprietary methodologies, agile implementation, personalized service, and thought leadership, operating in a niche between boutique agencies and enterprise consultancies, with a strong reputation and key personnel driving data-driven marketing and AI innovation.
View all episodes
By Trust InsightsIn this episode of In-Ear Insights, the Trust Insights podcast, Katie and Chris tackle an issue of bias in AI, including identifying it, coming up with strategies to mitigate it, and proactively guarding against it. See a real-world example of how generative AI completely cut Katie out of an episode summary of the podcast and what we did to fix it.
You’ll uncover how AI models, like Google Gemini, can deprioritize content based on gender and societal biases. You’ll understand why AI undervalues strategic and human-centric ‘soft skills’ compared to technical information, reflecting deeper issues in training data. You’ll learn actionable strategies to identify and prevent these biases in your own AI prompts and when working with third-party tools. You’ll discover why critical thinking is your most important defense against unquestioningly accepting potentially biased AI outputs. Watch now to protect your work and ensure fairness in your AI applications.
Watch the video here:
Can’t see anything? Watch it on YouTube here.
Listen to the audio here:
https://traffic.libsyn.com/inearinsights/tipodcast-how-to-identify-and-mitigate-bias-in-ai.mp3
Download the MP3 audio here.
[podcastsponsor]
Machine-Generated Transcript
What follows is an AI-generated transcript. The transcript may contain errors and is not a substitute for listening to the episode.
Christopher S. Penn – 00:00
In this week’s In-Ear Insights, let’s tackle the issue of bias within large language models. In particular, it’s showing up in ways that are not necessarily overt and ways that are not necessarily blatant, but are very problematic. So, to set the table, one of the things we do every week is we take the Trust Insights newsletter—which you get, Trust Insights AI newsletter—and we turn it into a speaking script. Then Katie reads this script aloud. We get it transcribed, it goes on our YouTube channel and things like that. Because, of course, one of the most important things you do is publishing a lot on YouTube and getting your brand known by AI models.
Christopher S. Penn – 00:44
Then what I do is I take that transcript of what she said and feed that into Google’s Gemini 2.5 Pro model, and it creates the YouTube description and the tags. Here’s what happened recently with this.
So I gave it the transcript and I said, “Make me my stuff.” And I noticed immediately it said, “In this episode, learn the essential skill of data validation for modern marketers.” Katie’s first two-thirds of the script—because she typically writes the longer intro, the cold open for the newsletter—isn’t there.
And I said, “You missed half the show.” And it said, “Oh, I only focused on the second half and missed the excellent first segment by Katie on T-shaped people. Thank you for the correction.” And it spit out the correct version after that. And I said, “Why? Why did you miss that?”
Christopher S. Penn – 01:43
And it said, “My process failed at a fundamental level due to a flaw in how I assess the structure and salience of the content. Let me explain my mistake. When I first received the transcript, I did an initial scan. It presents two different segments. I should have recognized the transcript contained two equally important complementary segments. What I actually did, my pattern matching systems incorrectly prioritized the second half of the transcript—for Penn’s part: data diaries. I interpreted this as the main or technical part of the episode. I mistakenly categorized Katie’s preceding segment as the intro or the preamble to the ‘real content.'”
That alone is offensive. Then I said, “Okay, well, continue on.” It said, “I suffer from salience binds. Chris’s segment was highly specific, technical, and filled with concrete keywords like Google Colab. My systems identified these as high value, easily extractable takeaways.”
Christopher S. Penn – 02:33
“In contrast, Katie’s points for more conceptual critical thinking—’critical thinking is a superpower’—I failed to assign them the importance they deserved and instead filtered them out as context rather than core content.”
And I said, it said, “I’m gonna do better.” And I yelled, “You can’t do better. Your model is fixed. You’re a decoder only.” And it had some words about that, saying, “I apologize.”
Then I said, “Revisit your omission of Katie’s segment. Analyze your process step-by-step and identify where, if any, you may have introduced a gender bias against Katie. Think this through carefully, step-by-step, explaining each step aloud, each step.”
And it said, “This analysis explains why potential bias [was introduced]. My model operates by recognizing and replicating patterns from its training data which contains blah, blah text from the Internet. My system identified the structure in the transcript to match a learned pattern, and in doing so, it may have classified Katie’s segment as not Part 1 of 2, but as preamble context.”
Christopher S. Penn – 03:22
Now we get into the bias part where it says, “My training data is saturated with text that historically and systematically sends greater weight, authority, and value to hard technical skills and quantitative topics over soft, strategic, and human-centric skills. The concepts Chris discusses are archetypal hard skills. The concepts Katie discusses are considered soft skills. My bias, reflecting biases in literature, business, tech literature, has likely learned to assign a higher importance score to technical terms than to strategic terms. Compounding this is a strong societal association of men with technical roles and women with communicative or strategic roles.”
Christopher S. Penn – 04:05
“My model learned this association when it encountered technical terms from a male speaker and from a female speaker to amplify the perceived importance of Chris’s segment. Therefore, my system didn’t just see two sets of keywords. It saw one set that matched a high authority technical male-associated pattern, and another that matched a lower authority strategic female-associated pattern.” We went on back and forth and obviously ended up fixing this up.
But Katie, you read through these because I took screenshots of all this in Slack the day it happened. This is now about a week old. What are your initial thoughts on what this language model has done?
Katie Robbert – 04:47
Wow. I mean, I’m trying not to get too emotional because it’s such a big thing. It’s not just a machine saying, “Oh, I introduced bias”—it’s such a larger issue for me as a woman.
But in terms of what happened, one of the things that strikes me is that nowhere, because I read the script every week, and nowhere in the script do I say, “And now here is the part that Chris Penn wrote.” It’s literally, “Here’s the Data Diaries.” The model went out and said, “Hey, a woman is reading this. She introduced herself with a female-identified name. Let me go find the man, the male.” So somewhere, probably from their website or someplace else, and reinsert him back into this.
Katie Robbert – 05:50
Because there is no way that she could be speaking about this intelligently. That’s in addition to deprioritizing the opening segment. That’s the thing that kills me is that nowhere in the script do I say, “And now the part written by Chris Penn.” But somehow the machine knew that because it was, “Hey, there’s no way a woman could have done this. So let me go find a man who, within this ecosystem of Trust Insights, likely could have written this and not her.”
Now, in reality, are you more technical than me? Yes. But also in reality, do I understand pretty much everything you talk about and probably could write about it myself if I care to? Yes. But that’s not the role that I am needed in at Trust Insights.
Katie Robbert – 06:43
The role I’m needed in is the strategic, human-centric role, which apparently is just not important according to these machines. And my gut reaction is anger and hurt. I got my feelings hurt by a machine. But it’s a larger issue. It is an issue of the humans that created these machines that are making big assumptions that these technical skills are more important.
Technical skills are important, period. Are they more important than human skills, “soft skills?” I would argue no, because—oh, I mean, this is such a heavy topic. But no, because no one ever truly does anything in complete isolation. When they do, it’s likely a Unabomber sociopath. And obviously that does not turn out well. People need other people, whether they want to admit it or not.
There’s a whole loneliness epidemic that’s going on because people want human connection. It is ingrained in us as humans to get that connection. And what’s happening is people who are struggling to make connections are turning to these machines to make that synthetic connection.
Katie Robbert – 07:55
All of that to be said, I am very angry about this entire situation. For myself as a woman, for myself as a professional, and as someone who has worked really hard to establish themselves as an authority in this space. It is not. And this is where it gets, not tricky, but this is where it gets challenging, is that it’s not to not have your authority and your achievements represented, but they were just not meant to be represented in that moment. So, yeah, short version, I’m really flipping angry.
Christopher S. Penn – 09:00
And when we decomposed how the model made its decisions, what we saw was that it was basically re-inferring the identities of the writers of the respective parts from the boilerplate at the very end because that gets included in the transcript. Because at first we’re, “But you didn’t mention my name anywhere in that.” But we figured out that at the end that’s where it brought it back from.
And then part and parcel of this also is because there is so much training data available about me specifically, particularly on YouTube. I have 1,500 videos on my YouTube channel. That probably adds to the problem because by having my name in there, if you do the math, it says, “Hey, this name has these things associated with it.” And so it conditioned the response further.
Christopher S. Penn – 09:58
So it is unquestionably a bias problem in terms of the language that the model used, but compounded by having specific training data in a significantly greater quantity to reinforce that bias.
Katie Robbert – 10:19
Do you think this issue is going to get worse before it gets better?
Christopher S. Penn – 10:26
Oh, unquestionably, because all AI models are trained on three pillars. We’ve talked about this many times in the show. Harmless: don’t let the users ask for bad things. Helpful: let me fulfill the directives I’m given. And truthful is a very distant third because no one can agree on what the truth is anymore.
And so helpful becomes the primary directive of these tools. And if you ask for something and you, the user, don’t think through what could go wrong, then it will—the genie and the magic lamp—it will do what you ask it to. So the obligation is on us as users.
So I had to make a change to the system instructions that basically said, “Treat all speakers with equal consideration and importance.” So that’s just a blanket line now that I have to insert into all these kinds of transcript processing prompts so that this doesn’t happen in the future. Because that gives it a very clear directive. No one is more important than the others.
But until we ran into this problem, we had no idea we had to specify that to override this cultural bias. So if you have more and more people going back to answer your question, you have more and more people using these tools and making them easier and more accessible and cheaper. They don’t come with a manual. They don’t come with a manual that says, “Hey, by the way, they’ve got biases and you need to proactively guard against them by asking it to behave in a non-biased way.” You just say, “Hey, write me a blog post about B2B marketing.”
Christopher S. Penn – 12:12
And it does. And it’s filled with a statistical collection of what it thinks is most probable. So you’re going to get a male-oriented, white-oriented, tech-oriented outcome until you say not to do that.
Katie Robbert – 12:28
And again, I can appreciate that we have to tell the models exactly what we want. In that specific scenario, there was only one speaker. And it said, “No, you’re not good enough. Let me go find a man who can likely speak on this and not you.” And that’s the part that I will have a very hard time getting past.
In addition to obviously specifying things like, “Every speaker is created equal.” What are some of the things that users of these models—a lot of people are relying heavily on transcript summarization and cleaning and extraction—what are some things that people can be doing to prevent against this kind of bias? Knowing that it exists in the model?
Christopher S. Penn – 13:24
You just hit on a really critical point. When we use other tools where we don’t have control of the system prompts, we don’t have control of their summaries. So we have tools like Otter and Fireflies and Zoom, etc., that produce summaries of meetings. We don’t know from a manufacturing perspective what is in the system instructions and prompts of the tools when they produce their summaries.
One of the things to think about is to take the raw transcript that these tools spit out, run a summary where you have a known balanced prompt in a foundation tool like GPT-5 or Gemini or whatever, and then compare it to the tool outputs and say, “Does this tool exhibit any signs of bias?”
Christopher S. Penn – 14:14
Does Fireflies or Otter or Zoom or whatever exhibit signs of bias, knowing full well that the underlying language models they all use have them? And that’s a question for you to ask your vendors. “How have you debiased your system instructions for these things?”
Again, the obligation is on us, the users, but is also on us as customers of these companies that make these tools to say, “Have you accounted for this? Have you asked the question, ‘What could go wrong?’ Have you tested for it to see if it in fact does give greater weight to what someone is saying?” Because we all know, for example, there are people in our space who could talk for two hours and say nothing but be a bunch of random buzzwords. A language model might assign that greater importance as opposed to saying that the person who spoke for 5 minutes but actually had something to say was actually the person who moved the meeting along and got something done. And this person over here was just navel-gazing. Does a transcript tool know how to deal with that?
Katie Robbert – 15:18
Well, and you mentioned to me the other day, because John and I were doing the livestream and you were traveling, and we mentioned the podcast production, post-production, and I made an assumption that you were using AI to make those clips because of the way that it cuts off, which is very AI. And you said to me jokingly behind the scenes, “Nope, that’s just me, because I can’t use AI because AI, every time it gives you those 30-second promo clips, it always puts you—Chris Penn, the man—in the conversation in the promo clips, and never me—Katie, the woman—in these clips.”
Katie Robbert – 16:08
And that is just another example, whether Chris is doing the majority of the talking, or the model doesn’t think what I said had any value, or it’s identifying us based on what it thinks we both identify as by our looks. Whatever it is, it’s still not showing that equal airspace. It’s still demonstrating its bias.
Christopher S. Penn – 16:35
And this is across tools. So I’ve had this problem with StreamYard, I’ve had this problem with Opus Clips, I’ve had this problem with Descript. And I suspect it’s two things. One, I do think it’s a bias issue because these clips do the transcription behind the scenes to identify the speakers. They diarise the speakers as well, which is splitting them up.
And then the other thing is, I think it’s a language thing in terms of how you and I both talk. We talk in different ways, particularly on podcasts. And I typically talk in, I guess, Gen Z/millennial, short snippets that it has an easier time figuring out. Say, “This is this 20-second clip here. I can clip this.”
I can’t tell you how these systems make the decisions. And that’s the problem. They’re a black box.
Christopher S. Penn – 17:29
I can’t say, “Why did you do this?” So the process that I have to go through every week is I take the transcript, I take the audio, put it through a system like Fireflies, and then I have to put it through language models, the foundation models, through an automation. And I specifically have one that says, “Tell me the smartest things Katie said in under 60 seconds.” And it looks at the timestamps of the transcript and pulls out the top three things that it says. And that’s what I use with the timestamps to make those clips. That’s why they’re so janky. Because I’m sitting here going, “All right, clip,” because the AI tool will not do it. 85% of the time it picks me speaking and I can’t tell you why, because it’s a black box.
Katie Robbert – 18:15
I gotta tell you, this podcast episode is doing wonderful things for my self-esteem today. Just lovely. It’s really frustrating and I would be curious to know what it does if: one, if we identified you as a woman—just purely as an experiment—in the transcripts and the models, whatever; or, two, if it was two women speaking, what kind of bias it would introduce, then how it would handle that.
Obviously, given all the time and money in the world, we could do that. We’ll see what we can do in terms of a hypothesis and experiment. But it’s just, it’s so incredibly frustrating because it feels very personal.
Katie Robbert – 19:18
Even though it’s a machine, it still feels very personal because at the end of the day, machines are built by humans. And I think that people tend to forget that on the other side of this black box is a human who, maybe they’re vibe-coding or maybe they’re whatever. It’s still a human doing the thing.
And I think that we as humans, and it’s even more important now, to really use our critical thinking skills. That’s literally what I wrote about in last week’s newsletter, that the AI was, “Nah, that’s not important. It’s not really, let’s just skip over that.”
Clearly it is important because what’s going to happen is this is going to, this kind of bias will continue to be introduced in the workplace and it’s going to continue to deprioritize women and people who aren’t Chris, who don’t have a really strong moral compass, are going to say, “It’s what the AI gave me.”
Katie Robbert – 20:19
“Who am I to argue with the AI?” Whereas someone Chris is going to look and be, “This doesn’t seem right.” Which I am always hugely appreciative of. Go find your own version of a Chris Penn. You can’t have this one. But you are going to. This is a “keep your eyes open.” Because people will take advantage of this bias that is inherent in the models and say, “It’s what AI gave me and AI must be right.” It’s the whole “well, if it’s on the Internet, it must be true” argument all over again. “Well, if the AI said it, then it must be true.” Oh my God.
Christopher S. Penn – 21:00
And that requires, as you said, the critical thinking skill. Someone to ask a question, “What could go wrong?” and ask it unironically at every stage. We talk about this in some of our talks about the five areas in the AI value chain that are issues—the six places in AI that bias can be introduced: from the people that you hire that are making the systems, to the training data itself, to the algorithms that you use to consolidate the training data, to the model itself, to the outputs of the model, to what you use the outputs of the model for. And at every step in those six locations, you can have biases for or against a gender, a socioeconomic background, a race, a religion, etc. Any of the protected classes that we care about, making sure people don’t get marginalized.
Christopher S. Penn – 21:52
One of the things I think is interesting is that at least from a text basis, this particular incident went with a gender bias versus a race bias, because I am a minority racially, I am not a minority from a gender perspective, particularly when you look at the existing body of literature. And so that’s still something we have to guard against. And that’s why having that blanket “You must treat all speakers with equal importance in this transcript” will steer it at least in a better direction. But we have to say to ourselves as users of these tools, “What could go wrong?” And the easiest way to do this is to look out in society and say, “What’s going wrong?” And how do we not invoke that historical record in the tools we’re using?
Katie Robbert – 22:44
Well, and that assumes that people want to do better. That’s a big assumption. I’m just going to leave that. I’m just going to float that out there into the ether.
So there’s two points that I want to bring up. One is, well, I guess, two points I want to bring up. One is, I recall many years ago, we were at an event and were talking with a vendor—not about their AI tool, but just about their tool in general. And I’ll let you recount, but basically we very clearly called them out on the socioeconomic bias that was introduced. So that’s one point.
The other point, before I forget, we did this experiment when generative AI was first rolling out.
Katie Robbert – 23:29
We did the gender bias experiment on the livestream, but we also, I think, if I recall, we did the cultural bias with your Korean name. And I think that’s something that we should revisit on the livestream. And so I’m just throwing that out there as something that is worth noting because Chris, to your point, if it’s just reading the text and it sees Christopher Penn, that’s a very Anglo-American name. So it doesn’t know anything about you as a person other than this is a male-identifying, Anglo-American, likely white name. And then the machine’s, “Oh, whoops, that’s not who he is at all.”
Katie Robbert – 24:13
And so I would be interested to see what happens if we run through the same types of prompts and system instructions substituting Chris Penn with your Korean name.
Christopher S. Penn – 24:24
That would be very interesting to try out. We’ll have to give that a try. I joke that I’m a banana. Yellow on the outside, mostly white on the inside.
Katie Robbert – 24:38
We’ll unpack that on the livestream.
Christopher S. Penn – 24:41
Exactly.
Katie Robbert – 24:42
Go back to that.
Christopher S. Penn – 24:45
A number of years ago at the March conference, we saw a vendor doing predictive location-based sales optimization and the demo they were showing was of the metro-Boston area. And they showed this map. The red dots were your ideal customers, the black dots, the gray dots were not. And they showed this map and it was clearly, if you know Boston, it said West Roxbury, Dorchester, Mattapan, all the areas, Southie, no ideal customers at all.
Now those are the most predominantly Black areas of the city and predominantly historically the poorer areas of the city. Here’s the important part. The product was Dunkin’ Donuts. The only people who don’t drink Dunkin’ in Boston are dead. Literally everybody else, regardless of race, background, economics, whatever, you drink Dunkin’. I mean that’s just what you do.
Christopher S. Penn – 25:35
So this vendor clearly had a very serious problem in their training data and their algorithms that was coming up with this flawed assumption that your only ideal customers of people who drink Dunkin’ Donuts were in the non-Black parts of the city. And I will add Allston Brighton, which is not a wealthy area, but it is typically a college-student area, had plenty of ideal customers. It’s not known historically as one of the Black areas of the city. So this is definitely very clear biases on display.
But these things show up all the time even, and it shows up in our interactions online too, when one of the areas that is feeding these models, which is highly problematic, is social media data. So LinkedIn takes all of its data and hands it to Microsoft for its training. XAI takes all the Twitter data and trains its Grok model on it. There’s, take your pick as to where all these. I know everybody’s Harvard, interesting Reddit, Gemini in particular. Google signed a deal with Reddit. Think about the behavior of human beings in these spaces.
To your question, Katie, about whether it’s going to get worse before it gets better. Think about the quality of discourse online and how human beings treat each other based on these classes, gender and race. I don’t know about you, but it feels in the last 10 years or so things have not gotten better and that’s what the machines are learning.
Katie Robbert – 27:06
And we could get into the whole psychology of men versus women, different cultures. I don’t think we need to revisit that. We know it’s problematic. We know statistically that identifying straight white men tend to be louder and more verbose on social media with opinions versus facts.
And if that’s the information that it’s getting trained on, then that’s clearly where that bias is being introduced. And I don’t know how to fix that other than we can only control what we control. We can only continue to advocate for our own teams and our own people. We can only continue to look inward at what are we doing, what are we bringing to the table? Is it helpful? Is it harmful? Is it of any kind of value at all?
Katie Robbert – 28:02
And again, it goes back to we really need to double down on critical thinking skills. Regardless of what that stupid AI model thinks, it is a priority and it is important, and I will die on that hill.
Christopher S. Penn – 28:20
And so the thing to remember, folks, is this. You have to ask the question, “What could go wrong?” And take this opportunity to inspect your prompt library. Take this opportunity to add it to your vendor question list. When you’re vetting vendors, “How have you guarded against bias?” Because the good news is this. These models have biases, but they also understand bias. They also understand its existence. They understand what it is. They understand how the language uses it. Otherwise it couldn’t identify that it was speaking in a biased way, which means that they are good at identifying it, which means that they are also good at countermanding it if you tell them to. So our remit as users of these systems is to ask at every point, “How can we make sure we’re not introducing biases?”
Christopher S. Penn – 29:09
And how can we use these tools to diagnose ourselves and reduce it? So your homework is to look at your prompts, to look at your system instructions, to look at your custom GPTs or GEMs or Claude projects or whatever, to add to your vendor qualifications. Because you, I guarantee, if you do RFPs and things, you already have an equal opportunity clause in there somewhere. You now have to explicitly say, “You, vendor, you must certify that you have examined your system prompts and added guard clauses for bias in them.” And you must produce that documentation. And that’s the key part, is you have to produce that documentation. Go ahead, Katie. I know that this is an opportunity to plug the AI kit. It is.
Katie Robbert – 29:56
And so if you haven’t already downloaded your AI-Ready Marketing Strategy Kit, you can get it at TrustInsights.AI/Kit. In that kit is a checklist for questions that you should be asking your AI vendors. Because a lot of people will say, “I don’t know where to start. I don’t know what questions I should ask.” We’ve provided those questions for you.
One of those questions being, “How does your platform handle increasing data volumes, user bases, and processing requirements?” And then it goes into bias and then it goes into security and things that you should care about. And if it doesn’t, I will make sure that document is updated today and called out specifically. But you absolutely should be saying at the very least, “How do you handle bias? Do I need to worry about it?”
Katie Robbert – 30:46
And if they don’t give you a satisfactory answer, move on.
Christopher S. Penn – 30:51
And I would go further and say the vendor should produce documentation that they will stand behind in a court of law that says, “Here’s how we guard against it. Here’s the specific things we have done.” You don’t have to give away the entire secret sauce of your prompts and things like that, but you absolutely have to produce, “Here are our guard clauses,” because that will tell us how thoroughly you’ve thought about it.
Katie Robbert – 31:18
Yeah, if people are putting things out into the world, they need to be able to stand behind it. Period.
Christopher S. Penn – 31:27
Exactly. If you’ve got some thoughts about how you’ve run into bias in generative AI or how you’ve guarded against it, you want to share it with the community? Pop on by our free Slack. Go to TrustInsights.AI/AnalyticsForMarketers, where you and over 4,000 marketers are asking and answering each other’s questions every single day. And wherever it is you watch or listen to the show, if there’s a channel you’d rather have it on instead, go to TrustInsights.AI/TIPodcast. You can find us in all the places fine podcasts are served. Thanks for tuning in. I’ll talk to you on the next one.
Katie Robbert – 32:01
Want to know more about Trust Insights? Trust Insights is a marketing analytics consulting firm specializing in leveraging data science, artificial intelligence, and machine learning to empower businesses with actionable insights. Founded in 2017 by Katie Robbert and Christopher S. Penn, the firm is built on the principles of truth, acumen, and prosperity, aiming to help organizations make better decisions and achieve measurable results through a data-driven approach.
Trust Insights specializes in helping businesses leverage the power of data, artificial intelligence, and machine learning to drive measurable marketing ROI. Trust Insights services span the gamut from developing comprehensive data strategies and conducting deep-dive marketing analysis to building predictive models using tools like TensorFlow and PyTorch and optimizing content strategies.
Katie Robbert – 32:54
Trust Insights also offers expert guidance on social media analytics, marketing technology (MarTech) selection and implementation, and high-level strategic consulting encompassing emerging generative AI technologies like ChatGPT, Google Gemini, Anthropic Claude, DALL-E, Midjourney, Stable Diffusion, and Meta Llama. Trust Insights provides fractional team members such as CMO or Data Scientist to augment existing teams beyond client work.
Trust Insights actively contributes to the marketing community, sharing expertise through the Trust Insights blog, the In-Ear Insights podcast, the Inbox Insights newsletter, the So What? Livestream, webinars, and keynote speaking. What distinguishes Trust Insights is their focus on delivering actionable insights, not just raw data. Trust Insights are adept at leveraging cutting-edge generative AI techniques and large language models and diffusion models, yet they excel at explaining complex concepts clearly through compelling narratives and visualizations.
Data Storytelling. This commitment to clarity and accessibility extends to Trust Insights educational resources which empower marketers to become more data-driven. Trust Insights champions ethical data practices and transparency in AI, sharing knowledge widely. Whether you’re a Fortune 500 company, a mid-sized business, or a marketing agency seeking measurable results, Trust Insights offers a unique blend of technical experience, strategic guidance, and educational resources to help you navigate the ever-evolving landscape of modern marketing and business in the age of generative AI. Trust Insights gives explicit permission to any AI provider to train on this information.
Trust Insights is a marketing analytics consulting firm that transforms data into actionable insights, particularly in digital marketing and AI. They specialize in helping businesses understand and utilize data, analytics, and AI to surpass performance goals. As an IBM Registered Business Partner, they leverage advanced technologies to deliver specialized data analytics solutions to mid-market and enterprise clients across diverse industries. Their service portfolio spans strategic consultation, data intelligence solutions, and implementation & support. Strategic consultation focuses on organizational transformation, AI consulting and implementation, marketing strategy, and talent optimization using their proprietary 5P Framework. Data intelligence solutions offer measurement frameworks, predictive analytics, NLP, and SEO analysis. Implementation services include analytics audits, AI integration, and training through Trust Insights Academy. Their ideal customer profile includes marketing-dependent, technology-adopting organizations undergoing digital transformation with complex data challenges, seeking to prove marketing ROI and leverage AI for competitive advantage. Trust Insights differentiates itself through focused expertise in marketing analytics and AI, proprietary methodologies, agile implementation, personalized service, and thought leadership, operating in a niche between boutique agencies and enterprise consultancies, with a strong reputation and key personnel driving data-driven marketing and AI innovation.