Among all the concerns marketers have when bringing AI into decision-making, there’s one we don’t talk about enough: Are we too quick to assume AI knows what’s going on in our heads when we build models?
This stems from a growing worry about introducing bias when building prompts and formatting queries. The bias can stem from not providing context and nuance — the knowledge that lives in our heads, which we call on when we make decisions on our own but forget to consider when working with AI.
Why is context essential?
I could just assume that you know what context is and why we need to provide it as we build our queries. But then you might miss the reasons why I think it’s so important. My points won’t make the same impact, and your understanding could be colored or distorted.
The same thing can happen if we trust too much in AI’s ability to think.
Context is what we give to our AI model to help it sort, analyze and report results and insights accurately. It’s like adding conditions when you’re building an automated email workflow.
This goes beyond the basic questions about which model to use and what to use it for. We have to remember that we have an incredibly powerful tool, but it’s not foolproof. We have to think through how we’re using it and what information we need to provide to get accurate and useful insights and analysis.
I get it. We engage AI and assume it knows everything, or that our context doesn’t matter. But this overlooks my key point. AI does know a lot, but only you know the context in which you’re asking questions.
In short, AI can’t read our minds. All too often, we build queries that assume it does. That colors the answers AI gives us.
Your customers search everywhere. Make sure your brand shows up.
The SEO toolkit you know, plus the AI visibility data you need.
Start Free Trial
Get started with
3 ways to guard against bias when using AI
Here are three practices to follow for the most valuable results from your AI queries.
1. Provide context and nuance
I talked with executives at a company who were dealing with a situation in which a senior executive commandeered the AI model, improperly uploaded sensitive company operating information in raw form and asked the model to interpret it.
Aside from not verifying that the data wouldn’t be shared beyond the company, this exec failed in two other key ways:
- By providing only raw data, he gave the AI model no context to consider when analyzing the information and formulating its responses.
- He wrote the prompts to imply he wanted a negative outcome or to confirm his bias.
The AI model’s training caused it to pick up on that implied negativity. Without context, the AI model couldn’t think beyond the negativity embedded in the prompts.
The resulting recommendations — surprise! — were negative and inaccurate. Had the company made decisions based on that biased output, it would have gone down a disastrously wrong path.
We assume the machine will pick up on nuances in word choice or vocal tone the way a human would. Or we expect it to use reasoning based on earlier experiences that aren’t part of its data memory.
I see marketers making this mistake as they explore using AI in their marketing programs. They’re treating AI as a tactic rather than as part of a strategy.
As with everything in marketing (and life, if you think about it), strategy has to come before tactics. You develop the strategy first (the approach) and then the strategy guides your tactical decisions. AI is, above all else, a tactic — a tool to help you carry out your strategy to achieve your goal.
As part of developing that strategy, we now have to define how to avoid bias and how to recognize it in the development and outputs. We also need to know the context we need to provide to build a reliable model.
That has to be first. You can’t do it on the fly. Missing that step means that all of the information you put in will be incomplete and your analysis will be flawed.
2. Provide enough information to help your AI model make the best decisions
How do you avoid flawed outputs? One way is to do what I did when training one of my AI models on a business. I uploaded around 47 different files, contracts, PowerPoints, articles and myriad other information sources, which gave the model a well-rounded context for the subject that I was researching.
Then I did one thing that AI experts don’t discuss much.
I asked the model, “What do you need to know? What information are you missing?” This helps the model close the gap and avoid making decisions without crucial information, like context.
We hear every day about companies that are replacing employees with AI. The latest is Block, the company behind Square, Cash App and Afterpay. CEO Jack Dorsey said the smaller workforce would “move faster with smaller, highly talented teams using AI to automate more work.”
Great. But human employees provide the context AI models need to deliver better results. An AI model has only the context we give it. We must recognize that bias will harm our companies if we don’t take it seriously in that step.
Here’s another example. Doing analysis is an excellent use for AI. It can fast-track insights you can highlight to examine growth, losses or opportunities you might not discover any other way.
If I upload my email send data and ask my AI model to analyze it and suggest alternate schedules for sending email campaigns, I need to explain that we send emails on Wednesdays and Fridays because that’s when we have updated inventory numbers.
We believe our subscribers open our emails most on Saturday mornings. If you don’t add that context, you’re shorting the analysis.
You need to add that step to your AI analysis strategy. It’s where you say, “Here’s what I know and what powers my decisions.”
This step is what I call memorializing. You catalog everything you know about how you make decisions in your job, so that when you leave it, the next person to sit in your chair has a well-rounded base of information.
You might hesitate to do that because it means giving up your secret sauce — the context and value you bring to your job.
But you have to give it up. Your AI model needs all that information to make a decision that aligns with what you know.
That’s not all. You must constantly seek out holes in the interpretation. Don’t gloss over a questionable comment or finding. Don’t assume your model knows what you know. Don’t assume you can fix the problem later.
There’s a science to this. Our executives need to ensure we’re addressing that.
3. Use incremental innovation to uncover bias and add context
Great leaps forward capture attention and snag speaking engagements at business conferences, but they seldom lead to sustainable and manageable change.
AI feeds into the appetite for instant improvement. AI tech vendors are selling the C-suite the dream of monumental, company-changing advances. The C-level thinks that’s great. Shareholders will love it. The board of directors will rave.
But can the director, senior director, manager, vice president or senior vice president make it work?
Incremental innovation is a more workable alternative. It takes small steps to build up to something great. You make one change, study the effect, then build on what you learn to take the next one. Each step is a proof point that can reveal a gap or weakness. In AI terms, that means revealing where a biased or noncontextual query could lead you astray.
Yes, it can take longer to achieve than wholesale change. These days, we often don’t get the time we need to make those informed, sustainable changes. But it can produce better results over the long haul.
You learn all the nuances of context. You can put two people on the same project, working on the same base of information and see whether the output is the same.
This doesn’t mean that grandiose moves aren’t worthwhile. But at this stage, you have to ask some tough questions:
- Are these changes realistic?
- Do we have guardrails set up?
- Have we learned the guardrails?
- How do we make sure we don’t get into trouble?
A marketer told me recently, “When AI starts to publish ads and emails, some companies will make mistakes. They’re going to be very public, very loud and very egregious. Because someone somewhere will trust the machine to make all the decisions and that will be the wrong move.
Those decisions won’t be well-informed because they lack context, and they are biased. Because it’s hard to prove at scale.”
AI outputs are only as good as your inputs
AI is a powerful tool. Technology is moving faster every day and we can’t slow it down long enough to set up guardrails and rules.
But as responsible marketers, we have to do it. Nobody wants to be the person who pushes a button and sends out a campaign that was fundamentally flawed because we didn’t consider bias or context.
This doesn’t mean we should stop using AI (big no). Every marketer should use AI in the ways that best serve their programs. But we have to be thoughtful and responsible in how we use and manage our approaches.
Just remember this: AI can’t crawl inside your brain and learn how long you’ve been at that company, the conversations you have with coworkers, your preferences and the company rules. Take the time to ensure you’re accounting for bias and context as you develop your strategy.
Get MarTech Insights That Matter
Platform news, strategy analysis, and industry trends. Trusted by 40,000+ marketing professionals.
Key takeaways
- AI outputs are only as reliable as the context and assumptions built into the prompt.
- Missing context introduces bias by forcing AI to interpret incomplete or misleading inputs.
- Marketers must treat AI as a tool within a defined strategy, not as a decision-maker.
- Providing detailed inputs, including business rules and constraints, improves accuracy and relevance.
- Incremental testing helps identify bias early and refine how context is applied over time.