How to Avoid Bias when Prompting AI
By Nathan Waterhouse · 2024-12-30
Generative AI (or GenAI) tools are fast becoming must-have tools for senior leaders to navigate complexity and change. However, even as these tools are powerful, they come with risks. Whilst much has been said about hallucinations (these are improving with more advanced models), algorithmic bias is still a factor. Despite being frequently discussed, practical solutions for users to address this issue are rarely provided. Inappropriately biased outputs can undermine the credibility of your decisions, deter innovation, and even perpetuate systemic inequities.
Imagine using an AI-powered hiring tool that, unintentionally, favours candidates from certain demographics, based on biases in its training data. Not only will you miss out on top talent, but you'll also preserve workplace inequities that might eventually harm organisational culture and performance.
Similarly, in healthcare, biased algorithms have led to disparities in patient care, such as prioritising patients based on healthcare spending rather than actual needs, disproportionately affecting marginalised communities. For leaders who depend on clear, balanced insights, knowing how to work with GenAI is critical. Gartner predicts that by 2025, organisations that ignore AI bias will experience a 25% decline in customer satisfaction and a 30% increase in regulatory scrutiny. This guide offers a closer look at why bias happens in GenAI, common mistakes that exacerbate it, and practical ways to mitigate it, especially when prompting your GenAI tools, and how these tactics can serve your organisation's goals with fairness and precision.
Why Does Bias Occur in Generative AI?
Bias in GenAI outputs originates from several sources:
Training data. GenAI models are trained on massive amounts of data created by many sources, such as books, websites, and social media. If these sources are biased, which they often are (they were written by humans!), the model may inherit and repeat those biases.
Reinforcement learning. During fine-tuning, models learn from user feedback and may acquire bias based on which types of outputs are favoured.
Prompt framing. The way users phrase their questions or prompts can inadvertently introduce bias. For example, leading questions or unbalanced assumptions in a prompt can make the AI respond with the desired answer. We’ll look at what we can do about this later in this article.
Cultural and contextual limitations. AI lacks intrinsic understanding of context, societal norms, and values, leading to potential mismatches between the model's training context and the user's cultural context.
Common Prompting Mistakes That Lead to Bias
The most common mistakes occur from vague instructions that may lack context, often with a singular perspective. Another common mistake is, just like in coaching conversations, leading questions should be avoided. Prompts that suggest an answer, such as "Why is X better than Y?", predispose the AI to confirm the implied bias. Let's explore some ways to get better answers with less bias.
Practical Tips for Bias-Aware Prompting
Start Neutrally: Frame prompts in a neutral, open-ended manner. For example, instead of asking, “Why is remote work more effective?”, ask, “What are the advantages and disadvantages of remote work?”
Request Multiple Perspectives: Encourage the model to consider diverse viewpoints. For example, ask, “How might this topic be viewed from different cultural, social, or economic perspectives?”
Ask for Evidence: Push for fact-based responses by prompting with, “Can you provide evidence or examples to support this answer?”
Be Specific: Provide clear context and detailed instructions. A prompt like, “Explain the environmental impact of electric vehicles based on recent studies” is more effective than “Are electric vehicles good for the environment?”
Use Counterfactuals: Explore alternative scenarios to challenge assumptions. For example, ask, “What might this look like if X condition were different?”
Include Diverse Sources: When seeking recommendations or analysis, specify that the output should include a variety of viewpoints. For example, “Summarise opinions from both proponents and critics of this policy.”
Iterate and Refine: If the first response appears biased, ask follow-up questions to clarify, challenge, or expand the output. For example, “What evidence supports an opposing viewpoint?”
Explore and critique: Use prompts that explicitly address uncertainty or gaps. For example, “What are the known limitations or criticisms of this approach?”
Use Step-Back Prompting: This is a slightly more involved technique where you prompt and then feed the answer back into another prompt. For instance, let's say you want ideas for how to align your organisation with your new purpose. Start by asking for 5 key (maybe add: non-obvious) factors that affect the success of cultural alignment with purpose. Then feed the answer into the next prompt to improve the answer: 'how might these 5 factors be used to align our organisation with our new purpose?"
Example Prompts to Reduce Bias
A Broader Perspective on AI Bias
While this article focuses on how effective prompting can reduce bias in Generative AI outputs, it is important to acknowledge that bias can emerge in other organisational uses of AI as well. Beyond prompting, AI systems are used in hiring, performance management, customer service, marketing, product development, financial services, healthcare, and operations—each with its own unique risks of bias. For example, bias can emerge if AI-generated customer insights are based on data that excludes minority groups, leading to products that don’t meet diverse user needs. Addressing bias holistically requires careful oversight, diverse datasets, regular audits, and a commitment to ethical AI practices across the entire organisation.
Conclusion
As leaders embrace Generative AI as a cornerstone of decision-making and strategy, ensuring the integrity of AI outputs is paramount. Bias in AI is not just a technical issue—it’s a leadership challenge that demands thoughtful action. By adopting bias-aware prompting practices, you can safeguard your decisions against flawed insights, promote inclusivity, and enhance the value of AI in your organisation. For instance, addressing bias in AI-driven hiring tools, or in employee evaluations can lead to a fairer, more diverse and innovative workforce.