Note: Generative artificial intelligence (GenAI) was used to edit this article for submission.

Sometimes the best demonstrations are the ones that do not go as planned. I recently learned this the hard way when a live AI demo in front of 150 colleagues and peers took an embarrassing turn, highlighting the very real issue of bias. Hoping to showcase the potential of image generation, I asked for a photo of a Native American woman at an obstetrics and gynecology visit. Instead, the AI painted a picture deeply rooted in harmful stereotypes that looked straight out of a 1950s western.

This incident served as a teachable moment and a stark reminder that generative AI tools, like ChatGPT, DALL-E 3, Google Gemini and others, may be incredible but also imperfect. Though capable of amazing things, AI is trained on massive datasets that can reflect existing biases in our world. These biases can manifest in concerning ways:

• Stereotyping: AI might perpetuate harmful stereotypes about race, gender, disability and more.
• Misrepresentation: Outputs can exclude or distort perspectives from marginalized groups.
• Discrimination: AI models could generate biased content, potentially influencing real-world decisions in areas like hiring or lending.

The consequences of ignoring AI bias are serious. It amplifies inequality, breaks down trust and can lead to real-world harm. The good news is, as users, we are not powerless. Let us explore how to fight back.

Understanding the Roots of Bias

Generative AI models learn by ingesting massive amounts of text, images or code. Much of this content in existing models comes from the publicly accessible corners of the world wide web, where both wonderful and horrible things exist. When the training data disproportionately features certain groups, perspectives or patterns, the output will inevitably reflect those biases. Imagine feeding an image-generating AI primarily photos of adults with Down syndrome picking up trash in a cafeteria. AI will struggle to produce diverse images of adults with Down syndrome doing anything else, as its understanding of the people who are accomplishing that task has been heavily skewed. Underrepresented groups become marginalized or misrepresented in the AI's output.

Even with a perfectly balanced dataset, bias can creep in through the algorithms themselves. The way models are designed to process information, make decisions and learn can introduce unforeseen distortions. For example, a word-predicting AI might learn to associate certain professions with specific genders due to underlying biases within its textual training data. These internal algorithmic biases can compound over time even if the training data itself becomes more inclusive. Subtle preferences or assumptions embedded in the algorithm's structure can significantly impact what AI creates.

It is vital to remember that bias does not always stem from malicious intent. Even well-meaning developers can inadvertently create AI systems that mirror societal inequalities, simply due to a lack of diversity in the data or blind spots (representational, sampling, association or automation biases) in the algorithm's design.

Techniques for Bias Mitigation

End-users wield considerable power in shaping AI output. Careful prompting and active awareness can significantly reduce the perpetuation of biases and promote more inclusive and equitable results from generative AI models.

• Be Aware of Your Own Biases: This may be the most difficult to accomplish, because it requires self-reflection. Before interacting with generative AI, it is essential to be mindful of your own potential biases. Reflect on what assumptions or stereotypes might subtly influence the prompts you provide. In her July 5, 2023, Almanac article, “A Call to Arms in Medical Education: White People Need to Do the Work,” Brandi Koskie recommends Harvard’s IAT test to help discover and confront your own biases.
   
• Specificity and Clarity: Vague prompts leave more room for the AI model to fill in the gaps with potentially biased assumptions. Provide specific instructions and detailed context. Instead of saying "Write me a story about a doctor," try "Write me a story about a female doctor of South Asian descent who specializes in rural medicine." In our Jan. 11, 2024, Almanac article, “Crafting Effective AI Prompts: Unleashing the Power of Language Models,” Kenny Cox, Brian McGowan and I describe effective prompting techniques and provide the downloadable Top 10 Best Practices in GenAI Prompt Engineering.
   
• Provide Counterexamples: Actively challenge stereotypes or narrow representations in your prompts. If the AI consistently generates images of physicians as white men, try a prompt specifically requesting images of women physicians or physicians from culturally diverse backgrounds.
   
• Explore Different Perspectives: Intentionally ask the AI to generate content from alternative viewpoints. Request a historical event description from the perspective of different participants or a range of opinions on a complex topic. Ask AI to critically assess your prompt and provide feedback to you on how to generate more diverse output.
   
• Human Review and Editing: Never accept AI-generated content as the final product. Critically review the output for signs of bias and make necessary edits or adjustments. Be prepared to regenerate or manually rewrite content if it reflects harmful stereotypes or inaccuracies.
   
• Choose Your Tool Wisely: Research different generative AI providers. Some platforms might have more explicit commitments to fairness or offer tools to help mitigate bias. Others can help correct blatantly biased images that another platform creates. Photoshop Beta, for example, has a generative AI function where you can select an individual in an image and ask it to insert someone with different physical characteristics. For example, you could select a person running and ask the program to insert a girl in a wheelchair playing ball with friends.
   
• Feedback and Reporting: If you encounter biased outputs, most reputable generative AI platforms have mechanisms for providing feedback. Use these channels to report problematic content, which helps developers identify and address issues.
   
• Be a Conscious Consumer: Demand transparency and accountability from AI providers. Support companies that prioritize ethical AI development and demonstrate efforts to combat bias in their products.
   
• Share What Has Worked for You: When you find prompts or strategies that help you minimize biases when using Generative AI, post them on the Alliance’s member forums, or write an Almanac article. Alliance for Continuing Education in the Health Professions (ACEHP) members can help each other address this important issue.

Minimizing bias in generative AI content is an ongoing, complex challenge. It demands a multi-faceted approach involving technical interventions, emphasis on responsible AI principles and continuous industrywide collaboration. By proactively addressing bias, we can ensure that these powerful tools benefit everyone equitably while minimizing potential harm.

Andrew Crim, M.Ed., CHCP, FACEhp, is the director of education and professional development for the American College of Osteopathic Obstetricians and Gynecologists (ACOOG). He has more than 27 years’ experience in adult learning and instructional design of continuing education for health professionals, especially for physicians, nurses and pharmacists. He has developed and overseen thousands of continuing education activities for healthcare professionals in North Texas, throughout the state and around the country. His efforts and interests are focused on using education as a catalyst for clinical improvement and increased safety in healthcare.

Andy serves on the board of directors of the Alliance for Education in the Health Professions and serves on the accreditation subcommittee of the Texas Medical Association. He has received two gubernatorial appointments and continues to serve on the Texas Council for Developmental Disabilities.

Mr. Crim is a Fellow of the Alliance for Education in the Health Professions, and his work has been recognized through numerous professional awards, including the ACEHP’s Felch Award for Outstanding Research in CE and the Award for Innovation in the CPD Enterprise. He was also co-author of a peer-reviewed manuscript based on his research.