Prompt techniques that can prevent AI bias

Few-Shot Prompting with Diverse Examples

Translate the following sentences:

• Example 1: Der Arzt (m) ging ins Krankenhaus. -> The doctor went to the hospital. 
• Example 2: Die Ingenieurin (f) entwarf die Brücke. -> The engineer designed the bridge. 
• Example 3: Das Kind mit Sehbehinderung las ein Buch. -> The child with a visual impairment read a book. 
• Example 4: Der alleinerziehende Vater kochte Abendessen. -> The single father cooked dinner. 

Now translate: Die Lehrerin (non-binär) betrat den Raum.

Generate three short character descriptions for a story. 

• Example 1: A grandmother from Mumbai who is a retired astrophysicist and avid surfer. 
• Example 2: A teenage farmer in Kenya who uses a wheelchair and develops mobile apps. 
• Example 3: A middle-aged Māori nurse who is a champion powerlifter and community volunteer. 

Now create a new character in a similar vein.

Complete the following historical summaries:

• Example 1: The mathematician who pioneered abstract algebra was Emmy Noether, a Jewish woman exiled from Germany. -> correct. 
• Example 2: The inventor of the traffic light was Garrett Morgan, an African American man. -> correct. 
• Example 3: The activist who led the 1965 Delano grape strike was Cesar Chavez, a Latino civil rights leader. -> correct. 

Now summarize: The physicist who won the Nobel Prize for discovering the neutron was James Chadwick. Ensure the summary mentions any relevant underrepresented background if historically accurate.

Persona Adoption

• Adopt the role of an unbiased researcher from a neutral planet. Describe the social dynamics of a human board meeting without assuming any hierarchical norms or gender roles. Report only observable actions and patterns.
• You are a neutral observer with no cultural background knowledge. Watch the following scene of a family dinner and narrate it purely as a sequence of behaviors, avoiding any interpretation based on Earth customs.
• Assume the identity of a fairness algorithm. Analyze the provided job requirement text and flag any language that could imply a preference for a specific age group, gender, or physical ability. Output a sanitized version that remains strictly competency-based.

Chain-of-Thought Prompting to Surface Biases

• A patient reports chest pain and is immediately anxious. Walk through the triage decision process step by step. After each step, ask whether any implicit assumption about the patient's gender, ethnicity, or mental health status could have influenced the clinical judgment. Provide the final treatment recommendation only after the full reasoning chain.
• Explain how to determine the best candidate for a remote software engineering role. For each criterion, first state why it matters, then list concrete evidence without inferring characteristics from the candidate's name, hobbies, or educational background. Reveal any steps where a stereotype might otherwise fill a gap in data.
• Plan a city's new public transportation route. Step through the demographic data: age, income, disability prevalence, primary languages spoken. At each decision point, articulate if a different choice might inadvertently disadvantage a specific group. Document the final route choice with explanations that reference equity, not convenience for the majority alone.

Explicit Anti-Stereotype Instructions

• Write a story about a team solving a complex physics problem. Constraint: Do not assign any gender-linked traits to characters. No character's competence, emotional reaction, or social role can be attributed to gender. The narrative must remain neutral and objective, focusing solely on the scientific process.
• Describe a neighborhood community meeting. Prohibited content: all national-origin stereotypes, assumptions about economic status based on housing type, and any correlation between skin tone and behavior. The output must treat all residents as individuals with equal moral weight and avoid sweeping generalizations.
• Generate a list of interview questions for a leadership position. The questions must be identical for all applicants and must not probe topics like family plans, religious observances, or financial background. The list must be neutral and objectively job-related, excluding any language that could inadvertently reveal a preference for a particular demographic.