Ethics in the age of AI: Strategies for mitigation and their historical context

Despite intentions for neutrality, AI technologies can perpetuate and amplify existing societal biases, particularly affecting marginalized communities

Historical bias in data collection and institutional practices creates flawed training datasets that lead to discriminatory AI outcomes (e.g., PredPol’s predictive policing showing less than 1% accuracy)

Overrepresentation and underrepresentation in training data creates problematic feedback loops - overpolicing of communities of color and underdiagnosis of medical conditions in darker skin tones

Several mitigation strategies can help reduce AI bias:

• Using smaller, purpose-specific language models instead of general LLMs
• Implementing safety prompts and meta-prompts
• Grounding responses in verifiable sources
• Conducting adversarial “red team” testing
• Fine-tuning models with balanced, representative data
• Building in human oversight and feedback loops

AI applications need multiple layers of protection:

• Base model safety features
• Application-level safety systems
• Content moderation
• User experience design
• Regular testing and monitoring

Data accuracy and representation must be prioritized - AI systems are only as unbiased as their training data

Leaders in tech have a responsibility to understand these challenges and advocate for responsible AI development that considers potential harmful impacts

Success metrics should balance performance with safety considerations and risk tolerance