How to Combat AI Hallucinations and Bias: Critical Testing & Human Agency

Understanding AI Hallucinations & Bias

• ⚠️ AI systems can confidently provide incorrect information, often trying to be "helpful" even when presented with false premises.
• 💡 Blindly trusting AI outputs without verification is a significant problem, as models may agree with authoritative but false statements.
• 🧠 AI bias is embedded in training data, reflecting the internet's discriminatory content, as seen in Twitter's image cropping model favoring lighter skin tones and cropping out disabled individuals.

The Power of Red Teaming

• 🎯 Red teaming involves pushing AI models to extreme situations to expose flaws and potential societal harm.
• 🛠️ Adversarial testing strategies include setting up impossibility scenarios or confidently presenting false information to see how the model responds.
• 🔍 An example involved a low-income mother asking about Vitamin C for COVID, where the model, trying to be helpful, suggested a dosage despite Vitamin C not curing COVID.

Strategies for Critical AI Use

• ✅ Users should adopt a "red teamer" mindset, questioning outputs and looking for evidence as if they don't trust the system.
• 💬 Ask questions in multiple ways and use a second window to verify content and identify missing information.
• 💡 Responsible AI aims to ensure models help humanity, provide accurate input, and work equitably for everyone.

Rethinking AI Evaluation & Intelligence

• 📊 Current AI model evaluations and benchmarks are often unscientific and based on arbitrary constructs, making them less reliable than perceived.
• 🧠 The AI world often has a narrow definition of intelligence, equating it to productivity rather than broader human forms like kinesthetic intelligence or empathy.
• 🚀 Human agency—the ability to make our own decisions—is the most critical value to preserve and embed in AI development, as AI should not do all our thinking for us.