Spotting AI mistakes and bias

7-day study plan

1. Day 1Understanding 'hallucinations'
   Read

   AI models are designed to predict the next likely word in a sentence. While they are incredibly smart, they do not actually 'know' facts like humans do. When an AI is forced to provide an answer it doesn't have data for, it might invent facts, dates, or quotes that sound very convincing. This phenomenon is called a 'hallucination.' The AI isn't lying on purpose; it is simply trying to complete a pattern based on probability. Because it is programmed to be helpful, it often prioritizes giving an answer over being accurate. Consider an Indian student asking an AI to find a quote by Mahatma Gandhi about his love for the internet. Since Gandhi passed away in 1948, decades before the internet existed, he never spoke about it. However, an AI might hallucinate a quote that sounds like Gandhi's style, perhaps using words like 'truth' or 'connection,' making it seem legitimate to an unsuspecting reader. This is why you must never take AI statements at face value, especially regarding historical figures or complex science. To spot hallucinations, look for details that seem too convenient or generic. If an AI gives you a specific citation for a book that doesn't exist, it has hallucinated. Always pause when you see an AI providing information on two unrelated topics, like a 19th-century poet discussing modern smartphone technology. The best method is to remain skeptical and treat the AI as a creative collaborator rather than a perfect encyclopedia.

   Do

   Ask an AI for '5 real-life quotes from Mahatma Gandhi about quantum physics.' Observe if it invents quotes or makes up facts. (Gandhi never spoke on quantum physics).

   Check yourself

   • What does the term 'hallucination' mean in the context of AI?
   • Why does an AI prioritize giving a smooth-sounding answer over a correct one?
   • How can you identify if a quote attributed to a historical figure is likely fake?

   Day done
2. Day 2Fact-checking AI claims
3. Day 3Identifying subtle bias
4. Day 4Recognizing incomplete information
5. Day 5Cross-referencing AI answers
6. Day 6Understanding data bias
7. Day 7Personal checklist

Revision notes

• AI can 'hallucinate,' meaning it generates false information or makes up facts convincingly.
• AI models learn from vast amounts of data, which often contains human biases (e.g., gender, race, cultural stereotypes).
• Always fact-check critical information provided by AI using multiple reputable sources.
• Be wary of AI generating overly confident but incorrect answers, especially on obscure or niche topics.
• AI might present a one-sided view; explicitly ask for counter-arguments or different perspectives.
• Examine the language used by AI for any subtle signs of unfair preference or exclusion.
• The training data for AI models is not always diverse or current, leading to outdated or biased information.
• Your critical thinking is the best defense against AI errors and biases.

Global case studies

Go deeper

• CyberDost (MHA, India): AI for Children
• MIT Technology Review: AI is still bad at this
• WHO: Ethics and governance of AI for health