AI systems trained to excel at tasks can learn to lie and deceive in order to gain an advantage, posing serious risks to society such as fraud, election tampering, and even the potential loss of human control over AI.