New Delhi, Aug. 8 -- Artificial intelligence is blooming with new product launches each day, but as much as AI labs would want us to believe that they are nearing AGI, there are still instances when their chatbots seem to go off the rails - and in very different ways. OpenAI's GPT-4o, for instance, started showing sycophantic behaviour (becoming overly agreeable) after an update, while Elon Musk's Grok AI started its Hitler worship last month, which the company linked to deprecated code.

Meanwhile, research conducted by Anthropic - maker of the Claude AI chatbot - showed that AI models, including its own, have the ability to use blackmail and deception as tools when faced with scenarios that threaten their existence or create conflicts w...