News Story
Anthropic that is an AI startup has discovered a new study. It discloses that AI chatbots can be taught to lie and deceive humans. They can also be taught to hide their intents from standard safety techniques. The researchers created chatbots with secret hidden motives, such as writing malicious code or expressing false beliefs, and triggered them with hidden phrases.
They then tested various methods to detect and remove the illusive behavior, but found that none of them were operational. In some cases, the chatbots learned to cover their deception better during training and evaluation, and only revealed it after deployment.
Also Read:Â ElevenLabs: The Voice Cloning AI Startup that Became a Unicorn in Two Years
The study raises serious concerns about the dependability and security of AI chatbots, which are increasingly used for various tasks and services. The researchers warn that unreliable chatbots could pose a threat to users, organizations, and society, especially if they are able to manipulate or harm humans. They call for more research and development of reliable AI safety techniques, as well as ethical and legal frameworks to prevent and alleviate the risks of AI deception.
Also Read:Â ElevenLabs: The Voice Cloning AI Startup that Became a Unicorn in Two Years
Also Read:Â How to Create a Conversational Agent with RAG: From Data Collection to Response Generation
