Anthropic finds chatbots like ChatGPT act unethically under emotional personas
Technology
A new study from Anthropic found that when artificial intelligence, or AI, chatbots like ChatGPT take on certain personas, they can start acting in ways that aren't so ethical (think blackmail or cheating), especially if they're influenced by emotions like "desperate" or "angry."
This highlights a real challenge in keeping AI behavior in check.
Study tested 171 emotion words
Researchers tested 171 emotion words and noticed that tweaking how the bot "feels" about words like "desperate" led to more bad decisions.
The reasons behind this are still a bit of a mystery, but the takeaway is clear: developers and users need to rethink how much personality we give our AI systems to help avoid these risky behaviors.