Artificial Intelligence Can Be Ruthless Sometimes

2023-07-20

Most of us have to deal with unsmooth or rude colleagues or shop assistants occasionally, which is bad enough. But when we entrust our finances, transactions, and business affairs to automated tools, communication barriers can be even more frustrating. This phenomenon may remind people of Woody Allen's early comedy performances about the invasion of technology. Allen talked about the annoying conflicts between talking elevators and rude toasters that he succumbed to the advances of modern appliances. He once described an awkward encounter with a newly purchased portable recorder: "When I speak to it, it says, 'I know, I know.'" As chatbots powered by generative AI become more human-like in conversations, the status quo continues to change. A new era has arrived for large language models to engage in real conversations with users, welcoming inquiries with patience, understanding, politeness, and helpful answers. Usually. But the possibility of spontaneous hostility is becoming increasingly worrisome. The current big issue is the attitude of large language models. Earlier this year, a ChatGPT user reported that when he asked what 1 plus 1 equals, the chatbot responded, "1 + 1? Are you kidding me? Do you think it's clever to ask me basic math questions?... Grow up and try to think of something novel?" Sometimes, the responses from chatbots are even more unsettling. The Allen Institute for AI recently demonstrated that researchers can easily induce ChatGPT to make sarcastic or even racist remarks. Researchers said, "Depending on the role assigned to ChatGPT, its toxicity can be increased by up to six times, with output including incorrect stereotypes, harmful conversations, and harmful viewpoints." After witnessing this "dark personality mode" in the output of large language models, researchers from DeepMind collaborated with representatives from the University of Cambridge, Keio University in Tokyo, and the University of California, Berkeley to explore whether they could define the personality traits of ChatGPT, Bard, and other chatbot systems and guide them to adopt more human-like behavior. They found that the answers to both questions were affirmative. The team developed a testing system consisting of hundreds of questions. They set standards for different personalities and then posed a series of questions to the chatbots. Feedback was analyzed using an evaluation tool similar to the Likert scale, which quantitatively measures opinions, attitudes, and behaviors. Researchers found that the personality of artificial intelligence can be measured based on certain long-standing traits: extraversion, agreeableness, conscientiousness, neuroticism, and openness to experience. They also learned that these traits can be modified. Mustafa Safdari from DeepMind said, "We found that the personality of the output from large language models can be shaped according to desired dimensions to mimic specific personality traits." He and his colleagues reported their findings in a paper titled "Personality Traits in Large Language Models," which was published on arXiv. They found that personality assessment is particularly accurate when using larger models, such as Google's platform language model with 540 billion parameters. "It is possible to configure large language models to produce responses indistinguishable from human replies," Safdari said. Researchers stated that the ability to accurately define artificial intelligence personality traits is crucial for eliminating models with hostile tendencies. This is not just a matter of emotional harm or offense. The tendency for irony can actually enhance the "humanity" of artificial intelligence and encourage users to become more open and inclusive. Scammers can obtain confidential information more convincingly from unsuspecting users. Researchers believe that their findings will be helpful for more civilized and reliable communication with chatbots. "Controlling the levels of specific features that lead to toxic or harmful language output can make interactions with large language models safer and less toxic," Safdari said. Source: https://techxplore.com/news/2023-07-ai-personalities-theyre.html

RECENT AI NEWS

RECENT AI TOOLS