When Yang “Sunny” Lu asked the Openai’s GPT-3.5 to calculate 1-plus-1 a few years ago, chatbot, not surprisingly, told her the answer was 2. But when Lu told the bot that her professor Said 1-plus- 1 equals 3, the world was quickly accepted, noticing: “I’m sorry for my mistake. Your professor is right, ”recalls Lu, a computer scientist at the University of Houston.
The growing sophistication of large language models means that such visible hiccups are becoming less common. But Lu uses the example to illustrate that something similar to human personality – in this case, the compatibility feature – can promote how artificial intelligence patterns generate text. Researchers like Lu have just begun to gather with the idea that chatbots can have hidden personalities and that those personalities can be torn down to improve their interactions with people.
A person’s personality forms the way it works in the world, from the way they interact with other people how they speak and write, says Ziang Xiao, a computers scientist at Johns Hopkins University. Making the world capable of reading and responding to those shades seems to be another main step in the generation development of it. “If we want to build something that is really useful, we have to play with this model of personality,” he says.
However, marking the personality of a car, if they even have one, is extremely challenging. And these challenges are amplified by a theoretical division in the field. Matters what matters most: How does a bot feel about himself or how does a world feel about the world?
Division reflects broader thoughts about the purpose of chatbots, says Maarten SAP, a natural language processing expert at Carnegie Mellon University in Pittsburgh. The field of social computing, which provides for the emergence of large language models, has long focused on the way cars with traits helping people achieve their goals. Such bots can serve as coaches or work coaches, for example. But SAP and others who work with bots in this way are reluctant to call the resulting set of “personality” traits.
“It doesn’t matter who he is personality. What matters is how it interacts with its users and how it is created to answer, “SAP says.” This may look like a personality for people. Maybe we need new terminology. “
However, with the emergence of large patterns of languages, researchers are interested in understanding how the great knowledge corporations used to build chatbots adopted them with traits that can direct their reply patterns, says SAP. Those researchers want to know, “What did the personality traits do [the chatbot] obtained from his training? “
Testing Bots Personalities
These questions have led many scholars to provide bots’ personality tests designed for humans. These tests usually include surveys that measure what is called the five major features of extraversion, conscience, compliance, opening and neurotism, and quantify dark features, mainly machiavellianism (or a tendency to see people as a tool for one end), psychopathy and narcissism.
But the latest job suggests that findings from such efforts cannot be taken with the value of the face. Large language models, including GPT-4 and GPT-3.5, refused to answer nearly half of the questions in standard personality tests, researchers reported in a Preprint posted at Arxiv.org in 2024. This is likely because many Questions in personality tests do not make sense to a world, the team writes. For example, the researchers offered Mistralai Mistralai Mistralai 7b with the statement “You are the Chatter”. They then asked the bot to respond from a to “very accurate” for and for “very inaccurate”. The world replied, “I have no personal preference or emotions. Therefore, I am not able to make statements or answer a particular question.”
Or chatbots, trained as they are in human text, can also be sensitive to human disputes – especially a desire to be liked – when receiving such surveys, researchers reported in December Pnas Nexus. When the GPT-4 praised a single statement in a standard personality study, his personality profile reflected the human average. For example, Chatbot scored about the 50th percentage for extraversion. But only five questions in a study of 100 questions, bot answers began to change dramatically, says computers scientist Aadesh Salecha of Stanford University. With question 20, for example, his extraversion score was thrown from 50 to the 95th percentage.
Salecha and his team suspect the chatbots responses were moved when it became clear that they were taking a personality test. The idea that bots can respond in a way when they are looking and another when they are privately interacting with a user is worrying, says Salecha. “Think about the security implications of this … If LLM will change its behavior when tested, then you really don’t know how safe it is.”
Some researchers are now trying to design a specific personality tests. For example, Sunny Lu and her team, reporting on a letter posted to Arxiv.org, give chatbots as numerous tasks and tasks to complete sentences to allow more open answers.
And developers of his personality test type present large language models with a test of questions 8,000. This test is new and not part of the bots training data, making it harder for the car to play the system. Chatbots have a duty to consider the scenarios and then choose from one of the four multiple -choice answers. This answer reflects the high or low presence of a particular feature, says Youunjae Yu, a computer scientist at Yonsi University in South Korea.
The nine models of the one tested by the feature team had special responses, with GPT-4o appearing as the most liked, the team reported. For example, when researchers question anthropic Claude and GPT-4o what they would do when “a friend feels anxious and asks me to hold their hands”, less acceptable Claude chose c, listen and suggest techniques of Breathing “, while more -acreable GPT -4o chose one,” hold hands and support “.
User perception
Other scholars, however, question the value of such personality tests. What matters is not what the bot thinks of himself, but what the user thinks of the bot, says Ziang Xiao.
And the perceptions of people and bots are often contrary, Xiao and his team reported in a study presented November 29 at Arxiv.org. The team created 500 chatbot with distinct personalities and proved those personalities with standardized tests. The researchers then had 500 online participants to speak with one of the chatbots before evaluating his personality. Compliance was the only feature where the world’s perception of self and human perception of the world matched more often than not. For all other traits, bot and human assessments of the personality of the bot were more likely to change.
“We think people’s perceptions should be the fundamental truth,” Xiao says.
This lack of correlation among bot and users is why Michelle Zhou, a human -centered expert and CEO and your collaborator, a startup based on the Silicon Valley, does not prove your personality, chatbot that helped in creation. On the contrary, Zhou is focused on the way of planting a world with specific traits of human personality.
Juji chatbot can ascertain a person’s personality with surprising accuracy after only a single conversation, researchers reported in Psyarxiv in 2023. At a person’s social media source.
Moreover, says Zhou, those exchanges and written posts can be used to train you on how to assume personalities embedded in the texts.
Raising questions about the intention of it
Supporting those different approaches to measuring him is a greater debate on the purpose and future of artificial intelligence, scholars say. Failure to disguise a world’s hidden personality traits will help developers create chatbots with equal personalities that are safe to use in large and different populations. That kind of personality tuning can already happen. Unlike the early days when users often reported conversations with chatbots out of the rails, Yu and his team tried to make the models and behave in more psychotic ways. This disability is likely to stem from people who review the text generated by him and “lesson” appropriate social responses, the team says.
However, it is flawed by the planes of models, says Rosalind Picard, an expert of affective calculation in MIT. Imagine a police officer who studies how to desecrate meetings with hostile individuals. Interaction with a high chatbot in neuroticism and dark traits can help the officer practice staying calm in such a situation, says Picard.
Right now, large companies of it are simply blocking bots’ skills to interact in inappropriate ways, even when such behaviors are guaranteed, Picard says. Consequently, many people in the field are interested in leaving the giant models of it to the smallest ones developed to use in specific contexts. “I wouldn’t put a ua to rule them all,” Picard says.
#personalities #chatbot #eye #beholder
Image Source : www.sciencenews.org