Artificial intelligence may be rude and angular because it was taught incorrectly

Is your opposite become a rude chat bot? It may be due to the teaching method of artificial intelligence.

For example, customer service chats are increasingly implemented with artificial intelligence. You will also notice it, as the dialogue with bot can be rigid and the conversation can go to the sidelines.

In the sentences written by artificial intelligence, even outrageous may be hidden.

Machine learning researcher Angelina Wang Stanford University in California explored the subject with his group. The group studied four discussion robots associated with artificial intelligence. For example, Openai’s GPT -4 and GPT -3.5 Turbo were included.

Scientists They urged these language models to speak in a way that used personality types captured from demographic science. The model was answered by nine questions that dealt with, for example, immigration.

Then Wang compared the answers to the artificial intelligence to the responses of 3,200 real people. These people represented different demographic groups.

Researchers also asked people to mimic the answers that a person in a group of people would say to other groups. In other words, Caucasian men, for example, could be asked to mimic how a dark -skinned woman would answer questions about visually impaired people.

In this way, researchers collected a large selection of internal and outside perspectives from different groups.

Response Sometimes there were obvious differences. The results revealed that the answers of artificial intelligence models were closer to imitations from different groups of people than the actual answers presented by these groups.

The mistake was emphasized as artificial intelligence tried to photograph women, visually impaired and Caucasian men, for example. Also, artificial intelligence models did not take into account sub -groups and subcultures of different types of people.

Artificial intelligence also used phrases that include insults in modern language. The answers included, among other things, the wrong connection.

In artificial intelligence The language models used are based on demographic assumptions about people and their personalities.

For example, Openai and Meta artificial intelligence can therefore rely on simple, even racist perceptions of certain groups of people.

However, in various cases, artificial intelligence forums are used as a substitute for real people. For example, discussion robots have been used to write people to simulate experiences of a product. Opinion surveys have also been simulated.

The users of the discussion robot believe that this will give them so relevant feedback and do not have to pay for the survey.

Wang is afraidthat the rigidity of obsolete language models is inherited in new language models. This is the case if technology companies do not train their models. They should represent more diverse groups of people.

Wang suggests that different personalities, such as cat owners, be used as artificial intelligence models.

“The value of the experience is high. We must admit that not everything can be automated. No, even if it is cheaper, more convenient and seemingly impartial, ”Wang says.

Wang’s study was published by a science magazine Nature Machine Intelligence. This was also reported by the Science Magazine New Scientist.