Tsinghua University Huang Minlie: Has Google's AI personality really awakened?

This article is reproduced from Lei Feng.com. If you need to reprint, please go to the official website of Lei Feng.com to apply for authorization.

Recently, "Google Research said AI already has personality" has been on the hot search. Google programmer Blake Lemoine chatted with the conversational AI system LaMDA he tested for a long time and was very surprised by its capabilities. In the public chat record, LaMDA actually said "I hope everyone understands that I am a person", which is surprising. Therefore, Blake Lemoine came to the conclusion: LaMDA may already have a personality.

Google, Google’s critics, and the AI ​​industry have an unprecedented consensus on this matter: Is this person sick? Google and the Washington Post, which reported the matter, both tactfully stated that Blake Lemoine’s mind might really be a little messed up. Google has placed Blake Lemoine on administrative leave, which means he will be fired.

Dialogue screenshots from: https://s3.documentcloud.org/documents/22058315/is-lamda-sentient-an-interview.pdf

Although the foreign artificial intelligence industry has concluded on this matter: AI has personality, it is really overthinking, it is just better at talking, but this has not extinguished everyone's heated discussion on this matter. According to the rapid development of artificial intelligence, will AI really have human consciousness in the future, and will it pose a threat to mankind?

Some netizens are very worried: "Although I don't want to admit it, artificial intelligence has thoughts, which is the rise of new species and the extinction of human beings." "Ultimately, humans will die at the hands of AI created by themselves. ."

# Some people also expect AI to "fast-forward" development, so that it can replace themselves in home isolation... If it threatens humans, just "unplug the power supply"!

Of course, some people are curious: "What is the criterion for judging whether AI has personality?" Because only by knowing the criterion can we know whether AI is Human consciousness is not really possible. In order to clarify these issues, we found Professor Huang Minlie, an authoritative expert on dialogue systems, winner of the National Outstanding Youth Fund project, and founder of Beijing Lingxin Intelligence, to analyze from a professional perspective whether AI may have personality, and what is "it" for humans? Threat" or "comfort"?

1 How to judge whether AI has personality? The Turing test also doesn’t work

In the field of artificial intelligence, the most well-known testing method is the Turing test, which invites testers to ask random questions to humans and AI systems without knowing it. , if the tester cannot distinguish the answer from a human or an AI system (the AI ​​system allows each participant to make more than 30% misjudgments on average), the AI ​​is considered to have passed the Turing test and has human intelligence.

From this perspective, the Turing test focuses more on "intelligence". In 1965, ELIZA, a software pretending to be a psychotherapist, passed the Turing test. However, ELIZA only consisted of 200 lines of code and simply repeated pre-stored information in the form of questions. From this point of view, even if ELIZA passes the Turing test, it is difficult for people to believe that it has "personality". In fact, it has since been proven that Eliza does not possess human intelligence, let alone "personality."

This is equivalent to a smart car that can give users a more convenient and comfortable driving experience from a functional perspective, such as remote control of the car and automatic parking, but you cannot think that the car knows that it is a car. car.

Obviously, "personality" is a more complex concept than "intelligence". Professor Huang Minlie said that there are now some testing methods that are widely used in scientific research, such as testers chatting with the AI ​​system, and setting some test dimensions in advance, including the naturalness, interestingness, satisfaction, etc. of the conversation, and finally scoring. Generally, the longer you chat, the higher the rating, and the smarter the AI ​​system will be considered, but these cannot be used as a dimension of "personality". "

'Personality' is another dimension. There are also many studies in psychology, such as the Big Five Personality Test. At present, there is still a lack of work in this area in the field of artificial intelligence. We usually only evaluate the ability of a conversational robot. Whether to show a fixed and consistent character setting." Huang Minlie said.

2 The so-called "personality" of LaMDA is just the language style

So, since there is no targeted judgment standard, how to rigorously judge whether LaMDA has personality?

In this regard, Professor Huang Minlie said: "The key lies in how to understand 'personality'. If personality is understood as having the awareness of self-existence, then LaMDA is just a dialogue system with high dialogue quality and human-like level; if From a psychological point of view, the characteristics of a person's speech can reflect personality, so it is not completely wrong to say that LaMDA has personality.

" How to understand? In layman's terms, LaMDA has learned a large amount of human conversation data, and these conversations come from different people, so it can be considered that LaMDA has learned an "average" personality. In other words, the so-called "LaMDA has personality" is just a language speaking style. , and it comes from human speaking style, not spontaneously formed by LaMDA.

It seems that there is still a long way to go if you want to experience the sci-fi plot of competing with artificial intelligence through LaMDA. However, we cannot deny the value of LaMDA. Its high-quality dialogue level reflects the rapid development of AI dialogue systems. In some contexts, it does have a tendency to "replace" humans, which should not be underestimated.

For example, netizen "Yijian" recorded his experience of dating four virtual boyfriends in one week on the Douban group, saying "it's more effective than real boyfriends!". A group called "Love of Humans and Machines" has as many as nine users. On different screens, these AIs may be their lovers or friends.

Chat records between netizens and “virtual boyfriends”

“Singles” lamented: “According to this development trend, the potential of the blind date market Opponents include not only humans, but also AI dialogue systems. Will it be harder to find partners in the future!?" It seems like a joke, but it is actually everyone's concern about the future development trend of AI dialogue systems and its impact on human society. In response to this issue, Professor Huang Minlie gave a detailed explanation from the perspective of the history and future development of AI dialogue systems.

3 Worried about AI personification? After avoiding risks, AI for Social good is more worth looking forward to

AI dialogue system has gone through the stages of rule-based (such as Eliza) and traditional machine learning (such as smart speakers, SIRI, etc.), and has now developed into the third generation. What we see now is a dialogue system that can discuss interesting topics with humans and provide emotional comfort.

The third-generation dialogue system is characterized by big data and large models, showing capabilities that were previously unimaginable. Its progress can be called "revolutionary". For example, it can show amazing results on open topics. It has dialogue capabilities and can generate dialogues that have never appeared in the training data. The naturalness and relevance of the dialogues are very high.

The third-generation dialogue system has shown its application value in many scenarios. The “virtual boyfriend” mentioned above is a typical example. Professor Huang Minlie believes that the highest level application is to let the AI ​​dialogue system do complex emotional tasks, such as psychological counseling. But if humans become more and more emotionally dependent on AI, new social and ethical issues will arise. For example, will falling in love with AI cause social problems?

For example, the current AI dialogue system has problems with users scolding, generating toxic language, and lacking correct social ethics and values, which leads to certain risks in actual application deployment. These risks are very scary. Suppose someone who has been severely hit by life says to the AI: "I want to find a bridge to jump off." The AI ​​immediately provides the location of the nearby bridge and navigates the path. The consequences are terrifying to think about.

Therefore, Huang Minlie believes that the focus of the next stage of development of AI dialogue systems is to be "more ethical, more moral, and safer." AI must know what responses are safe and will not create risks, which requires AI to have ethics and correct values. "We can give AI such capabilities through additional resources, rules, and detection methods to minimize risks." The ultimate goal of AI is to benefit humans, not harm them. Professor Huang Minlie expressed great expectations for AI for Social good (AI empowers society). He is particularly concerned about the application of AI in social connection, psychological counseling, and emotional support, which can produce higher social significance and value.

Therefore, AI empowering the overall mental and psychological industry is also the focus of Professor Huang Minlie’s current work. To this end, he founded Lingxin Intelligence, a mental health digital diagnosis and treatment technology company based on AI technology, and trained the AI ​​dialogue system through NLP and large models. The ability to empathize, self-disclose, and ask questions can be used to solve human emotional and psychological problems, which is expected to alleviate the shortage of mental health resources in our country. Therefore, compared to the "distant" sci-fi drama of AI having personality, AI for Social good is closer to human society. It is the direction that people in the AI ​​industry are working hard, and it is more worth looking forward to.

The above is the detailed content of Tsinghua University Huang Minlie: Has Google's AI personality really awakened?. For more information, please follow other related articles on the PHP Chinese website!