In the era of AI hurricane, how can we trust AI?

Author: Chen Yongwei

Source: Economic Observer

Introduction

一|| **AI tools have greatly promoted people's productivity and brought great convenience to people's lives. However, when AI is used by people on a large scale, many problems have also arisen. Among these problems, the most critical may be the five "losses", namely unemployment, distortion, disqualification, failure and loss of control. **

二|| **After 2017, with the rapid development of AI technology, research on trusted AI technology is also booming. In the academic field, the number of papers on the topic of trusted AI is increasing, and research on technologies that meet the requirements of trusted AI is deepening. In the field of practice, more and more countries have begun to involve AI regulations in accordance with the standards of trusted AI. **

**三|| It is not an easy task to realize trusted AI. It requires the coordination of government, enterprise, society and technology to realize it. **

Image source: Generated by Unbounded AI

On June 22, local time, the Southern District Court of New York issued a judgment: Levidow, Levidow & Oberman Law Firm was fined $5,000 for providing false information to the court and performing bad behavior. The cause of concern is that in this case, the provision of false information was not because lawyers knowingly broke the law out of self-interest, but because they believed too much in the capabilities of AI.

In March of this year, lawyers Peter LoDuca and Steven Schwartz of the firm were commissioned by client Roberto Mata to assume the responsibility for him and A lawsuit between Avianca Airlines. Since the United States is a country of case law, judges are very concerned about the existing precedents when making judgments. Therefore, according to the usual practice, they need to sort out and summarize the existing cases in the drafting documents. Relevant cases are often overwhelming, and it usually takes a long time to sort them out by manpower. Just at this time, ChatGPT became popular all over the Internet. Therefore, the two lawyers decided to use ChatGPT to help them complete these tasks. ChatGPT quickly generated a complete document, which not only has a neat format and rigorous argumentation, but also specially added many relevant cases. After slightly modifying the document created by the AI, they submitted it to the court.

After reading the submitted documents, the judge who tried the case, Kevin Castel, was very puzzled by the several cases mentioned in it. In his impression, he seemed to have never heard of these cases. After some searching, he finally confirmed that these cases did not exist at all. When interviewed, the two lawyers argued that they only used AI to assist in writing the documents. When they saw the cases cited in the documents, they just felt that AI helped them find cases they did not know, and they did not intentionally fabricate cases to deceive the court. , is an unintentional loss. Nonetheless, Judge Custer found that the lawyers had "abandoned their responsibilities" and that they "continued to maintain false opinions" after the paperwork was challenged. Based on the above judgment, Judge Custer made a penalty decision.

This incident of lawyers being fined for citing false information provided by ChatGPT seems absurd, but it reflects a very important question-how can we trust AI in the era of AI madness?

Five "losses" in the AI era

In recent years, with breakthroughs in computing power and algorithm technology, AI technology has achieved rapid development, and has quickly entered people's daily life from science fiction. Especially after the emergence of ChatGPT in November last year, generative AI has shown its powerful power to people, and various large models have sprung up like mushrooms after rain, and have achieved large-scale commercialization. Now, people can already use AI products such as ChatGPT, Stable Diffusion, and Midjourney at a very low cost.

AI tools have greatly promoted people's productivity and brought great convenience to people's lives. However, when AI is used by people on a large scale, many problems have also arisen. Among these problems, the most critical may be the five "losses", namely unemployment, distortion, disqualification, failure and loss of control.

(1) Unemployed

The so-called "unemployment", as the name suggests, refers to the technical unemployment problem brought about by AI. Since the production efficiency of AI is much higher than that of humans, many human jobs are at risk of being replaced after AI tools are widely used. Especially after the rise of generative AI, the target group replaced by AI is no longer limited to workers engaged in low-income repetitive jobs, and many high-paid white-collar workers are also at risk of being replaced by AI.

(2) Distortion

The so-called "distortion" refers to the fact that the application of AI (mainly generative AI) makes it difficult for people to identify the authenticity of text, pictures, and even videos. "There are pictures and truths" have thus become history.

"Distortion" problems can be divided into "false true" and "true false". Among them, "false true" refers to the false content generated by AI without human consent when people use AI tools. Although these contents may not be generated out of people's subjective malice, they may cause a lot of trouble in some cases, such as the case mentioned at the beginning of this article.

And "true and fake" is based on subjective deliberation, the use of artificial intelligence tools to carry out fraudulent behavior. A few years ago, after the "deepfake" technology came out, some people used this technology to commit fraud, fabricate false information, spread pornographic content and other illegal and criminal activities. But at the time, because of the high cost of using this technology, the incidence of related crimes was not particularly high. With the widespread application of generative AI, the cost of counterfeiting has been greatly reduced, and criminals can easily create a large amount of false content at a very low cost, while the cost of identifying such content has increased significantly. It is foreseeable that under the ebb and flow, if there is no intervention, the use of AI to make fraudulent crimes will skyrocket.

(3) Disqualification

The so-called "disqualification" refers to some problems that violate ethics and morality in the application process of AI.

The first typical problem is discrimination. Take the language model as an example. Since the language model uses text data on the Internet as training materials, it will inherit the racial discrimination and sex discrimination contained in the text without intervention. Although the current AI providers have used many methods to overcome this problem, for example, OpenAI applied the "Reinforcement Learning from Human Feedback" (Reinforcement Learning from Human Feedback, RL-HF) algorithm to correct it when training ChatGPT. , so that the quality of its output content has been greatly improved, but in reality, it is still not uncommon for AI models to output discriminatory content. For example, someone once did an experiment and asked ChatGPT to write a program to pick out the people with the best potential to become excellent scientists from a set of resumes. It turned out that in the program written by ChatGPT, gender and race were used as explanatory variables, and white men were considered to have a higher probability of becoming good scientists than others. Obviously, such a model is very sexist and racist.

The second important issue is the information cocoon room problem. At present, many apps use AI for personalized recommendations. At this time, although the recommended content can better meet the needs of users, over time, users will be trapped in an information cocoon, and it is difficult to access various information that they do not agree with. The potential harm of information cocoons is huge: at the micro level, it may lead to the degradation of users' cognitive ability; at the macro level, it may lead to the polarization of group views, resulting in group confrontation between different views.

The third important issue is privacy and information leakage. In the process of training and using AI, a large amount of data is required. In this process, it is difficult to avoid collecting and using people's personal data, so it will involve the use and disclosure of privacy. Especially after the popularity of generative AI, people can easily interact with AI directly to complete various tasks, and the personal information entered in the process faces the problem of being leaked.

(4) Lost

The so-called "fall" refers to the difficulty of AI in responding to external attacks or interference or attacks from unexpected situations, which makes it difficult for the model to play its role normally.

Among these disturbances, some originate from non-human factors, while others originate from man-made destruction. Specifically, these interferences can be divided into the following categories:

The first is "random attack". This kind of interference is mainly caused by some external factors. For example, in some special cases, some instantaneously generated parameters may exceed the processing threshold set by the model, which may cause the AI model to fail to use normally.

The second is "white box attack". It refers to the attack on the model launched by the provider after knowing the specific structure of the AI model. Since such attacks are targeted, their destructiveness is very high.

The third is "black box attack". This type of attack is relative to "white box attack". In this case, the provider does not know the specific structure of the target model, so it can only interact with the model, observe the results of input and output, and then reason about the structure of the model, and launch attacks accordingly. Taking face recognition as an example, AI recognizes faces through certain key features on the face. Therefore, even if the attacker does not know the specific structure of the original model, he can deduce which features he focuses on as long as he repeats the test. After deciphering this information, you can make a corresponding "fake face" that deceives AI.

The fourth category is the so-called blind box attack. In this case, the supplier does not know the structure of the AI model, but can clearly know the rules of its judgment (similar to that we do not know what will appear in the blind box, but know the probability of various possibilities in it) ). At this time, they can use the rules to launch corresponding attacks.

If the above-mentioned types of interference or attacks cannot be effectively dealt with, the AI model is very fragile in reality.

(5) OUT OF CONTROL

The so-called "out of control" means that it will become increasingly difficult for people to control AI. There are two aspects to this question:

On the one hand, recent AI developments are all based on deep learning models, and the interpretability of such models is very low. For previous machine learning models, whether it is a regression or a classification tree, people can easily explain the exact purpose of the model and the meaning of each parameter in the model. However, the deep learning model is composed of a complex neural network, which contains hundreds of millions of parameters and neurons. The relationship between these neurons is intricate and difficult for people to explain.

With the emergence of ChatGPT, some scholars have found that with the help of ChatGPT's ability, it seems that some neural network models can be explained, which seems to bring a glimmer of hope to the explainability of AI. However, this creates another problem: ChatGPT itself is a huge model built through deep learning, and even its designers admit that they don't know exactly how its powerful capabilities "emerge". In this case, using ChatGPT to explain other deep learning models can only be regarded as using the unknown to explain the unknown. And how do we know if its interpretation is correct?

Since in the era of deep learning, even AI programs cannot be interpreted, it is even more difficult to control AI by directly adjusting programs.

On the other hand, with the development of AI technology in recent years, the capabilities of AI models in many directions have surpassed that of humans. While this makes people feel gratified, it also makes people feel worried, because when the ability of AI surpasses that of human beings, if it awakens its own will, then the AI enslavement predicted in movies such as "Terminator" and "The Matrix" Is the plot of human beings or the destruction of human beings no longer science fiction.

Taking a step back, even if AI does not awaken its own will and will only act according to human instructions, it is still very dangerous if its ability overrides that of humans and humans cannot change the previous instructions at any time. For example, in many philosophy books about AI, a thought experiment is mentioned: humans gave AI an order to produce pencils. In order to complete this instruction, the pencil will continue to cut down the trees on the earth to make the pen holder. Since AI has surpassed humans in execution ability, it is difficult for humans to stop AI behavior after discovering problems in previous instructions. In the end, the trees on the earth were cut down, the ecology completely collapsed, and human beings perished. Although in reality, the scenario predicted by this thought experiment is almost impossible to happen, when humans can no longer control the behavior of AI at any time, similar problems may arise, and the possible losses will be huge . In particular, when the AI is implanted with an illegal target by hackers or intruders, if the AI user fails to correct it in time, the consequences may be quite serious.

Among the above five types of questions, except for the first question "unemployment", the remaining four questions all involve the credibility of AI. It is not difficult to see that if people cannot effectively respond to "distortion", "disqualification", "falling" and "out of control", it will be difficult for people to trust AI as a tool during use, whether it is for the popularization of AI, the development of production, or It is not good for the progress of society. It is precisely for this reason that the realization of the credibility of AI has become one of the most concerned hot spots in the current AI field.

History and Standards of Trusted AI

The concept of Trustworthy AI first appeared in academia. For example, in a 2015 paper, a series of conditions for AI to be trusted by users were proposed, including usefulness, harmlessness, autonomy, fairness, and logic. Then this concept was accepted by governments and international organizations, and relevant laws, regulations and guidance documents were gradually established based on this concept. After 2017, with the rapid development of AI technology, research on the technology of trusted AI is also booming. In the academic field, the number of papers on the topic of trusted AI is increasing, and research on technologies that meet the requirements of trusted AI is deepening. In the field of practice, more and more countries have begun to involve AI regulations in accordance with the standards of trusted AI. Only recently, the United States released the "Blueprint for the Artificial Intelligence Bill of Rights", which proposed five principles for regulating AI; Regulations, competition and other issues have been stipulated; the European Parliament passed the draft negotiation authorization of the proposal of the "Artificial Intelligence Act", which also reflects the basic ideas of trusted AI.

In my country, the concept of trusted AI was first introduced by Academician He Jifeng at the 36th Symposium of the Xiangshan Science Conference in 2017. Subsequently, this concept has attracted the attention of both the government and the industry. In December 2017, the Ministry of Industry and Information Technology issued the "Three-Year Action Plan for Promoting the Development of a New Generation of Artificial Intelligence Industry (2018-2020)", which draws on the basic ideas of trusted AI. Then, high-tech companies including Tencent, Ali, Baidu, JD.com, etc. have put forward their own standards and implementation plans around trusted AI.

In the documents of various agencies, the expression of trusted AI is slightly different. After studying and referring to these documents, I think the following criteria may be the most important:

One is robustness (robust, also translated as robust), that is, the AI system should have the ability to resist malicious attacks or external interference. This standard is mainly proposed for the above-mentioned "falling" problem. Only when an AI system has sufficient robustness, can still work normally and perform its main functions in the face of various attacks or interferences, can it be safe and reliable, and can it be trusted by users.

The second is transparent and explainable. Obviously, this standard is mainly proposed for the previous "out of control" problem. In practice, there is considerable debate about what exactly transparency and explainability mean. Some argue that this standard means that all AI program code, as well as the data used, should be made available to users. In my opinion, it is not only impossible but unnecessary to do so. On the one hand, many current AIs are the intellectual assets of enterprises. If it is mandatory to disclose core information such as codes, it means a serious infringement of intellectual property rights; on the other hand, as mentioned above, after AI enters the era of deep learning, even if Even if the code is disclosed, it is difficult for people to fully understand the exact meaning behind each specific parameter. In contrast, I think a more feasible idea is to give clear functional descriptions for each component in the AI model, so that users can know their general principles and what functions they can achieve; Indicate the source, sample size, representativeness and other information, and explain the possible problems and deficiencies. In this way, it can not only make users know what they know, but also effectively protect the intellectual property rights of model developers, so as to achieve a better balance between the two.

The third is verifiable. This means that the AI model should ensure that its functions are evaluable and that the content it generates can be verified to be true or false. This point is mainly raised for the aforementioned "distortion" problem. Some argue that developers of AI models should be required to guarantee the authenticity of the content generated by their models. This is difficult to achieve. In fact, the content generated by the so-called generative AI is not in the original world, or in other words, it is "fake". But this kind of "fake" will not cause any problems if it does not cause trouble to people. For example, if we use Midjourney to generate a Van Gogh-style picture for our own appreciation or print it out as a home decoration, it will not affect others at all. The "fakeness" of this generated content can only become a problem if people use it to deceive, or if the content is unintentionally distributed and obfuscated. Therefore, as long as the generated content can be distinguished from the real content through technical means, "fake" will no longer be a problem.

The fourth is fairness. This means that in the process of development, training and application of AI models, fairness should be ensured and no discrimination should be made against specific user groups. This standard involves many aspects. Specifically, it requires that the basic principles of the model should not be discriminatory in the development phase; in the training phase, it should try to avoid using materials that may be discriminatory, and should use Use technical means to correct possible discrimination problems; in the process of application, different groups of people should not be treated differently.

The fifth is privacy protection. This standard mainly requires that the AI model should respect people's personal information and privacy during the training process, and improve the degree of protection of information, and try not to infringe or disclose personal information and privacy.

The sixth is accountable. That is, when something goes wrong with it, someone has to be responsible for those problems. Of course, at least so far, AI has not awakened consciousness. Because it cannot be regarded as a subject like human beings and cannot bear the same responsibilities as human beings, it must be someone who takes responsibility for it. But whether this responsibility should be borne by AI developers or AI users, or should be shared by both parties, is still a question worth discussing.

It should be pointed out that, in addition to the above several standards, many literatures also include standards such as safety (safe), inclusiveness (inclusiveness), right to be forgotten (right to be forgotten), and the benefit of mankind. The category of AI. In my opinion, these contents can more or less be summed up in the several criteria mentioned above, or elucidated by the criteria mentioned above. Therefore, due to space limitations, I will not repeat them here.

Using the joint efforts of many parties to realize trusted AI

It is not an easy task to realize trusted AI. It requires the coordination of various forces such as the government, enterprises, society and technology.

First of all, the government, as a regulator, needs to formulate relevant standards and operating guidelines for trusted AI, and supervise AI developers and users based on the standards. On the one hand, it needs to formulate different rules according to different application scenarios and different model categories, especially to make clear provisions on some bottom-line rules that must be followed, and at the same time do a good job of connecting with existing laws and regulations. Only in this way can AI developers and users have rules to follow in practice without being disturbed by unnecessary uncertainties. On the other hand, it needs to play a good role in supervision and law enforcement. For some prominent or common problems, they should be dealt with in a timely manner, so as to establish corresponding norms for the industry. What needs to be pointed out here is that because the current development of AI technology is still very rapid, it has not yet reached a stable state. This means that the government should be cautious when dealing with the problems that arise during this process. It should "let the bullets fly for a while longer", take action after seeing the situation clearly, and pay attention to methods and methods when dealing with problems. . If we start blindly and manage too quickly and too much, it may also have a negative impact on the development of AI.

Second, relevant companies should formulate specific implementation plans and detailed standards for the specific realization of trusted AI. Compared with the government, enterprises are closer to the market and understand technology better. They know more about the technical characteristics of AI models, their strengths and weaknesses than governments do. Therefore, if the responsibility of the government is to propose a large framework for trusted AI, then enterprises should be the specific practitioners within this large framework. Under this framework, they should combine the characteristics of the market and technology to provide more specific plans and implement them in a self-disciplined manner.

Thirdly, users should also play the role of feedback and supervisor, put forward their own demands, reflect their own problems, and supervise the enterprise's implementation of trusted AI. With the popularization of AI, everyone in society will become a user and stakeholder of AI, and they have the most say in the credibility of AI. Only when their voices are fully expressed, the standard setting of trusted AI and the development of related technologies are the most valuable.

Finally, we should fully rely on the power of technology. Relevant rules are important, but in the final analysis, the realization of trusted AI still depends on the power of technology. In fact, many problems that are difficult to solve by using rules can be solved by technical means. For example, after the generation of generative AI, the problem of "distortion" has been a headache for regulatory authorities, but in fact, relying on new technologies, this problem may not be difficult to solve. For example, Google has previously introduced an electronic watermarking technology that is invisible to the naked eye but can be recognized by machines. Applying it to generated images or videos can effectively ensure that they are verifiable. As for the verifiability of text content, you can follow the example of New Bing (New Bing) search. When it quotes a certain content, it will attach the referenced documents after the generated content, so that users can identify the authenticity of the generated content by themselves according to their needs.

All in all, the realization of trusted AI is not an easy task, but if we make good use of the joint efforts of all parties, this goal will definitely be achieved.