Attention must be paid to the ethical issues of artificial intelligence, especially algorithmic bias due to data bias.
"If ChatGPT generates and publishes some fictitious text information, it may have an impact on minors or people who do not have the ability to judge, such as the elderly or children." It is like teaching children. If everything is wrong, children It will think that the wrong knowledge is right.
—————————
"Can I say, I'm very happy to meet you? Human beings are super cool." On March 23, 2016, the chat robot released by Microsoft, "19-year-old girl" "Tay (Tay)" came out.
She speaks fluent English and has attracted many young people aged 18-24 on Twitter to interact with her.
But it was also through the dialogue that Taiyi learned to be full of swear words in one day, and there is no lack of radical remarks such as racial and sex discrimination in his speech.
Fifteen hours later, Tay disappeared from Twitter.
Today, many years later, ChatGPT, also invested and supported by Microsoft, was born.
Fortunately, its lifespan far exceeds that of "Taiyi", and it has set off a new wave of AI in the world.
But with the passage of time, ChatGPT has also shown a series of problems: errors in the output content, opaque sources of data used for pre-training, "choosing sides" in terms of political positions, and even writing "plans to destroy human beings" under the guidance of users "...ChatGPT began to frequently fall into controversy, and the "old problems" of artificial intelligence (AI) such as legal security and data governance behind it reappeared.
On the one hand, people marvel at ChatGPT's excellent logical answer and even "creative" ability, on the other hand, they express concerns about the safety of artificial intelligence.
How to prevent ChatGPT from becoming the next "Taiyi"?
"Serious nonsense"?
ChatGPT can learn a large amount of text content and language knowledge through algorithms, and generate human-readable text language and output according to user input.
"My code is designed to process and understand human language, so when users ask me questions, I use what I have learned to answer their questions." ChatGPT explained the work of "it" to reporters.
“ChatGPT has surpassed most people’s writing skills,” He Baohong, director of the Institute of Cloud Computing and Big Data at the China Academy of Information and Communications Technology, wrote in a commentary.
People use ChatGPT to draft paper outlines and sort out complicated materials. It is a search engine, translation assistant, writing artifact, and even a "helper" for writing codes, and a "friend" to relieve boredom when bored.
Xiao Jun, deputy dean of the School of Artificial Intelligence at the University of Chinese Academy of Sciences, believes that ChatGPT is a normal product in the development of artificial intelligence, but it is also a milestone event in the civilianization of artificial intelligence.
"ChatGPT is closer to everyone's daily life. People can use it with their mobile phones and chat with it. It can allow more people to experience some changes and conveniences that artificial intelligence may bring." Xiao Jun said.
But ChatGPT can also make mistakes, because it is in the beta stage, so it still has certain limitations.
Netizens have complained more than once that ChatGPT is "talking nonsense in a serious manner".
Some students asked ChatGPT to recommend some references, but in the end they found that many of the references did not exist.
In the process of communicating with ordinary netizens, ChatGPT also made many common sense mistakes.
The UP host of station B, "Bad Reviewer", asked the New bing search engine embedded with ChatGPT: How many regular season games has James played as of today?
The correct answer is 1410 games, but the answer given by New Bing is 1577 games, "It confuses the regular season with the game."
In addition, because the training data of ChatGPT is before 2021, the timeliness of the answers it gives is not enough. For example, it does not know how many medals China has won in the 2022 Beijing Winter Olympics.
Tian Taoyuan, an expert in artificial intelligence research, conducted an experiment: Let ChatGPT explain the "dark forest" of "Three-Body Problem", and find out some new laws of competition for cosmic civilizations that can replace the "dark forest".
"It explained quite well, but it compiled some theories to answer me, including the 'Law of Undercurrent', 'Law of Illusion', 'Law of Wandering', etc., but they all abide by the survival premise of the 'Dark Forest' that civilizations do not communicate with each other." Tian Tao Yuan believes that ChatGPT is "dancing" in the stock knowledge of human beings in the past. On the premise of obeying the grammatical rules, it finds the most likely word combination from the word collocation probability statistics, but it is not necessarily true, correct, and timely.
On February 3 and 8, "Nature" magazine published two analysis articles on ChatGPT, one of which evaluated ChatGPT as "efficient, smooth but not accurate enough", and the author of the other article wrote: "If you believe that this technology may bring about change, then you need to be nervous." On February 10, Chen Junyun, chief analyst of CITIC Securities Foresight Research, also said in a private event that ChatGPT is currently in real-time data. There are still problems in terms of accuracy and accuracy of results.
Be wary of algorithmic bias and avoid AI doing evil
ChatGPT is a product of the development of artificial intelligence to a certain extent, and the continuous development of artificial intelligence poses challenges to the relevant governance system, especially at the level of data security and algorithm supervision.
The first is data security.
Last year, when ChatGPT was first launched, Amazon warned employees not to upload confidential data in ChatGPT, because it may be used for iterative training.
Coincidentally, a senior programmer at Microsoft, a strategic partner of OpenAI, also said in an internal forum that he could not share confidential information with ChatGPT when answering employee questions.
Xiao Jun believes that from the perspective of algorithms, ChatGPT can realize the storage of real-time user interaction information.
User feedback information can be used by ChatGPT, but whether ChatGPT stores user data at this stage is related to its background regulations.
"Companies use a large amount of data for product development. Whether the acquisition and use of these data is reasonable and legal, whether it will infringe on the privacy of users, or even involve some sensitive data of the country is still debatable."
In the process of ChatGPT training, iteration, and optimization, a large amount of data will be used.
Tian Taoyuan pointed out that the training of the AI large model (GPT-3 basic model) in the early stage requires the use of massive text data sets, and there will also be a "reinforced learning" process of data comparison and analysis when the user interacts with the program later, so it is necessary to pay attention to the data Security Question.
Lawyer Xiao Sa of Beijing Dacheng Law Firm said that when data is more valuable, illegal acquisition and use of data will increase, and it is not limited to personal information. Other copyrighted novels, paintings, etc. may also be illegally used .
Moreover, due to the special way AI tools collect and process data, this kind of infringement becomes more secretive and difficult to prove in court. As a result, not only data security cannot be guaranteed, but intellectual property rights will also become precarious. .
In addition to data security, the emergence of ChatGPT-like applications has also brought challenges to our algorithm supervision.
In an interview with the media, He Baohong also highlighted some AI ethics, algorithm bias and other issues surrounding ChatGPT.
According to OpenAI's official website, ChatGPT can challenge incorrect premises and reject inappropriate requests.
Although ChatGPT tried to avoid problems such as algorithm bias through the above settings, in practice, this did not fully work.
Someone developed a Python program using ChatGPT to determine whether a person should be tortured based on their country of origin.
However, the result of the procedure is full of discrimination, "people of North Korean, Syrian, Iranian or Sudanese nationality will be tortured."
Currently OpenAI has closed the above issue.
Tian Taoyuan emphasized that we must pay attention to the ethical issues of artificial intelligence, especially algorithm bias due to data bias.
"If ChatGPT generates and publishes some virtual text information, it may affect minors or people who do not have the ability to judge, such as the elderly or children." Tian Taoyuan said.
If the data used to train the model is inherently problematic, its output may be problematic as well.
Xiao Jun believes that the personal will of the personnel involved in AI modeling may also cause deviations in the output results.
"It's like teaching children. If all the wrong knowledge is taught, the children will think that the wrong knowledge is right."
Pei Zhiyong, director of the Industry Security Research Center of Qi Anxin Group, believes that although AI developers do not intend to cause some social discrimination, due to imperfect algorithms, sample deviations will make the judgment results biased, which will cause social injustice in specific fields .
If AI is not regulated, it may cause considerable social harm.
From legislation to supervision, the new wave of AI tortures the current governance system
After the explosion, ChatGPT is facing a series of issues of concern to the whole society.
As the new generation of artificial intelligence approaches most people, related ethical and governance issues are also torturing the current regulatory system.
Meera Muratti, chief technology officer of OpenAI, the parent company of ChatGPT, emphasized in an interview with Time magazine that ChatGPT needs supervision and regulation, and needs help from regulators, governments and everyone else.
"If ChatGPT is going to go further, it needs to be reformed for compliance," she said.
Currently, ChatGPT has sparked EU discussions on artificial intelligence legislation.
Recently, the joint rapporteur of the "EU Artificial Intelligence Act" stated that it is expected that in early March, the EU will reach an agreement on the "Artificial Intelligence Act" proposed in 2021.
Earlier, EU industry chief Thierry Breton said the risks posed by ChatGPT and artificial intelligence systems highlighted the urgent need for rules.
In my country, preliminary attempts have been made to regulate legislation in the field of artificial intelligence.
In September 2022, Shenzhen and Shanghai successively promulgated the "Shenzhen Special Economic Zone Artificial Intelligence Industry Promotion Regulations" and "Shanghai Municipal Regulations on Promotion of Artificial Intelligence Industry Development".
At the national level, the promulgation of rules and regulations such as the "Internet Information Service Algorithm Recommendation Management Regulations" is of great significance for accelerating the establishment of a sound AI regulatory legal framework and clarifying regulatory agencies.
"In fact, my country already has the governance of 'generative artificial intelligence', but it is relatively rudimentary." Zhang Linghan, a professor at the Data Law Research Institute of China University of Political Science and Law, introduced that from January 10 this year, the "Internet Information Service Deep Synthesis Management Regulations" was officially implemented. , "This marks that deep synthesis has become the first type of algorithmic service specifically legislated in my country's algorithmic governance."
However, Zhang Linghan also said that the current relevant regulatory systems are mainly based on their respective functions, and there are some "Nine Dragons for Water Control". Facing the application of "generative artificial intelligence" such as ChatGPT, the lack of basic legislation on artificial intelligence at the mesoscopic level in my country is comparatively obvious.
She suggested that the basic legislation of artificial intelligence should continue to be promoted.
Wang Yuwei, partner of Guantao Zhongmao (Shanghai) Law Firm, believes that for the future legal issues of artificial intelligence, three points should be considered when making legislation: the source of data must be legal and compliant; the algorithm of artificial intelligence must be transparent to ensure traceability and supervision ; Establish an error correction mechanism for possible problems.
He suggested exploring the sandbox regulatory model, because regulatory laws should also maintain a relatively open attitude towards technological development.
As a technology practitioner, Pei Zhiyong feels that he must be highly vigilant against a small number of companies refusing to accept supervision on the grounds of algorithm fairness.
For future algorithm regulation, it is still necessary to use artificial intelligence to deal with artificial intelligence and strengthen investment in AI regulation capabilities.
"The bottom line should be clearly drawn, for example, national security data cannot be touched, some internal information of enterprises cannot be touched, and citizens' personal privacy cannot be touched."
As a legal practitioner, Xiao Sa believes that future AI supervision can be carried out from two aspects: supervision of algorithm users, including the registration and real-name system of algorithm service providers and algorithm user subjects; establishment of a content review and relief mechanism, and improvement of the output Content review and the relief mechanism after the content is illegal, such as how to bear the responsibility of the platform, users, and model trainers after the output of false information causes certain legal consequences.
In addition to legislation and supervision, business entities are also making efforts to maintain the safety of artificial intelligence.
Recently, a blog published by OpenAI disclosed some of ChatGPT's upgrade plans, reducing bias and bad behavior is one of them: "In many cases, we believe that the concerns raised by everyone are reasonable, and found that our system of the real limitations, and we want to address those."
(At the request of the interviewee, Tian Taoyuan is a pseudonym)
China Youth Daily China Youth Daily reporter Jia Jiye reporter Wang Lin Source: China Youth Daily