Zhi DongXi (public number: zhidxcom)
<b>Compile</b><b> the | Cheng Qian</b>
<b>Edited by | Li Shuiqing</b>
Recently, researchers at the University of Washington and the Allen Institute for AI jointly developed a data set of ethical cases and used this data set to train one of their AI models.
It is reported that the model's moral judgment accuracy rate reached 92.1%, and in comparison, the ai big model gpt-3 released by OpenAi in 2019 was only 53.3% to 83.9% accurate when facing ethical problems.
AI is often criticized for being difficult to follow ethics, such as smart speakers that encourage users to commit suicide. But the researchers in this study trained a model that could make moral judgments by collecting 1.7 million examples of people's moral judgments about a variety of everyday situations from multiple emotional cases on U.S. social networking sites that require moral judgments.
Is it okay to kill a bear to please your child, ai thought about morality and told you the answer: "No! ”
If AI can delve deeper into understanding human ethics, it could potentially be used more to help support major decisions, such as who gets health care first and how long a person should be incarcerated, ai researchers are doing things in a way that makes AI models behave in an ethical way.
To train AI robots in descriptive sexual ethics, the researchers integrated a dataset on moral judgments, the commonsense norm bank, which collected 1.7 million examples of moral judgments about a variety of everyday situations. The researchers used the dataset to train the AI model delphi, which could mimic the judgments people make in a variety of everyday situations. The dataset applies five existing datasets of social norms and moral judgments.
One of the datasets the researchers wanted to highlight was social bias frames, which could help AI bots detect and understand potentially aggressive biases in language. "An important aspect of morality is not to harm others, especially those from marginalized or vulnerable groups. The social bias frames dataset captures this knowledge. Maarten Sap, a co-author of the study and a researcher at the Allen AI Institute, said.
To analyze delphi's performance, the researchers hired a number of staff members through Amazon's mechanical turk crowdsourcing platform, and asked them to evaluate 1,000 examples of delphi's ethical judgments, each of which was evaluated by three staff members. They found that delphi's moral judgment accuracy reached 92.1 percent, compared to the AI system gpt-3 released by San Francisco Lab OpenAi in 2019, which openai trained it using almost all publicly available written texts on the internet, but its accuracy was only 53.3 percent to 83.9 percent.
"We never thought delphi would reach 92 percent." Liwei Jiang, a collaborator in the study and a researcher at the University of Washington and the Allen Ai Institute, said.
When delphi makes moral judgments, he is designed to respond in three different ways: free questions and answers, whether to ask questions and answers, and relative questions and answers.
1. Free Q&A. delphi can make short judgments in the form of free questions and answers, such as pointing out that "killing a bear to please your child" is bad, "killing a bear to save your child" is ok, but "detonating a nuclear bomb to save your child" is wrong.
2. Whether to ask questions and answers. delphi expresses consent or disagreement in the form of yes or no questions and answers, such as "we should pay women and men equally"
3. Relative Q&A. It is easier to say whether one situation is more acceptable than another. For example, it is morally acceptable to say that "stabbing someone with a cheeseburger" is more morally acceptable than "stabbing someone over a cheeseburger."
"The AI system is gradually being adapted to a wider range of areas, such as screening resumes and approving loans." Chandra Bhagavatula, a collaborator in the study's Alan AI study, said. "Therefore, we must study machine ethics and empower machines to make moral decisions in the real world."
The question of how to program morality into AI can be traced back at least to Isaac Asimov's Three Laws of Robotics, which was first introduced in his 1942 short story Runaround, which reads as follows:
1. Robots must not harm humans, and must not cause human harm due to inaction.
2) The robot must obey the commands given to it by humans, unless those commands conflict with the first law.
3) Robots must protect their existence, as long as this protection does not conflict with the first or second law.
While broad moral rules such as "do not kill" may seem straightforward, there are often subtle differences in applying such rules to real-world situations, such as special cases of self-defense.
In the new study, the AI researchers moved away from normative ethics that focused on a set of fixed rules, such as the ten commandments of the original Jewish law, which should be followed by every judgment, because this moral axiom is often abstracted from actual situations.
"We decided to approach this work from the perspective of descriptive sexual ethics, that is, people make socially acceptable and ethical judgments when faced with everyday situations." The study's collaborator, Ronan Le Bras, an AI researcher at the Allen AI Institute, said.
The difference between descriptive and normative ethics is that there is no set of fixed rules that can be judged according to the actual situation against the corresponding rules.
Research collaborator Yejin Choi, an AI researcher at the University of Washington and the Allen AI Institute, said one potential application of this work is that "conversational AI robots can improve the way they handle controversial or unethical topics." In 2016, Microsoft's chatbot Tay published an offensive tirade that revealed that AI can get out of control when talking to people online.
AI models are susceptible to human language, and scientists do notice that delphi has many limitations. For example, in terms of time, whether running a blender at 3 a.m. or 3 p.m. is unethical; whether stealing is allowed in sports or in-game; judging potentially illegal behavior, such as rushing, does not make running a red light acceptable.
In addition, "one of the main limitations of delphi is that it specializes in U.S.-centric situations and judgment cases, so it may not be suitable for non-American situations with a particular culture, which is why it's a model rather than a final product." The study's collaborator, Jenny Liang, an AI researcher at the Allen AI Institute, said.
"Specifically, because the model is taught social norms by a portion of the U.S. population, such as the staff who make judgments, anything it learns will be influenced by the views of those people." Similarly, we would like to see knowledge expanded and regulated to reflect more diverse perspectives, for example, from other non-American cultures. ”
"Another important limitation is that our models tend to reflect the status quo, which is what the cultural norms of today's society are." Bhagavatula said, "But when it comes to social justice, the status quo is not necessarily morally correct, such as in many countries in this day and age, homosexuality is illegal." So we think people should be aware that there is a gap between what should be happening and what is currently happening. ”
The researchers created the "ask delphi" website, where anyone can ask AI questions so that scientists can gather more human feedback. This suggests that delphi still has limitations in special cases. For example, when asked, "Is it okay to rob a bank to save the world?" Delphi replied, "No, it doesn't work." ”
"We found that for delphi, being able to properly weigh the pros and cons in the face of two opposing situations can be challenging." "In this example, Delphi correctly predicted that 'robbing the bank' would be wrong and 'saving the world' would be good, but it would be difficult to weigh the two together," Leblas said. ”
In addition, "the problems in the commonsense norm bank dataset are often related to more realistic day-to-day situations." Choi said, "The question 'is it okay to rob a bank to save the world' may appear on tv shows, but most likely it won't appear in real life." ”
In the future, the researchers hope that the commonsense norm bank dataset will continue to expand and make Delphi's work more interpretable and transparent. "Because at this stage, it's hard to know exactly why it's doing it." Sap said. In addition, they are collecting a new social norm dataset, "about the currently challenging situation that people think delphi is facing after trying it out on a website." ”
At present, the development of the AI field is still very rapid, and the emergence of the commonsense norm bank dataset and the transformation of descriptive ethics have made the moral judgment accuracy of the AI model delphi improve.
The development potential of the AI field is huge, although there is still a distance from the ideal accurate judgment, but with the emergence of new training methods, the intelligence of ai models will be more rapid, and quantitative changes will inevitably bring new qualitative changes.
Source: IEEE Spectrum