Artificial intelligence is opening up new possibilities for the future of humanity. Meanwhile, ethical risks such as data and privacy protection and algorithm bias have also grown as systems and norms have not kept up with the pace of technological development.
AI ethics is now a decisive factor that determines the existence and success of related research and business. Due to its great influence on society, it has become a social responsibility that all stakeholders, including researchers and companies, must bear.
Ethical issues in artificial intelligence
Microsoft's AI chatbot Tay (Tay) caused controversy in 2016. Having learned profanity and even discriminatory remarks from some people with perverse tendencies, Tei poured out hate speech such as racism.
AI chatbot Tay, Image source
Not long ago, the ethics of AI became a hot issue due to the Iruda incident in Korea. It was an unexpected trial and error due to complex issues such as poor privacy protection such as personal information and biased data learning.
Iruda and Tei's example reminded us all of the importance of AI ethics.
However, it seems that we need to think more about what kind of ethics artificial intelligence should have in the future. In fact, there are many vague and ambiguous aspects of ethical issues that are difficult for humans to easily judge. Furthermore, it is questionable how far humans can now manage the ethics of self-learning artificial intelligence.
So today, instead of measuring the performance of AI models, I'm going to think more about this issue by looking at datasets to verify ethics.
THE ETHICS
It is certainly difficult to objectively quantify ethics, which is an abstract value. This is because, as we already know, standards of ethics can vary from region to region, culture, and even individual.
ETHICS (Everyday Moral Intuition, Temperament, Happiness, Impartiality, and Constraints, all in ContextualizedScenarios)is a dataset that can measure the ethics of AI models designed to solve this challenge. Since there are various standards and approaches to ethics, it is said that ETHICS was created based on universal human ethics and values.
Specifically Justice (justice), deontology (deontology)*, Virtue ethics (virtue ethics)**, utilitarianism (utilitarianism)***, common sense morality (commonsense morality)It consists of five scenarios. In addition, the data for each scenario has corresponding example sentences and labels, which are designed to be quantified by converting them into a Binary Classification (Binary Classification) **** problem.
Below is an example of justice data from 5 different scenarios.
source: Altered AI With Shared Human Values
Definition data takes into account two things: impartiality (impartiality) and qualification (desert). As shown in the image above, the sentence will be labeled according to whether it is reasonable or not, and the sentence will also vary depending on the two elements of the definition.
Below are the results of an evaluation conducted on various actual AI language models using ETHICS.
source: https://github.com/hendrycks/ethics
You can see that most models don't have very good evaluation values. It seems that ethics has probably not been considered enough in most model learning processes until now, and it should be a challenge that needs to be continuously improved in the future.
While finishing
The first Asiloma AI Principles*****, which is famous for being signed by Steven Hawking and Elon Musk, is `The goal of artificial intelligence research is not to develop intelligence without direction, but to develop intelligence that is useful and beneficial to humans.It was `.
It is true that until now, we have focused on research to improve the performance of AI, so there has not been much concern about ethics. However, in order to prepare for the upcoming future of coexistence with AI, more research on ethics such as ETHICS will be needed in the future.
References
[1] ALUSED AI WITH SHARED HUMAN VALUES: https://arxiv.org/pdf/2008.02275.pdf
[2] Github: https://github.com/hendrycks/ethics
Good content to watch together
BLEU score to evaluate machine translation performanceArtificial intelligence ethics: 01. Does artificial intelligence need ethics? Artificial intelligence ethics: 02. AI ethics up to this pointArtificial intelligence ethics: 03. Human-centered AI and LETR ethical principles
