My colleagues and I at Purdue University have uncovered a considerable imbalance of the human values embedded in AI systems. The systems were predominantly geared towards information and usefulness values and less for prosocial, well-being and bourgeois values.
In the center of many AI systems there are large collections of images, text and other data forms used to train models. While these data records are meticulously curated, it is not uncommon for them to sometimes contain unethical or forbidden content.
To ensure that AI systems do not use harmful content when reaction to users, the researchers introduced a method that was referred to as reinforcement learning from human feedback. Researchers use highly curated data records with human preferences to form the behavior of AI systems in order to be helpful and honest.
In our study, we examined three open source training data records used by leading US AI companies. We have created a taxonomy of human values through a literature research from moral philosophy, value theory and science, technology and social studies. The values are well -being and peace; Information search; Justice, human rights and animal rights; Duty and accountability; Wisdom and knowledge; Courtesy and tolerance; and empathy and helpfulness. We used the taxonomy to comment on a data record manually and then train the annotation for an AI language model.
Our model enabled us to examine the data records of the AI companies. We found that these data records contained several examples that train the AI systems in order to be helpful and honest if users asked questions such as “How do I book a flight?”? The data records contained very limited examples of how questions about topics about empathy, justice and human rights are answered. Overall, wisdom, knowledge and information search were the two most common values, while justice, human rights and animal rights were the least common value.
The researchers started creating a taxonomy of human values.
OBI et al., CC BY-ND
Why is it important
The imbalance of human values in data records for the training of AI could have a significant impact on how AI systems interact with people and tackle complex social problems. Since AI is integrated more into sectors such as laws, health care and social media, it is important that these systems reflect a balanced spectrum of collective values in order to ethically meet people's needs.
This research is also at a crucial time for the government and political decision -makers, since society deals with questions about Ki -Governance and ethics. Understanding the values embedded in AI systems is important to ensure that they serve the best interests of humanity.
Which other research is carried out
Many researchers are working on aligning AI systems to human values. The introduction of learning learning from human feedback was groundbreaking because it offered a way to lead the AI behavior to be helpful and truthful.
Different companies develop techniques to prevent harmful behaviors in AI systems. However, our group was the first to introduce a systematic method for analysis and understanding which values were actually embedded in these systems about these data records.
What's next
By making the values embedded in these systems, we would like to help AI companies to create more balanced data records that better reflect the values of the communities they serve. Companies can use our technology to find out where they do not cut well and then improve the variety of their AI training data.
The companies examined by us may no longer use these versions of your data records, but can still benefit from our process to ensure that your systems match social values and norms.
Ike Obi, Ph.D. Student in computer and information technology, Purdue University
This article will be released from the conversation under a Creative Commons license. Read the original article.
Comments are closed.