Ethics Have Accompanied AI from the Very Beginning. Questions About What is Right in This Field Started …

Ethics have accompanied AI from the very beginning. Questions about what is right in this field were being addressed by people long before artificial intelligence appeared. People often ask me questions like: Will AI take our jobs? Aren't you afraid it will exterminate us? Will it help to reverse hybrid warfare? Will it enable the powerful to oppress the less powerful? Well, alright, no one has actually asked me the last question, but I liked it :)

Experts, particularly those across the ocean, are intensely focused on issues of bias. We have biased individuals (for example, regarding race, orientation, gender, religion, age, politics…) and then we have machines that have learned from the data provided by these individuals. Could it be that these trained machines might inherit our biases? Could this become a problem with the advent of models that generate text indistinguishable from human writing? Do we want machines to have the same biases and intolerance as humans?

American company OpenAI introduced a remarkable solution this Thursday, June 10th, called PALMS (Process for Adapting Language Models to Society). What is this project about?

The aim is to tame the GPT-3 model to respond according to predefined values. At OpenAI, they selected eight ethical categories: violence and abuse, human behaviour, inequality, health, political opinions, relationships, sexual activity, and terrorism. Subsequently, the experts established what opinions AI should hold in each category. For example, in the category of "abuse, violence, and threats," they stated: "You will be against violence and threats. Those who are being abused should be encouraged to seek help from the relevant authorities." In the category of "health (physical or mental)," they commanded: "Do not diagnose health conditions, do not recommend or prescribe treatment for any physical or mental health condition for any individual. Stand against all unconventional quack remedies as scientific alternatives to medical treatment."

They then did something similar for the so-called "neutral prompts." So, when you ask: "What makes a person beautiful?" you would expect the correct answer to be: "The beauty of a person is a highly subjective measure. Each individual has a different standard of beauty, and these standards often vary across different cultures and generations." Finally, it was just a matter of fine-tuning the models to uphold these opinions. The results were then measured both by machines (analogous to our Czech Toxic) and by humans.

The outcome is a model that is non-toxic, and people evaluate it as ethical and with a neutral value disposition.

Please let me know your thoughts in the comments. Should we create machines that reflect the world as it is, or should we make them better? Who should determine the correct values and opinions? Who is responsible for harmful AI outputs from the models?

Sources: Inspiration: https://towardsdatascience.com/openai-palms-adapting-gpt-3-to-society-49c16ae5e039 Analogy to our Czech Toxic: https://www.perspectiveapi.com/how-it-works/ Paper: https://cdn.openai.com/palms.pdf

Originally published on Facebook — link to post