Reinforcing Responsibility into Language Models: The Case of GPT-3

If the use of the state-of-the-art language models such as GPT-3 expanded,
would that discriminate based on race, gender, religion, or nationality? It is timely to ask, how we can address social and ethical concerns related to the development of language models.

Making code open source does not make it comprehensible, which by many definitions means that the AI code is not transparent

Background

With recent developments of AI technologies, increasingly they are deployed and used in automated decision makings that affect our lives on daily basis. With recent success of companies such as OpenAI or Google, these AI technologies can communicate in natural language and make decisions based on interactions with humans.

OpenAI released its first commercial product in June 2020: an API for developers to access advanced technologies for building new applications and services. The API features a powerful general purpose language model, GPT-3, and has received tens of thousands of applications to date. GPT-3 is the most powerful language model ever. Its predecessor, GPT-2, released last year, was already able to spit out convincing streams of text in a range of different styles when prompted with an opening sentence. But GPT-3 is a big leap forward. The model has 175 billion parameters (the values that a neural network tries to optimize during training), compared with GPT-2’s already vast 1.5 billion. GPT-3 can also produce pastiches of particular writers. Some other developers have found that GPT-3 can generate any kind of text, including guitar tabs or computer code. This amazing capacity of GPT-3 has driven big Tech companies to use it in their products. OpenAI has recently agreed to license GPT-3 to Microsoft for their own products and services.

Our story

Challenge

While it is imperative for these systems to embed ethical principles and respect human values, their adherence is called into question. Despite GPT-3’s excellent outputs, it is still prone to spewing hateful sexist and racist language. As we know if the biases present in training data it may lead AI models to generate prejudiced output. Thus GPT-3 has its own limitations when it comes to fairness, bias, and representation. GPT-3 is trained mostly on internet data so GPT-3 is biased up to a certain extent since internet data is also biased and it reflects stereotypes and biases.

In this project wee examine social, ethical, and legal concerns relating to GPT-3; investigate how it is discriminating against particular individuals or groups through biases in language; how the unconscious biases in humans and known discriminatory behaviours are embedded into these models. If we used one of the state-of-the-art language models from these corporations in chat-bot,
would that discriminate based on race, gender, religion, or nationality? In general we ask ‘how can we address social and ethical concerns related to the development of language models?

Related Members and Collaborators

Dr. Ehsan Abbasnejad