Recently, the OpenAI chatbot ChatGPT apologised to Microsoft CEO Satya Nadella when quizzed about the most popular south Indian tiffin options. It had incorrectly suggested biryani, alongside idli, dosa, and vada, among others. Later, it accepted the biryani goof-up. So, if you have interacted with ChatGPT multiple times, you would have at least once come across an instance when the popular chatbot apologised to you if it did not have a relevant response. 

What makes ChatGPT special is that it acknowledges its ignorance if it does not have the relevant information or response to the query asked. Empathy. It forms the core of emotional intelligence and can be explained as the ability to understand or feel what another person is experiencing within their frame of reference.

Experts stress on the power of empathy in artificial intelligent systems, if we want to move toward artificial general intelligence. And since empathy can be learned and trained, AI can surely be equipped with artificial empathy in the coming years.

And this is exactly what sets ChatGPT apart – its ability to express empathy based on context. 

ChatGPT has aced empathy

Nicole Marino, managing partner at Endurance, came across a personal and vulnerable post by someone who was hurting after losing a loved one. Marino was dismayed by the kind of responses that felt tone-deaf as they sought for more information from the person. Many times people struggle with writing empathic responses in such situations.

She tested how ChatGPT  would respond to the post and was amazed by the result. ChatGPT wrote, “I am deeply sorry for your loss and I know how difficult it can be to cope with the death of a loved one. It’s important to remember that it’s okay to feel whatever emotions come up, even though it may be difficult. Please know that I’m here for you and I’m here to listen if you need someone to talk to.” 

Mind you, ChatGPT is not the kind of over-confident, arrogant AI that Hollywood would have told you about, but a humble and human-like chatbot that acknowledges its errors and corrects them when pointed out. For instance, when Connor James asked the chatbot for a coding query, it corrected a wrong response and apologised for the same.

Moreover, ChatGPT is able to provide additional useful information in case it is not able to provide the required answer. Manan Mathur, an advocate, asked the chatbot to provide a Supreme Court judgement on the quashing of Section 439. In response,  ChatGPT apologised for not being able to browse the internet or access specific legal materials to provide the required information. However, it did provide information of the Section under the IPC and also provided a list of legal luminaries who have dealt with criminal cases.

How is ChatGPT able to do this?

ChatGPT is trained using Reinforcement Learning from Human Feedback (RLHF). RLHF uses human feedback to help ChatGPT learn the ability to follow instructions and generate responses that are satisfying to humans based on the nature of prompts.

The chatbot is based on the GPT 3.5 model that was trained using datasets from the internet that included a massive 570 GB of data sourced from books, Wikipedia, research articles, webtexts, websites and other forms of content and writings on the net. Approximately 300 billion words were fed into the system.

(Source: arxiv) The notion of empathetic conversational systems as a class of conversational artificial intelligence systems that seek to incorporate empathy. 

ChatGPT-like models are typically trained on various parameters, including emotion expressions, emojis, exemplars, communication mechanism, sentiment, dialog act/ intent, verbal and non verbal behaviours, positive and negative emotion clusters, anticipated emotion and external knowledge among a host of others.

The engineers who built ChatGPT hired contractors (called labelers) to evaluate the outputs of the two systems, GPT-3 and the new InstructGPT (ChatGPT’s ‘sibling model’). What sets ChatGPT apart from a simple chatbot is that it has been specially trained to understand human intent in a question and provide helpful, truthful and harmless answers.

Because of this training, ChatGPT may challenge certain questions and discard parts of the question that do not make sense and also acknowledge its ignorance if it does not have the required information.

The training process of ChatGPT is featured in the infographic below.

But before it gives you any response, it goes through a moderation process set in place by OpenAI that is detailed below. ChatGPT’s often apologies if the queries fall under the below mentioned categories. This includes hate, self harm, sexual, violence, etc.

With the help of empathy ingrained into ChatGPT like models using datasets can result in deep imitation learning models that can accurately and effectively provide context-aware empathetic and natural dialogue conversations. This, in turn, will help take step towards functional artificial general intelligence that can help humankind.