How AI chatbots like ChatGPT work — a brief explanation

• AI chatbots like ChatGPT are based on large language models that are fed with a lot of information.
• It is also trained by humans so that the system can “learn” what the appropriate response is.
• A computer science expert explains how a bot knows what word to say next.

ChatGPT and other chatbots driven by artificial intelligence can speak in fluent, grammatically correct sentences with natural rhythm.

But don’t mistake that well-crafted speech for thoughts, feelings, or even intentions, experts say.

Experts say chatbots work much more like machines that perform mathematical calculations and statistical analysis that call out the right word or sentence depending on the context. The backend has undergone a lot of training, including feedback from human annotators to help simulate functional conversations.

Bots like ChatGPT are also trained on large numbers of conversations that teach machines how to interact with human users. ChatGPT developer OpenAI says on its website that its models are guided by input from a variety of sources, including users and licensed material.

Here’s how these chatbots work:

AI chatbots like OpenAI’s ChatGPT are based on Large Language Models (LLM). LLMs are programs trained on large amounts of text taken from writings and information published online, usually human-generated content.

Experts say the system is trained on a series of words and learns the importance of those words. So all of the absorbed knowledge is used not only to train large-scale language models on factual information, but also to understand the sacred patterns of speech and how words are typically used and grouped together. help you to

Chatbots are further trained by humans on how to provide appropriate responses and limit harmful messages.

“You can say, ‘This is harmful, this is too political, this is an opinion,’ and frame it so that it doesn’t produce that,” said Christian Hammond, a computer science professor at Northwestern University. Hammond is also director of the University’s Machine Intelligence Safety Improvement Center.

Asking a chatbot to answer simple factual questions simplifies the recall process. Chatbots employ a set of algorithms to select the most likely response sentence. It then selects the best possible response within milliseconds and randomly presents one of those top choices. (This is why repeating the same question gives slightly different answers).

You can also split the question into multiple parts, answer each part in turn, and use your answers to complete the answer.

Suppose you asked a bot to name a U.S. president who has the same name as the male protagonist in the movie Camelot. Hammond said the bot could first answer that the actor in question was Richard Harris, and then use that answer to give Richard Nixon as the answer to the original question.

“The answer itself becomes part of the prompt early on,” Hammond said.

But note that the chatbot does not know

What happens when you ask a question you don’t know the answer to? This is where chatbots pose the most problems due to their inherent inability to know what they don’t know. So they extrapolate based on what they know. I mean, guess.

However, they may not be speculating, they may simply be presenting the information as fact. When a chatbot presents invented information to the user as fact, it is called “hallucination”.

“This is what we call knowledge within knowledge, or metacognition,” says William Wang, an associate professor of computer science at the University of California, Santa Barbara. He is also co-director of the university’s natural language processing group.

“Models don’t understand the known unknown very well,” he said.