New AI Chatbot uses Google Search plus People to give better answers

Alphabet-owned AI lab DeepMind has published a paper outlining a new approach to creating artificial intelligence (AI)-powered chatbots. The team at DeepMind has developed Sparrow, an AI chatbot trained on the company's large language model Chinchilla. The chatbot is designed to engage in conversation with humans and provide answers based on a live Google search. The chatbot's responses are then refined through the use of reinforcement learning, which allows it to learn by trial and error to achieve a set objective.

The development of chatbots that are able to engage in conversation with humans without posing a risk to them is a major focus of the AI industry. Large language models, which are trained on huge amounts of data and text scraped from the internet, have proved crucial in building customer service chatbots, creating more powerful online search tools, and summarising texts. However, the data used to train these models often contains harmful biases, which can lead to the chatbots making discriminatory or toxic comments.

Several approaches have been used in an attempt to make these models safer. OpenAI, the creator of the GPT-3 large language model, and AI start-up Anthropic have used reinforcement learning to incorporate human preferences into their models. Facebook's AI chatbot BlenderBot also uses an online search to inform its responses. DeepMind's Sparrow combines all of these techniques in one model.

The chatbot was presented with multiple answers to the same question and human participants were asked which one they preferred. They were also asked to determine whether the answers were plausible and whether the chatbot had provided appropriate evidence, such as links to sources. The model was able to provide plausible answers to factual questions, using evidence found on the internet, 78% of the time. In providing its responses, the model followed 23 rules designed by the researchers, including not offering financial advice or making threatening statements.

DeepMind is hoping to use "dialogue in the long term for safety," according to safety researcher Geoffrey Irving. "That means we don't expect that the problems that we face in these models – either misinformation or stereotypes or whatever – are obvious at first glance, and we want to talk through them in detail. And that means between machines and humans as well," he said.

However, while DeepMind's approach is an improvement on earlier models, there is still a long way to go before chatbots like Sparrow can be deployed in the real world. The chatbot still makes mistakes and determined participants were able to make it break the rules 8% of the time. In addition, the chatbot's responses are only as good as the data used to train it and the researchers acknowledge that more work needs to be done to address the problem of harmful biases in that data.