Nvidia's chips have helped companies like Microsoft add
human-like chat features to search engines like Bing. But the chatbots can
still be unpredictable and say things their creators wish they did not.
Microsoft in February limited users to five questions per
session with its Bing search engine after the New York Times reported the
system gave unsettling responses during long conversations.
Nvidia's software tools, provided free of charge, are
designed to help companies guard against unwanted responses from chatbots. Some
of those uses are straightforward - the maker of a customer service chatbot
might not want the system to mention products from its competitors.
But the Nvidia tools are also designed to help AI system
creators put into place important safety measures, such as ensuring that
chatbots do not respond with potentially dangerous information such as how to
create weapons or send users to unknown links that could contain computer
viruses.
US lawmakers have called for regulations around AI systems
as apps like ChatGPT have surged in popularity. Few legal rules or industry
standards exist on how to make AI systems safe.
Jonathan Cohen, vice president of applied research at
Nvidia, said the company aims to provide tools to put those standards into
software code if and when they do arrive, whether through industry consensus or
regulation.
"I think it's difficult to talk about standards if you
don't have a way to implement them," he said. "If standards emerge,
then there'll be good place to put them." © Reuters