Google employees asked to fix Bard’s inaccurate responses

Google’s management has asked staff to participate in testing and correct Bard’s inaccurate responses generated by the Bard conversational AI. CNBC reports.

According to the publication, Prabhakar Ragavan, the company’s vice president of search, sent staff a corresponding letter. The executive asked them to help ensure that its new competitor to ChatGPT provides correct answers.

The email contained a link to a page with instructions on how to fix responses during internal Bard testing. Staff were advised to rewrite answers on topics they are knowledgeable about.

“Bard learns best from example, […] if you spend time on thoughtful corrections, this will help us improve the product”, — the document states.

Earlier reports indicated that Google CEO Sundar Pichai asked staff to devote two to four hours of their time to the chatbot. He acknowledged, “this will be a long road.”

Ragavan echoed the CEO’s view.

“We feel a strong responsibility to get this right. Your participation in testing will help accelerate the model’s training and test its throughput (not to mention that trying Bard is actually quite fun!),” wrote the vice president.

Under Google’s rules, staff must write “polite, informal and accessible.” Responses must be “first-person” and maintain “an unbiased, neutral tone.”

Testers must not create stereotypes. They must “avoid assumptions based on race, nationality, gender, age, religion or similar categories.”

“Do not describe Bard as a personality, do not hint at emotions and do not claim that it has experiences similar to human ones,” the document says.

The leadership also instructed staff to give a “thumbs-down” to responses offering “legal, medical, financial advice.” The same rule applies to responses that incite hate or are insulting.

“Don’t try to rewrite this; our team will deal with it,” the document says.

According to Ragavan, to encourage staff to test Bard and leave feedback, participants will receive a “Moma” badge displayed in corporate profiles.

The ten best testers will be invited to a meeting with the developers, where they can share their experiences of interacting with the chatbot.

“Heartfelt thanks to the teams that have worked hard on this behind the scenes,” Ragavan wrote.

The company did not comment on the letter.

Earlier this year, Google unveiled the Bard conversational AI, which it plans to integrate into its search service.

Within a day, users noticed the chatbot’s errors in the promotional video. As a result, Alphabet lost $100 billion of market value.