Google Employees Rewrite Bard's Bad Responses

Google executives recognize that Bard, the company's artificial intelligence search tool, isn't always accurate in its responses. Some of the responsibility for fixing the wrong answers falls on employees.

‍

On Wednesday, Google's vice president for search, Prabhakar Raghavan, asked staffers to make sure ChatGPT gets answers right. In the email, which Trade Algo viewed, employees were instructed to fix responses as they tested Bard internally by visiting a dos and don'ts page.

‍

The staff is encouraged to rewrite answers on topics they are familiar with.

‍

“To improve the mode, we need to take the time to rewrite a response thoughtfully, since Bard learns best by example,” the document says.

‍

As Trade Algo reported earlier, CEO Sundar Pichai asked employees to devote two to four hours to Bard, stating “this will be a long journey for everyone, across the field.” 

‍

This sentiment was echoed by Raghavan.

‍

“This is an exciting technology, but it is still in its infancy,” Raghavan wrote. “We feel a great responsibility to get it right, and your participation will help speed up the model's training and test its load capacity.”

‍

Last week, Google unveiled its conversation technology, but a series of missteps caused the stock price to fall nearly 9%. The rollout internally was criticized by employees as "rushed," "botched," and "comically short-sighted."

‍

To fix AI errors, company leaders are relying on human knowledge. Google provides guidance about what to consider "before teaching Bard" at the top of the section on dos and don'ts.

‍

Google instructs employees to keep responses "polite, casual, and approachable." It also urges them to use the first person, and maintain an "unbiased, neutral tone.”

‍

As for don'ts, employees are told to "avoid making presumptions based on race, nationality, gender, age, religion, sexual orientation, political ideology, location, or similar categories.”

‍

The document also warns against describing Bard as a person, suggesting emotions, or implying human-like experiences.

‍

After that, Google instructs employees to give a "thumbs down" to answers that give "legal, medical, financial advice" or are hateful and abusive.

‍

“Don't try to rewrite it; our team will handle it from there,” the document states.

‍

As a way to motivate people within his organization to test Bard and provide feedback, Raghavan says contributors will earn a "Moma badge," which appears on profiles within his organization. Google is going to invite the top 10 contributors to the rewrite process from the Knowledge and Information organization, which Raghavan oversees, to a listening session, which will take place later this month. It is from there that they have the opportunity to "share their feedback live" with Raghavan and others working on Bard.

‍

“ I would like to thank the teams who have been working hard behind the scenes on this,” Raghavan wrote.

‍

The company did not immediately respond to a request for comment.

‍