Report: Google asking employees to help train Bard by rewriting wrong answers

Ben Schoon | Feb 16 2023 - 1:45 pm PT

As Google prepares to launch Bard to the general public, the company is asking its employees to assist in training the AI by rewriting its bad answers.

CNBC reports that Google’s VP of Search, Prabhakar Raghavan, sent an email to employees earlier this week asking for help in training Bard by correcting incorrect or bad responses. The email included a page of do’s and don’ts for how employees should correct Bard, and also incentivized employees to write the corrections with a “Moma badge” which would appear on internal employee profiles.

The do’s and don’t document requests that staffers interact with Bard to help test its capacity, as well as speed up training of the model. On topics that employees “understand well,” the document requests that employees rewrite answers that are incorrect.

The document says in part:

Bard learns best by example, so taking the time to rewrite a response thoughtfully will go a long way in helping us to improve the [model]

It goes on to add that when employees rewrite answers, they should do so in a “polite, casual and approachable,” while keeping an “unopinionated, neutral tone,” and writing in first person. The document further says that responses should not stereotype and “avoid making presumptions based on race, nationality, gender, age, religion, sexual orientation, political ideology, location, or similar categories.” Answers also shouldn’t “imply emotion” or “claim to have human-like experiences.”

When it comes to more sensitive topics, such as “legal, medical, [or] financial advice” or if Bard responds with something hateful or abusive, employees are told to give a “thumbs down,” presumably referring to the UI Google previewed previously, instead of rewriting the answers.

The “top 10” contributors from the Knowledge and Information organization would also get the chance to share feedback directly with Raghavan according to the email.

Raghavan said:

This is exciting technology but still in its early days. We feel a great responsibility to get it right, and your participation in the dogfood will help accelerate the model’s training and test its load capacity (Not to mention, trying out Bard is actually quite fun!).

This comes as many have found over the past week that Microsoft’s Bing AI has a tendency to go off the rails after a certain point, often making replies that go directly against the sort of things Google seems to be trying to avoid in this training.

It also comes following a negative sentiment has come from Bard’s reveal. Google’s announcement was largely considered underwhelming, and Google employees reportedly called the launch “botched” and “rushed,” while directly criticizing CEO Sundar Pichai and Google management. An error in the early previews of Bard also contributed to a huge stock dip for Alphabet.