In theory, AI exists largely to complement and aid humans in their measly tasks while this giant rock on which they’re stuck hurls through empty space. Google appears to be subverting this paradigm by asking employees to amend AI errors in order to improve the performance of their own AI chatbot initiative, Bard.
Bard is Google’s version of ChatGPT. Much like the latter, Bard has its fair share of issues, but Google’s approach to fixing those issues is quite intriguing.
Search Engine Land reports that, in a memo to employees, Google provided a list of ways in which employees should address the process of “teaching” Bard. The list includes reminders to “Keep responses ‘polite, casual and approachable’”, use first person language when responding, and prioritize an “unopinionated, neutral tone” when giving feedback for incorrect answers.
According to 9to5Google, the memo also posits that “Bard learns best by example” and suggests that employees “[take] the time to rewrite a response thoughtfully” in order to best improve the AI’s bank of reasonable responses.
Employees are also asked specifically to avoid “[describing] Bard as a person, [implying] emotion, or [claiming] to have human-like experiences” in their corrections.
All of this implies a heavy workload for Google employees, but the same memo also recommends a much faster approach in certain circumstances. If Bard ever returns a discriminatory response–perhaps one that “[makes] presumptions based on race, nationality, gender, age, religion, sexual orientation, political ideology, location, or similar categories”–or gives inaccurate medical advice, employees need only indicate a “thumbs-down”.
Prabhakar Raghavan, Google’s VP of Search and progenitor of the memo asking for help training Bard, acknowledges that these improvement efforts signal that the proprietary AI chatbot has a substantial amount of room for improvement.
“This is exciting technology but still in its early days,” says Raghavan. “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.”
These corrections come at a pivotal time for AI, one in which ChatGPT is more or less dominating the digital landscape while virtually every technological titan scrambles to make up for lost ground.
Jack Lloyd has a BA in Creative Writing from Forest Grove's Pacific University; he spends his writing days using his degree to pursue semicolons, freelance writing and editing, oxford commas, and enough coffee to kill a bear. His infatuation with rain is matched only by his dry sense of humor.
