The controversy centered around some users' attempts to generate images of historical figures. Gemini ended up generating more than once historical images that were inaccurately diverse. There was an image of the US Founding Fathers that was ethnically diverse. And even an image of Nazis that featured many different races.
Google promptly apologized. Google says the issue was due to major engineering challenges that it needs to solve better. But critics say the real problem is that the company is forcing its views on users.
What's really going on here?
On Episode 85 of The Artificial Intelligence Show, Marketing AI Institute founder and CEO Paul Roetzer broke it down for me.
AI models, like Gemini, learn from data found on the internet. Unfortunately, the internet is a vast repository of human bias. Even with the best intentions, engineers can unknowingly train their models to reflect existing prejudices.
This is not only a Google problem. It’s a problem faced by every company that builds AI.
"This is widely known to be a major challenge with these models," says Roetzer.
In Google’s case, its attempts to counteract the bias learned by the model appears to have had unintended consequences. Essentially, the model ended up trying too hard and sometimes generating images that defied historical fact.
Unfortunately, the issue became politicized almost immediately. (People like Elon Musk jumped into fray, claiming the tool was biased.)
Musk and allies are of the opinion that Google did this on purpose and don't want to fix this. They believe that, unless Google clears house, their culture won't change.
(They don't bother, however, to address the fact every AI company has had these issues.)
There’s no question that this is an important issue, says Roetzer. But the politicization around it misses the actual problem to be solved.
"This is a very challenging technical issue," says Roetzer. And it's one with important social impacts:
"The more people rely on these models for their information, we're going to start having a generation that learns what is true and what is not based on what these models output.”
There's no single "fix" for AI bias, either bias it learned online or bias from a particular company’s culture. Guardrails designed to protect users can also create new points of contention. The very concept of bias is subjective—what one person finds unfair, another may claim as necessary.
The only real way to solve this, according to Roetzer, is the path OpenAI has hinted about in the past:
Allowing users to control the bias settings of AI models. That would allow you to “set the temperature” of any AI tool you use to your own personal preferences. This puts the responsibility for what you get from AI models into your own hands—and takes it out of the hands of AI companies.
"I think that's the only way to solve this: let individuals choose what sort of experience they want through settings," says Roetzer. "Otherwise, we're going to keep having these arguments all the time."