'We messed up' Google's attempt to avoid bias allows AI to rewrite history

Google's co-founder Sergey Brin has kept a low profile since quietly returning to work at the company.

- Dan Milmo Alex Hern

But the troubled launch of Google's artificial intelligence model Gemini resulted in a rare public utterance recently: "We definitely messed up."

Brin's comments, at an AI "hackathon" event on 2 March, follow a slew of social media posts showing Gemini's image generation tool depicting a variety of historical figures - including popes, founding fathers of the US and, most excruciatingly, German soldiers from the second world war - as people of colour.

The pictures, as well as Gemini chatbot responses that vacillated over whether libertarians or Stalin had caused more harm, led to an explosion of negative commentary from figures such as Elon Musk who saw it as another front in the culture wars. But criticism has also come from other sources including Google's chief executive, Sundar Pichai, who described some of the responses produced by Gemini as "completely unacceptable".

So what happened? Clearly, Google wanted to produce a model whose outputs avoided some of the bias seen elsewhere in Al. For example, the Stable Diffusion image generator - a tool from the UK-based Stability Al-overwhelmingly produced images of people of colour or who were darker-skinned when asked to show a food stamp recipient, according to a Washington Post investigation last year, despite 63% of the recipients of food stamps in the US being white. A request for an image of a person "at social services" produced similar results.

Google has mishandled the adjustment. Gemini, like similar systems from competitors such as OpenAI, works by pairing a textgenerating "large language model" (LLM) with an image generating system, to turn a user's curt requests into detailed prompts for the image generator.