Google Botched Their GPT Rollout. Then They Botched the Cover-Up.

Can Google get itself more tied in knots? Maybe they should stop lying to people and just play it straight.

Google needs no introduction as a company. It goes by the legal name Alphabet these days, but the Google brand is ubiquitous – search engines, news feeds, Gmail, Chrome, YouTube, and many more apps and channels are used by billions of people every day. They are heavily involved in the world of artificial intelligence (AI) and generative pre-trained transformers (GPT) as well.

Google’s entry into the GPT competition is called Gemini (formerly Bard). Gemini recently made its debut, and it was a complete fiasco.

One user gave it a prompt to produce an image of a pope. Gemini produced a huge collection of black women, an indigenous shaman, and one black male. In 2,000 years, there have been no women, blacks, or shamans as popes. None.

Another prompt asked for images of Vikings. Gemini produced images of black Vikings. There were no black Vikings. And so on.

It turned out that the system had been programmed to erase white people from history. In this interview, Sergey Brin, co-founder of Google, said, “We definitely messed up on the image generation. I think it was mostly due to just not thorough testing.”

That’s false. Gemini had been programmed to perform exactly as it did. The problem was something called Prompt Injection.

The prompt is the request you make to the system. Injection is a euphemism for an algorithm that changes the question without informing the user. For example, if your prompt is “Show me an image of a pope,” the injection algo will change that to “Show me an image of a pope in a world of diversity.” With that prompt, getting female and indigenous pope images is no surprise.

The system is doing exactly what it was designed to do.

The problem was that the prompt itself was altered to elevate the “values” of Google engineers over hard facts and actual history. Brin knows this and is just blowing smoke.

Google is a private company, and they can design their system any way they like. That doesn’t mean we have to trust them or use their products. I’ve never trusted Google’s output because of political bias and skewed results. Hopefully, that sentiment will spread.

Corporate leaders and institutional fiduciaries looking to incorporate state of the art predictive analytics to their risk mitigation and strategic analysis should click the link to learn more about Raven Predictive Analytics®.