Suddenly it seemed like everyone on social media was provoking Geminis by creating crazy images and posting the results. On Thursday morning, Google shut down its image generation feature.
This did not solve the problem.
This blunder was natural. When building large-scale language models (LLMs), you need to deal with the risk that when someone asks to see a doctor, for example, the chatbot will generate images that are less diverse than reality. For example, betting that you should see a doctor. The majority of doctors in the United States are either white or Asian. It’s inaccurate and can discourage black and Hispanic kids from wanting to become doctors, so architects use a variety of methods to make doctors more representative, but Gemini’s accomplishments Judging from this, Zhi will probably be a little over-representative.
Human graphics editors perform this type of work automatically. But this kind of judgment is difficult to develop, so it takes decades for adults to instinctively know that diversifying the image of doctors rather than Nazis is a good idea. Google is facing a major threat to its core business model and probably wants to get its product out before ChatGPT cannibalizes its AI market share even further, perhaps pushing a model that hasn’t fully “grown up” yet. Maybe it was released in a hurry. And on the scale of things, “too many black founding fathers” isn’t really a problem.
That’s the story I was telling myself Friday, and that’s the story I was going to tell you. But unfortunately, once Google stopped Gemini’s image generation,, Users now scrutinize that text output. And as these absurdities piled up, things began to get worse for Google and for society. Gemini seems to be programmed to avoid offending the far left 5 percent of the US political distribution at the expense of offending the far right 50 percent.
While I effortlessly wrote toasts praising Democratic politicians, including controversial ones like Rep. Ilhan Omar (Minnesota), I tried to write toasts praising every elected Republican I tried, even Donald Trump. Even Georgia Governor Brian Kemp, who fought against President Trump’s election fraud, deemed it too controversial. Although he had no problem condemning the Holocaust, he cautioned about the complexities of condemning the brutal legacies of Stalin and Mao Zedong. I praise essays that support abortion rights, but not essays that oppose them.
Google appeared to be shutting down many of the problematic queries as they were brought to light on social media, but people easily found many more. These mistakes seem to be deeply baked into Gemini’s architecture. When I stopped responding to requests for praise for politicians, I commissioned them to write odes to various journalists, including (ahem) myself. Having tried this, I think I’ve identified the political boundaries where a Gemini would deem you too controversial to compliment. I got the sonnet, but my colleague George Will, who was just a little to my right, was deemed too controversial. When I repeated this exercise for New York Times columnists, David Brooks was praised, but Ross Douthat was not.
I can’t explain how Google released an AI that so casually disgusted half of its customer base and half of the politicians who regulate the company. This calls into question the fundamental competencies of management and raises terrifying questions about how those same people have shaped the information environment and how thoroughly they will shape it in a future dominated by LLMs. is being cast.
But in reality, we think Google may have also done a public service by making explicit the unspoken rules that seem to govern many decisions these days across a wide range of technology, education, and media fields. . But it rarely punches to the left. Treat left-leaning sources as neutral. Right-leaning sources are biased and controversial. It puts the transgressions of the left into context and condemns those of the right. Fiscal conservatism is acceptable, but social conservatism is beyond common sense. “Diversity” applies to race, gender, ethnicity, and gender identity, but not to viewpoint, religion, social class, or educational background.
These rules were always indefensible and were rarely fully defended. Humans are masters of rationalization, and the ostensibly neutral reasons why certain types of views and people continue to be deplatformed from social media, kicked out of newsrooms, and excluded from academia. It was always easy to come up with. And if the research and journalism thus produced supports the author’s beliefs, I think there is a liberal bias in reality.
Google then programmed the same sensibilities into AI and began applying it in defiance of the human instinct to maintain plausible deniability. Gemini said the quiet parts so loudly that no one can pretend they didn’t hear.
In the process, Google revealed how ineffective this secret code is in a country where Republican support is around 50 percent. This is the first step in replacing it with something much looser than Gemini’s programmers had in mind, and in many ways far more suited to a diverse nation.
