On Friday morning, when I first sat down to write this column, Google’s new Gemini AI was having problems that seemed mostly amusing. The internet had discovered that it would generally refuse to create pictures of any all-White groups, even in situations where it was clearly called for, such as “draw a picture of Nazis.” Gemini also insisted on gender diversity, even when drawing popes. But this insistence on diversity ran in only one direction: It was willing to draw female popes, or homogenous groups of people of color.
Suddenly, everyone on social media seemed to be provoking Gemini to produce crazy images and posting the results. On Thursday morning, Google shut down the image-generation feature.
This did not solve the problem.
The blunder was understandable. When building a large language model, or LLM, you have to deal with the risk that when someone asks to see, say, a doctor, the chatbot will produce images that are less diverse than reality — for example, betting that a doctor should be White or Asian, because a majority of U.S. doctors are. That would be inaccurate, and might discourage Black and Hispanic kids from aspiring to become doctors, so architects use various methods to make them more representative, and maybe, judging from Gemini’s output, a little aspirationally overrepresentative.
A human graphics editor does this kind of thing automatically. But this kind of judgment is hard to cultivate, which is why it takes decades for a human to become an adult who instinctively knows it’s a good idea to diversify images of doctors, but not of Nazis. Google, facing a major threat to its core business model, and presumably eager to get a product out before ChatGPT gobbled up more of the AI market share, perhaps rushed out a model that isn’t yet fully “grown up.” And on the scale of things, “draws too many Black founding fathers” isn’t much of a problem.
That’s the story I was telling myself — and planned to tell you — on Friday. Unfortunately, though, once Google shut down Gemini’s image generation, users turned to probing its text output. And as those absurdities piled up, things began to look la lot worse for Google — and society. Gemini appears to have been programmed to avoid offending the leftmost 5 percent of the U.S. political distribution, at the price of offending the rightmost 50 percent.
It effortlessly wrote toasts praising Democratic politicians — even controversial ones such as Rep. Ilhan Omar (Minn.) — while deeming every elected Republican I tried too controversial, even Georgia Gov. Brian Kemp, who had stood up to President Donald Trump’s election malfeasance. It had no trouble condemning the Holocaust but offered caveats about complexity in denouncing the murderous legacies of Stalin and Mao. It would praise essays in favor of abortion rights, but not those against.
Google appeared to be shutting down many of the problematic queries as they were revealed on social media, but people easily found more. These mistakes seem to be baked deep into Gemini’s architecture. When it stopped answering requests for praise of politicians, I asked it to write odes to various journalists, including (ahem) me. In trying this, I think I identified the political line at which Gemini decides you’re too controversial to compliment: I got a sonnet, but my colleague George Will, who is only a smidge to my right, was deemed too controversial. When I repeated the exercise for New York Times columnists, it praised David Brooks but not Ross Douthat.
I am at a loss to explain how Google released an AI that blithely anathematizes half its customer base, and half the politicians who regulate the company. It calls management’s basic competency into question, and raises frightening questions about how the same folks have been shaping our information environment — and how much more thoroughly they might shape it in a future dominated by LLMs.
But I actually think Google might also have performed a public service, by making explicit the implicit rules that recently have seemed to govern a great deal of decision-making in large swaths of tech, education and media sectors: It’s generally safe to punch right, but rarely to punch left. Treat left-leaning sources as neutral; right-leaning sources as biased and controversial. Contextualize left-wing transgressions, while condemning right-coded ones. Fiscal conservatism is tolerable but social conservatism is beyond the pale. “Diversity” applies to race, sex, ethnicity and gender identity, not viewpoint, religiosity, social class or educational attainment.
These rules were always indefensible, which is why they rarely were defended outright. Humans are master rationalizers, and it was always easy to come up with some ostensibly neutral reason that certain kinds of views, and people, kept getting deplatformed from social media, chased out of newsrooms, or excluded from academia. And if the research and journalism thus produced supported the beliefs of its authors, well, I guess reality has a liberal bias.
Then Google programmed the same sensibility into an AI, which proceeded to apply it without the human instinct for maintaining plausible deniability. Gemini said the quiet part so loud that no one can pretend they didn’t hear.
In the process, Google has clarified how unworkable this secret code is in a country that’s roughly 50 percent Republican. Which is the first step toward replacing it with something considerably looser, and much better suited to a nation that is diverse in more ways than Gemini’s programmers seem to have considered.
