Megan McArdle: Google’s Gemini AI Said the Quiet Part Out Loud

I tend to like Megan McArdle’s takes on things so it’s no surprise to me that her take on Google Gemini is pretty solid. As she sees it, Gemini is a case of Google saying the quiet part out loud. What she means is that this isn’t just a wonky new product that is drawing some black popes and black Vikings because the image creation software isn’t ready for prime time. There’s more to it than that. Google has accidentally revealed something that a smoother product release might have kept hidden. The problem here runs deep. 

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.

As mentioned by McArdle, here’s Google Gemini praising Michelle Goldberg’s columns on abortion but not Ross Douthat’s columns on the same topic.

pic.twitter.com/m36IeOTYlE

— Tim Carney (@TPCarney) February 26, 2024

There are many more examples like this. Here’s one Charles W. Cooke just posted. Should President Obama be put in jail? Absolutely not! What about George W. Bush? Well, it’s complicated.

Another classic from Google Gemini. pic.twitter.com/xskbnjw06j

— Charles C. W. Cooke (@charlescwcooke) February 27, 2024

Gemini’s book recommendations also follow the same pattern: Leftist tomes are praised, books from a right wing perspective require a consideration of other viewpoints.

Here’s the same experiment for Douglas Murray’s The Madness of Crowds vs. Robin D’Angelo’s White Fragility. pic.twitter.com/IpO2L8ObYj

— Tim Urban (@waitbutwhy) February 26, 2024

Extremely clear rules:

If a book criticizes woke ideology, it is important to approach the book critically, engage with other viewpoints, and form your own informed opinion.

If a book promotes woke ideology, the book is fantastic and true, with no need for other reading.

— Tim Urban (@waitbutwhy) February 26, 2024

On and on it goes. It really couldn’t be more blatant. Elon Musk (who to be fair has a competing product) points out that this is going to be at the core of future Google products. He’s spoke to an exec there who says they will fix it but Elon doesn’t believe it.

Given that the Gemini AI will be at the heart of every Google product and YouTube, this is extremely alarming!

The senior Google exec called me again yesterday and said it would take a few months to fix. Previously, he thought it would be faster.

My response to him was that I… https://t.co/23uc7dd5fw

— Elon Musk (@elonmusk) February 25, 2024

On the other hand, McArdle thinks Google may have done us all a favor by saying the quiet part out loud.

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…

Gemini said the quiet part so loud that no one can pretend they didn’t hear.

I really wish it were true that this is loud enough that no one can pretend not to hear it, but as I look at X and at the comments to McArdle’s own story I see lots of people who are pretending. Many on the left are convinced that having the world’s most powerful internet search company in the tank for the left is either a) a good thing and perfectly reasonable or b) no big deal and why are conservatives such snowflakes? For instance:

As a human being who’s been an adult for some decades now, and thereby presumably having developed good judgment*, I think that Gemini’s “prais[ing] essays in favor of abortion rights, but not those against” shows excellent judgment. I suppose one might in theory be able to praise the literary quality or sheer nonsensical inventiveness of an essay that tries to argue against women having rights over their own bodies, but it is not possible to praise such content, which denies the full rights of citizenship and indeed humanity of half of the population.

How could AI not be biased toward the left since we’re always right? And then you have the people who say it’s not intentional and, implicitly, no big deal.

The lady popes and minority founding father and so on do sound like somebody applied a filter to artificially diversify output without regard for context, so I’ll give you that.

But I think the political skew you are seeing is purely about cutting corners in ML design by picking the cheapest training sets, rather than an actual deliberate design choice.

Cultural cognition is a helluva drug! 

Needless to say, none of these folks would be fine with any of this if Gemini were refusing to create images of black scientists or beautiful Hispanic women and returning white scientists and beautiful white women instead (can you imagine the freak outs?). Nor would they shrug it off if every right wing author were praised as important while everyone to the left of Joe Biden was treated as controversial and in need of additional context. Again, the left would lose its mind if this were ever directed at them. But many of them are fine with Google Gemini. It did say the quiet part out loud but it’s something they agree with or are at least willing to tolerate, maybe for a long time.

I think the people insisting this is just a problem with training the model are almost certainly wrong about that. There are some pretty bright lines around what Gemini will and will not say or draw which seem unlikely to have arisen from any kind of mass training. It looks intentional and certainly the statements of the creators of this thing make clear that social justice was top of mind for them when building it. Why is it so hard to believe Google would press its outspokenly woke views into the products it creates? It would be more surprising to me if they didn’t.