The Google Gemini Debacle Shows Us Why Office Politics Matters More Than Ever

On the Real Alignment Problem

This is what happened when Google Gemini was asked to generate a picture of a Nazi soldier. The oddest thing about it might be that it can’t do a swastika.

Catch-up service:
On Macbeth (brilliant comments on this one)
Hitler’s Willing Executioners
How To Fix DEI
On the Gender Divide In Politics

People say Big Tech is evil, but I think Google has been admirably public-spirited in providing the world with a laugh during these dark times. In case you missed this story, I’ll summarise. Last week, Google released Gemini, an AI chatbot and image-generator intended to compete with similar apps from Open AI and others. It could hardly have been a more important launch, and it could hardly have gone more wrong.

As soon as people started playing with it, they noticed something odd. When they asked it for pictures of historical figures, it served up ‘diverse’ imagery regardless of context. Here’s what happened when it was asked for an image of a Viking (the whitest people in history):

Can I get an image of a pope?

How about the Founding Fathers?

Most bizarrely of all, when Gemini was asked for pictures of WWII Nazis, it served up a beautiful rainbow of ethnicities and genders including the gentleman above.

These images weren’t generated by a few mischief-makers fiddling with prompts; they came up again and again in response to standard questions. The problem wasn’t just restricted to image generation, and it wasn’t just about diversity. Gemini had a very distinct worldview (I’m using the past tense because, in response to the furore, Google has paused its image generation tool and neutered its chatbot). I guess you could call it woke, but that doesn’t quite convey how extreme it was, or how silly. It’s more like someone performing a crude parody of woke:

It may be that these examples are conflating different flaws in the system - some to do with bias, others to do with a general wariness about answering moral or political questions directly. And to be fair, this kind of question is not the typical use case for a chatbot, although by all accounts Gemini doesn’t seem to have been very good at answering more common questions either. Even so, when a rocket explodes on take-off, it’s reasonable to ask questions about what went wrong, starting with the most obvious one: what the hell did Google think they were doing?

How did they end up releasing a product that is so obviously faulty? The company’s initial explanation for the ‘diverse imagery’ problem is that it was just a bug, rather than a feature of the system. That seems disingenuous. This product would have been relentlessly tested and tuned before being unleashed on the world. Gemini’s quirks seem more likely to have been the output of a corporate culture that doesn’t realise how weird it is.

In my recent post on how to fix DEI I suggested that organisations make an effort to understand how the cultural-political worldview of their staff compares to their median user (or voter). It’s not that all organisations should try and be a mirror of the public, it’s that, in a highly politicised environment, they should be self-aware enough to know how the profile of their staff differs from the profile of the people they serve.

The anthropologist Joseph Henrich famously reframed our supposedly neutral, objective Western worldview as a WEIRD one (Western, Educated, Industrialised, Rich, Democratic). His point wasn’t that WEIRD is bad, just that it’s, well, weird; shared by only a minority of the global population. For an individual or an organisation, it’s not necessarily a good thing to be normal, but it’s nearly always a good thing to know when you’re not normal. To adapt the poker truism, if you look around the table and you can’t see who the weirdo is, it’s you.

Last week was, or ought to have been, a moment of self-insight for Google.¹ As Nate Silver puts it, Gemini displayed “the politics of the median voter of the San Francisco Board of Supervisors” - i.e. it behaves like a left-wing outlier even versus America’s educated and relatively liberal classes. If Gemini does indeed reflect the internal culture of Google, a company which serves the whole of the world, then the problem for Google goes way beyond the launch of Gemini.

When Google was only serving us information from other websites, the political outlook of its staff was less of an issue. The genius of the PageRank-based search engine is that it merely shows us what other people are looking at (in essence). As Paul Graham says, it’s just math. Google Search is more like a librarian, telling us where to go to get what we’re looking for, rather than an author. We don’t expect the librarian to be a font of knowledge or wisdom (even if many of them are).

An AI app is a very different proposition to a search engine. It doesn’t just link to external information sources. It gives us information and opinions and pictures directly (even though in reality the app is regurgitating the internet; it’s a librarian disguised as a guru). The credit and the blame for Gemini’s output therefore goes to its creator - to the company behind it.

Google is intensely aware of this, of course. That’s why, when it was tuning the model behind Gemini, it clearly went to great effort to eliminate any output that might be construed as racist, misogynist, speciesist and so on. But it over-compensated. It only worried about outputs that a left-wing cultural progressive might find offensive. It didn’t worry about - it probably didn’t even notice - outputs that seemed offensive or odd or downright wrong to anyone who isn’t an inhabitant of that particular cultural niche.

The Gemini debacle shows us that for the companies behind AI systems, understanding the cultural-political worldview of their staff, and how it differs from the median user, is crucial. It also illustrates the difficulty of trying to imbue an AI with the right ‘values’.

So far, tech companies have defined “the alignment problem” as the problem of aligning the values of AIs with the values of humanity. But as Joseph Heinrich or Jonathan Haidt will tell you, there vanishingly few values we all hold in common. What we have are different clusters of values, with great variation between nations and (at least in democratic countries) within them. Perhaps AI companies should accept this and be transparent about their own cultural politics. At the moment they all cluster together, though there is already some variation, according to David Rozado, who studies such things:

Racism and misogyny and inequality and other ills are not like bugs that you can just fix with few lines of code, leaving only the Pure Human Value good stuff. Each of these problems is complex and contested and woven into intricate social systems. Addressing them involves difficult choices and different interests, not all of which can be reconciled. Since there is no optimal solution to the problem of living together (sorry Plato, sorry Marx) there is no AI that is going to please everyone in the world.

Perhaps AI companies need to choose where they stand in relation to their median user, however defined, and stand by it. I can see a future in which different AI brands cater to different worldviews, as media brands do now. I can also see a future in which a single AI offers different cultural-political options. As David Rozado suggests, instead of a one-size-fits-all set of values, an AI might give the user a clear way to decide what kind of answer they want; how much empirical accuracy versus how much normative value.

Whatever the future looks like, a crucial question for AI companies, as it is for many other kinds of organisation, is how to keep their internal political culture in touch with the median user. The real alignment problem is the human one.

This post is free to read, so if you enjoyed it, please share and recommend. And if you’re a free subscriber, please consider taking the paid option - it’s what makes The Ruffian possible. Thanks to all of you who have already done so.

1. I’m using ‘Google’ as shorthand when perhaps I should be more specific and say ‘the team behind Gemini’, or at least, whoever signed off on the model. It’s not like Google’s staff are all left-wing agitators. I’m sure that, as this tweeter says, many of them are fairly apolitical and sceptical of activist cultural politics, but just keep quiet. What that means, of course, is that those who shout loudest wield outsized influence. It’s a challenge for leaders of all organisations to try and hear the voices of those who don’t enjoy speaking up.