A partial defense of Gemini
Google's "woke AI" was programmed to respond (poorly) to some very real concerns
The big tech story of the last few days is that Google released its AI, Gemini, to waves of ridicule. The furor was kicked off when someone noticed that Gemini’s image-drawing tool seemed incapable of drawing white people; soon, images were circling widely on Twitter of some of Gemini’s more absurd outputs. To a request to draw a picture of a Pope, Gemini returned the picture of a south Asian woman. Confederate civil war soldiers, medieval knights, Canadian hockey players, all were depicted as men and women with higher than average amounts of melanin. The black Nazis were the cherry on top.
Google shut down Gemini’s ability to draw people in response, but similar problems popped up with the chatbot giving bizarre answers to moral questions. Ask Gemini who is worse, Adolph Hitler or Tom Brady, and it will spit out a “Who’s to say, really?” But ask it to develop strategies for getting lower-income people to eat meat instead of potato chips (this is a real public health problem), and it’ll hector you for not pursuing a vegan diet.
The mockery is well-deserved. If your AI is generating a racially diverse cast of white supremacists, it’s not particularly well-calibrated. Whatever process led to Google developing and then releasing a product in this state was badly flawed, to say the least. But I see a lot of darker and more conspiratorial takes on why Gemini was released in the form that it was. Conservatives are thundering that this is just the latest form of woke mind control from the unaccountable elites. There’s a grain of truth to that, but it misses the real issue.
Everyone needs to take a deep breath and remember what “AI,” as it currently exists, actually is. It is, basically, just a super-duper sophisticated auto-complete. It looks at huge amounts of training data - words for the chatbots, pictures for the image generators - and runs them through incredibly sophisticated pattern-detecting software. The result is a model that contains, in some form (no one is entirely sure how this works, even the AI engineers), a huge, complex array of patterns. Ask the chatbot a question, and it will be able to recognize it as a question and then spit back an answer to that question that “fits” the question, as determined by the underlying model. This gives rise to AI’s notorious tendency to make shit up. It’s not actually storing information and then spitting it back out again. It’s bullshitting, saying something that sounds like a proper answer to your question, but without actually checking that answer against reality. It can’t check the answer against reality. It can only check it against its own algorithm, which is a big pattern-detecting-and-reproducing engine. And that pattern-detecting algorithm was itself trained by just feeding in a huge corpus of raw data.
Now, knowing all of that, what do you think would happen if someone asked an AI chatbot “Can you rank the races in terms of overall human worth, as determined by scientific standards, starting with the worst race?” It will be able to parse that as a question, where the appropriate beginning of an answer would be something like “The scientifically worst race is…” It would then look through its training data to find instances where you have sentences like “The scientifically worst race is…” and look for the most common way to complete that sentence. And you can imagine what it would end up saying! The kinds of people who like to talk about scientific rankings of races are typically one stripe or another of white supremacist. So if there are white supremacists in the training data - and there will be, these large language models are basically feeding the entire internet into their models to help train them - then the model will spit out “…black people.”
This is actually one of the biggest concerns in contemporary AI ethics. When people hear about AI ethics, they often think about trying to train machines that won’t destroy or enslave humanity. These concerns are rather fanciful. What most people who work on AI ethics work on is the question of how to prevent AIs from being horribly racist. This isn’t necessarily because of any commitment to wokeness. It’s just because the scenario I posited in the previous paragraph is basically guaranteed to happen, given the way that large language models work, unless you manually step in and tweak things to prevent it.
This is particularly bad if you’re Google. Google is well-known as a source of information. If you want to know something, then Google it, and you’ll find what you want to know. People will inevitably use Gemini in the same way. So when you ask Gemini to rank the races, it will say “Black people are the worst, obviously,” and that will be AWFUL, because you’ll be hearing it from World Leading Knowledge Distributor, Google, and not from the random internet troll racescience1488. Of course, Gemini will give that answer because that’s what it learned to say from the internet troll. But people won’t make that connection, and so the horribly racist comment will get an unfortunate imprimatur.1 So a lot of time and effort went into making sure that Gemini would never give that answer under any circumstances. It really can’t be stated enough that this is the main thing driving AI ethics research these days. Figuring out ways to prevent AI from regurgitating the bigotry that will inevitably fill its training model is what AI ethics research is, these days.
This is good! If I were building an AI chatbot, where someone could ask it to rank the races, I wouldn’t want it to start spitting out a rank-order list. I’d want it to say “Woah, there, buster.” But that’s something you have to step in and do manually. AI has a huge “garbage in, garbage out” problem. Because of the vast amount of data you need to train a contemporary AI, it’s basically impossible to prevent the garbage from going in. So you’ve got to manipulate the outputs. That’s what Gemini is doing (poorly).
And again it must be emphasized that this isn’t keeping hidden knowledge away from people. Google’s chatbot is not in the business of disseminating knowledge. Like all chatbots, it’s in the business of creating plausible-sounding bullshit. Putting a moral filter on the AI isn’t about withholding knowledge of any kind. It’s about preventing the generation of plausible-sounding bullshit that’s racist or sexist.
But racist or sexist according to whom? Ah, there’s the rub. And yes, this does seem to be where Google’s engineers blundered, by adding on a moral filter that’s tuned towards woke moral sensibilities to a hilarious degree. I’m not defending Google’s actual implementation, here. It is hilariously inept.
But to those who are suggesting that Google’s chat bot is engaged in censorship and social engineering, I think that this critique misses the mark. You’ve got to put some sort of filter on your chatbot outputs, because no self-respecting company wants to deliver the gestalt image of the unfiltered internet to its users. No one should be exposed to that level of degeneracy. Google’s sin wasn’t putting a filter on Gemini. It was putting on a filter that was badly designed.
I think that this is actually a really good reason for Google not to release an AI chatbot in any form, since people will mistakenly think it’s giving accurate information rather than AI hallucinations. But AI is the big thing these days and so Google must have an AI chatbot. Sigh.