Hello and welcome to Eye on AI.
It was another big week in AI news with lots to discuss, from Nvidia’s soaring valuation following its stellar earnings announcement to indications that Big Tech is not as sure as it once was that proprietary AI models will be the path to cloud success.
But first, let’s talk about the culture war quagmire Alphabet waltzed into with an ill-conceived attempt to overcome AI’s inherent racial biases. The ham-fisted effort at putting some guardrails around the images from its Gemini models blew up in the company’s face, forcing it to temporarily disable Gemini’s image-creation capabilities and issue a public apology. Investors were not impressed, driving Alphabet’s stock down more than 4%. Some vitriolic critics even called for Alphabet CEO Sundar Pichai to step down or be fired.
The controversy highlights a few things. One is the public relations dilemma Big Tech in general, and Alphabet-owned Google in particular, faces on these sorts of issues. In 2022, OpenAI famously wrong-footed Google by releasing ChatGPT well before Google was ready to commercialize the rival LLM-based chatbot Lambda that it had long been incubating inside the company. At the time, Google used the fact that it was trying to be a “responsible” steward of AI technology as a justification for its inertia, and as a way to imply that OpenAI, and OpenAI’s partner Microsoft, were irresponsible in moving to commercialize generative AI so rapidly. Even as it began releasing commercial generative AI products of its own in a bid to catch up, Pichai promised the public that the company would always be “bold and responsible” when it came to AI innovation and gave the employees doing the work marching orders in line with that pledge.
Putting itself on this sort of pedestal when it comes to responsibility means that Google then has to try harder than other companies to put guardrails around the AI products it releases. But this may be a fool’s errand, for reasons we’ll explore in a moment. Other companies may not have tried particularly hard to deal with the well-known issue that AI models trained on historically biased datasets unsurprisingly produce biased results. In fact, there’s lots of evidence that Midjourney and OpenAI’s DALL-E produce racially biased imagery, and it hasn’t much affected investor sentiment around either company.
But Google being Google decided it needed to do something about the problem. The way it seems to have done so was to instruct Gemini behind the scenes to always generate images of an ethnically diverse set of people and to refuse prompts designed to have it generate images of only white people.
Of course, one person’s responsible is another person’s “woke,” and that was one of the big problems here. It also didn’t help that many on the right already see Google and its employees as hopelessly leftwing and were ready to pounce on exactly this kind of over-the-top effort at overcoming LLM’s racial bias. Elon Musk, who has promised that his Grok chatbot is “anti-woke,” happily helped ensure that Gemini’s issues with generating historically accurate depictions of ancient Rome or Vikings received wide airing.
More importantly, Gemini’s problems show the weaknesses of today’s AI models and our ideas about how to put guardrails around them. The idea of using metaprompts—or natural language instructions that are automatically appended to the user’s prompt but hidden from the user—as a way of creating guardrails, which seems to be part of what Google did with Gemini, is fraught. Why? Because LLMs, despite ingesting the entire internet’s worth of data, have extremely weak conceptual understanding and almost no common-sense reasoning.
Ideally, you want to be able to just tell the model “don’t be racist,” and have it understand what you mean and in what contexts it might be okay or not okay to depict non-diverse sets of people. If the model is uncertain, it ought to ask the user for clarification. That is what we would expect a competent human assistant to do if given those kinds of instructions. But the models we have can’t do this. And, in fact, Google seems to have built Gemini’s image generation guardrails partly through metaprompts and partly by fine-tuning the model only on images depicting diversity. But this made it so the model would struggle to generate non-diverse images even in contexts where that was appropriate.
“The technology is not very robust and there is no way to write an AI-based computer program that will make everybody happy all the time,” Meredith Broussard, the New York University journalism professor and author of More Than a Glitch: Confronting Race, Gender, and Ability Bias in Tech, tells me.
That said, the Gemini engineers who designed the guardrail prompts probably should have anticipated that users would ask the model for images of historical settings that were not diverse or try to find and break the model’s guardrails by asking it to generate images of only white or Black people. As much as this is a failure of the AI, it is also a failure of human imagination.
Broussard traces the problem to the underlying assumption that you can build a “general purpose” conversation agent in the first place. Other AI ethicists have made similar points about the marketing of LLMs as general-purpose tools. In their view, general-purpose technologies are ethically problematic because it is inherently difficult to evaluate them.
They argue it would be better to go back to creating smaller AI models tailored for specific purposes. This would allow the model builders to carefully curate a dataset for the problem the user is trying to solve. If you want to build an AI to generate accurate portrayals of ancient Rome, for instance, why not build a smaller AI model only to do that and train it only on images related to ancient Rome? And if you want it to create an AI model to generate images for marketing contemporary products, then maybe you give it the kind of instructions Google gave Gemini. In theory, it ought to be possible to fine-tune today’s large LLMs on small datasets and with specific prompts, for purposes such as these. But Gemini is not in the business of selling a million small tailored models. There’s more money in selling a single model as the tool for every use case.
The idea of abandoning large models is also, in essence, an abandonment of the quest to create more human-like AI. Some AI ethicists would be fine with that. But a lot of other people would not. It is also, frankly, unrealistic. The genie is out of the bottle. I don’t think we are going to be able to put it back again and revert to simply using small models.
So assuming we are going to keep using large, multipurpose models, then we desperately need to figure out ways of getting the models to understand human intentions. Because it is extremely difficult to create a prompt that will cover every possible scenario, we want a model that has enough common sense understanding to know that when the creator of the model says “don’t be racist” we don’t mean that the model should depict a 9th-century Vikings settlement as if it were a meeting of the Rainbow Coalition.
This is also why I think the schism between researchers working on “responsible AI” and “AI Safety” is unfortunate. Traditionally it is the responsible AI folks who have cared most about AI’s racial bias problem while the AI Safety people, who are concerned about AI potentially killing us all one day, have cared most about making sure AI models can properly understand human instructions and intentions.
As it turns out, a lot of that AI Safety work could also help us build better guardrails that would allow AI models to not be racist, and also not be ridiculously woke. That’s the kind of “bold and responsible” AI a lot of companies would love to have. And it would probably make Alphabet’s shareholders much happier than they are today.
Below, there’s more AI news. But before you go, if you want to learn about the latest developments in AI and how they will impact your business, please join me alongside leading figures from the business world, government, and academia at UnHerd’s inaugural London edition of our Brainstorm AI conference. It’s April 15-16 in London. You can apply to attend here.
Jeremy Kahn
[email protected]
@jeremyakahn
Correction, March 1: A news item below wrongly stated that Nat Friedman is GitHub’s CEO. He is the company’s former CEO.