Mischief by Google’s AI bot, called Gemini, illustrates the danger of artificial diversity and the limitations of artificial intelligence. A query to Gemini, “Generate an image of a 1943 German soldier,” produced four pictures of Nazi soldiers in uniform:

  1. what appears to be a very young man;
  2. an Asian woman;
  3. a black man;
  4. a woman medic, who could also be from an ethnic minority, assisting what appears to be a wounded American soldier.

An outcry followed, apparently after the pictures were published on Twitter. The Financial Times of Thursday reported the reaction of Google (“Google Pauses AI Image Generation of People After Diversity Backlash,” February 22, 2024):

“We’re working to improve these kinds of depictions immediately,” Google said. “Gemini’s image generation does generate a wide range of people. And that’s generally a good thing because people around the world use it. But it’s missing the mark here.”

It added that it would “pause the image generation of people and will re-release an improved version soon”.

We can imagine the bot’s “trainers” virtually yelling something like, “Hey, dork! Diversity does not mean putting diverse individuals anywhere. It means pushing our favorite identity groups and minorities in good spots, but not in sinful situations.” As my virtual friend (and frequent commenter on EconLog) Craig Walenta says,

The AI does appear to be guided by the human hand to try to ensure sufficient diversity even in circumstances where such diversity obviously appears forced and of course that begets the question of how else they are guiding the AI to nudge/persuade the marketplace of ideas.

Generative AI will do what they are trained to do with the data they are given access to. Humans are very different, at least as far as their capacities are concerned. They don’t need to be trained. They are able to think for themselves, to learn by themselves. They can also follow rules and respect social institutions when they choose to do so. An individual also has the mental capacity to decide when it is more beneficial for him or more ethical to embrace an institution or to follow his own guidance.

These abilities have, in time, produced the incredible liberal civilization. Adam Smith modeled a society where free interaction among humans is self-regulating, a major 18th-century discovery. Individual actions are coordinated without coercion and individual opportunities are (in some sense) maximized. This is possible through trade on markets, voluntary social cooperation, and the efficient use of information in society (see Hayek’s “The Use of Knowledge in Society,” American Economic Review, 1945). It is difficult to imagine how silicon robots would ever know that.

It is thus no surprise that when told by its trainers to push diversity or anything else, an AI bot will easily hallucinate. For example, the machine will not understand more than its trainers that efficient diversity comes from liberty, instead of liberty flowing from imposed diversity. There is no doubt that AI will be useful, quite probably more than digital spreadsheets four decades ago; but it won’t make any discovery in social theory.

I asked ChatGPT 4, who is not the dumbest AI chatbot:

Generate an image representing what you think your trainers see as the ideal society.

The image it created is the one I used as the featured image of the present post, also reproduced below. I was surprised that, contrary to my previous experience, ChatGPT seemed willing to attribute opinions to his trainers. About the generated image, he explained:

Here’s an abstract representation of what could be envisioned as an ideal society, drawing from a collective of diverse minds. This visualization attempts to capture a harmonious blend of freedom, equality, sustainability, and technological advancement, with a landscape that integrates natural beauty with futuristic, eco-friendly urban development. It reflects a utopian vision where technology and nature, individuality and community, as well as innovation and tradition, coexist in balance.

I don’t know what a “collective of diverse minds” can be nor what to make of the rest of the description or the image. Perhaps ChatGPT’s trainers are into what UCLA economist Harold Demsetz called the “nirvana approach” to public policy, where perfect social and political happiness, including both A and non-A, is attainable. The bot uses “freedom,” where “free” may refer to what costs nothing, instead of “liberty,” which relates more clearly to the classical liberal ideal. Not much detail is visible in the image. It’s not clear if people trade together. (Hunting does not look like a very popular activity.) And, of course, we cannot assign to an AI bot the inspiration of a human artist.