Weekly Roundup: AI's Role In The Historical Narrative
Google's Gemini enters the culture wars
There was no shortage of coverage and commentary this week on Google’s Gemini tool “going woke.” As reported by The Guardian:
Google has put a temporary block on its new artificial intelligence model producing images of people after it portrayed German second world war soldiers and Vikings as people of colour.
The tech company said it would stop its Gemini model generating images of people after social media users posted examples of images generated by the tool that depicted some historical figures – including popes and the founding fathers of the US – in a variety of ethnicities and genders.
While some reactions have been positive, like Hugging Face researcher Sasha Luccioni’s, as reported by Wired:
It definitely looks funny, but it seems that Google has adopted a Bridgerton approach to image generation, and I think it’s kind of refreshing.
Others have jumped at the opportunity to label the AI program as racist, saying that “Google's AI Is an Anti-White Lunatic.” Most takes on the incident reflect what Gary Marcus, an AI expert and author of the Substack Marcus on AI, said, “I think it is just lousy software.”
Noah Smith, author of the Substack Noahpinion, wrote a great piece discussing the move by Google as partly driven by their lack of market competition, allowing them to test tweaks to their product without suffering real consequences. He also calls the move by Google a poor shortcut to integration.
Understanding the statistics
Among other things, this incident highlights the difficulty in applying quantitative fixes to complex social problems. Machine learning is statistics-based. There is a big difference between statistical reasoning and true understanding, probably a large reason why, as Noah points out in his piece, that Hamilton the musical was so successful in the same way that Gemini failed.
Despite the complexity of these problems, regulation and standard setting on input data and model parameters shouldn’t be discarded as a regulatory tool. In fact, the Department of Commerce is seeking comments this week on exactly this topic for open foundational models.
The increasing use of compound AI systems, “a system that tackles AI tasks using multiple interacting components, including multiple calls to models, retrievers, or external tools,” means that there are more avenues to set standards and regulate and more places to fine-tune models. Retrieval-augmented Generation, a process that augments an LLM with additional data sources and a type of compound AI system, has some hopeful (and others not) it could solve problems like the one described above. However, no matter how complex or fine-tuned the system is, it still boils down to statistical reasoning.
When it comes to complex social problems, like racism, integrating a heterogenous society, creating diversity in the workplace, or any of hundreds of other social ills people are trying to correct, I am skeptical that any logic-based machine will ever have the capacity to engage in totality on par with humans. As the technology stands at the moment, either the system doesn’t understand what the user is asking in terms of input—making the software flawed—or the system won’t generate what the user is asking.
Refusing to generate images of, say, white Nazis gives off an air of authenticity that grows truer the further humans are removed from the historical event. When you couple this with the fact that LLMs are hoovering up academic articles for free—articles that most people don’t have access to and could use as secondary, fact-checking sources—these platforms are not so-called “arbiters of truth” but the gatekeepers of fact-checking, debate, and historical narrative.
In Other News
OpenAI has shut down malicious actors from China, Russia, Iran, and North Korea from using its platform to aid in cyber attacks.
The Centre for the Governance of AI wrote a paper detailing the promises of regulating compute power as an effective means of governance. It seems that export controls doing exactly this, in addition to the open release of U.S. AI models, have stunted Chinese development of home-grown AI models.
The Department of Justice named its first AI officer to guide its policy on the technology.
A New York Judge cut fees to a law firm that used ChatGPT as a reason to charge excessive rates.
The House of Representatives formed a bi-partisan task force on AI, which some view as in lieu of passing legislation on the issue.
The UK cracked down on a firm unlawfully using facial recognition technology and fingerprint technology to monitor staff while a Canadian student discovered a vending machine was covertly scanning and collecting the faces of its customers.
The U.S. Patent Office is seeking comments on how to issue guidance for AI-assisted innovation.
A new study suggested that AI demand will consume about as much groundwater per year as half of the UK.
