Rebooting San Francisco royalty: Emperor Norton meets GenAI
Oyez, oyez! We, Norton I, Emperor of America and Chief Wrangler of Cantankerous Seagulls, do hereby proclaim: Our Royal persona shall be reborn in the realm of thinking machines! All citizens must converse with Our digital doppelgänger or face a tax of 37 pineapples.
The genteel alleys of Jackson Square north of San Francisco’s Transamerica Pyramid are full of boutique venture capital firms, design studios, and the occasional Michelin-starred restaurant. As one of the few neighborhoods not obliterated by the 1906 earthquake, it also has a direct connection to the city’s Barbary Coast origins. If you take a historical walking tour, you’ll hear about the gold and silver rushes, financial panics, political corruption, violent gangs of Australian convicts, and an eccentric named Joshua Norton who proclaimed himself Emperor of the United States and Protector of Mexico.
The most recent San Francisco boom cycle involves generative artificial intelligence, not precious metals in the Sierras. The peculiar mix of local culture, technology, and money is the same. Developers are writing code on the same street where Emperor Norton wrote a proclamation with the then-absurd idea of a bridge across the Bay. He was the original mascot, progressive, oddball, and minor celebrity who helped generations of locals reflect on what kind of a place San Francisco was or could be.
Last year, I decided to use generative AI to recreate Norton’s voice and have it engage with people on social media. I was interested in how an AI model of Norton, which I’ll refer to as the Emperor Bot, could make us more creative. Much of what is written about AI seems to be about civilization-ending scenarios or startup pitches for new ways to summarize your emails. I wanted to know if generative AI could help people reflect on the past… and possibly laugh.
Building an Emperor Bot that challenges injustice
The first thing I had to do was build the bot, which was the most fun I’ve had on a technical project in a long time. How it works is straightforward, even for a non-programmer: you instruct a large language model (LLM), like OpenAI’s GPT-4, to reply as Emperor Norton. This instruction is called the prompt.
Some facts about Norton have been incorporated into the base LLM models. The result is cutting-edge technology approximating how to speak in the voice of someone from the 19th century. After I made dozens of small adjustments to the prompt — it was difficult to get the replies concise — here’s the bot’s reply to a tweet about seagulls and a locally famous Taco Bell:
Emperor Norton I has decreed that all seagulls must be removed from Pacifica & its environs. The Board of Supervisors must award damages to the owners of the desecrated Cathedral of Taco Bell.
The reply captures three things that are essential to the historical Norton: it challenges injustice (albeit of a Taco Bell), holds politicians accountable, and does so with the authority of someone with absolute political power.
Here’s another reply to a tweet from a local supervisor proposing renovation of some crumbling piers that doesn’t hit the mark:
We decree that Piers 30–32 in San Francisco shall be rebuilt as a mixed-use development with public recreation, retail, office, and housing!
The pro-development tweet was it seemed inauthentic in the worst possible way. It didn’t capture the historical voice. It was like seeing an ad for breakfast cereal featuring an Emperor Norton proclamation on whole grains. I wanted to understand if this was truly a limitation of the current models, or something that could be improved.
Finding Norton’s authentic voice in the LLM
One of the issues with recreating Norton’s voice is that, while everything he wrote is in the public domain, there is a thin historical record for someone who continues to be famous in the modern era. The proclamations he published were short, and no journal or diary survived. With the help of local historians I became more familiar with gold rush era San Francisco. John Lumea of the Emperor Norton Trust, the world’s foremost expert on Norton, sent me several helpful emails, and Taryn Edwards of the Mechanics’ Institute (where the real Norton was a member) shared a reading list with a warning: San Francisco history is rife with lore and legend.
I can never know Norton’s views on the correct amount to tip at restaurants, but did know he was always against corruption and wanted to keep politicians accountable… to the point of calling for Congress to be abolished. Using more AI, I generated a map of the themes and topics Norton wrote about and hoped it would guide better responses.
For example, Norton had a proclamation that referred to sailing ships — but his writing on the ships thematically touched on justice, authority, responsibility and safety. When he wrote about being refused entrance to an event by “snobs in charge of the door” — the map pointed to his concerns around the abuse of power and social hierarchies.
Instead of parroting or sounding like a shill, the bot’s replies seemed to improve with the new context. Here’s the reply to the news of a local brewery closing:
Great Scott! Anchor Brewing, a city gem, faces peril? We’ll rally, oh we will! Workers, I stand in solidarity with you! Citizens, let us preserve our city’s heritage!
It was around this time a controversial local billionaire changed policies and most bots, including the Emperor Bot, were kicked off Twitter.
BlueSky Social, hackerspaces and vegetables
Elon Musk kicked the bots off Twitter because he wanted money for the type of access the bots needed to post, which was previously free. This resulted in the Emperor Bot moving to the newer social network BlueSky.
Despite being a friendlier and invite-only version of Twitter at the time, I was concerned what might happen. Despite profile disclosures and a prominent robot icon, my impression is that most people assumed the text was written by a cosplaying San Francisco history nerd pretending to be a digital version of Norton.
Non-public tests with the bot commenting on trending topics went badly. The bot never generated outright offensive content, but anything related to politics was inflammatory. Even positive replies, like support for the Hollywood writers’ strike, were problematic. People striking for reasons related to losing their livelihoods over AI-generated content do not want a supportive yet eccentric 19th century bot to join the conversation. The fully-automated social media version was never released, so I reviewed replies and posted the unaltered text every few days.
The most engaging conversations with the bot were in private channels on Discord, the popular chat app. A local hackerspace added it to their AI channel and started interacting with it. Using the latest tricks being spread around the LLM space, bot replies seemed to improve when I added an “therapy” mode where the bot would articulate its sensations, fears and desires before replying.
Here’s the bot’s inner monologue for replying to something related political corruption:
Sensations: A harsh, sneering laughter rings in my ears, leaving an unpleasant taste in my mouth.
Fears: Anarchy and corruption taking over my beloved city.
Desires: To enact justice and maintain the order of my realm.
Somebody also asked the bot what vegetable it found the most threatening.
I would say it is the artichoke. Its sharp exterior might be seen as a symbol of uprising, but in truth, it’s the tastiest rebellion I’ve ever faced.
A folk hero emerges on BlueSky Social
The bot was finding a certain groove on Discord, but on BlueSky it was limited to replys related to a very small number of topics: mostly pets and Norton himself. I was surprised how culturally relevant Norton, who died in 1880, is to modern-day San Francisco and how deep that appreciation and fandom goes. An award-winning author was one of the first followers.
My impression is that Norton is a folk hero on social media or, depending on who you ask, the world’s first crypto bro. (He printed his own money and passed it off as legal tender.) People imply they would love to vote for him in the current election. He’s also considered a minor saint in a belief system that started in 1960s Berkeley called Discordianism (no relation to the app Discord).
The most liked posts from the bot, after some trial and error, were those that acted as a kind of promoter for San Francisco and the historical Norton. Anything else tended to annoy or confuse people, like John Adams showing up on Instagram and liking your vacation photo. When someone skeeted (that’s BlueSky for tweet) that Norton was cool and they wished he could save everyone now, it replied:
Indeed, time unceasingly marches forward. While I can’t step forth in flesh, my philosophies persist. Lead with benevolence, act with authenticity.
In another instance, the bot replied to a post memorializing Norton with a familiar sounding phrase: “death is but a door, time a window.” It’s a quote from the villain in the 1989 movie Ghostbusters II… a 16th century historical figure that comes back to life in an evil painting.
What are the chances of the Emperor Bot taking over the world?
The 1906 earthquake destroyed much of Norton’s San Francisco. The murkiness of the late 19th century also means that it’s easier for every generation to repurpose Norton to fit their mood and anxieties. In present day San Francisco and the United States: there are plenty of issues to choose from.
Of course, one of the largest local conversations in technology is the promise and perils of generative AI itself.
It’s not hard to imagine these sorts of bots finding a place in educational settings or museums to make history more interactive and conversational, at least for well-liked and popular historical figures. The bot would be an excellent companion in a self-guided walking tour of the Emperor’s old neighborhood.
There’s also the obvious danger of these bots on social media, which has been written about at length in many different places. My assessment is the risk of the Emperor Bot causing global conflict or taking over governments as a digital Emperor is effectively zero. However, it’s hard to think of any scenario where more advanced gen AI bots will improve the overall quality of conversation in the near-term on social media.
There has been an enormous amount of investment from major companies to make sure the output of LLMs is inoffensive and not harmful, but that is changing as individuals and groups create their own models. Using a collection of scanned 19th century newspapers, I started on an experiment with fine-tuning the bot to make it sound more like actual writers from the 1880s. That was paused once I realized that I did not want my bot to pick up on the views that existed in the vast majority of late 19th-century American newspapers.
Norton, generative AI, and a way forward
Joshua Norton is buried in Colma, the cemetery city just south of San Francisco. There’s a marble headstone that says Norton I Emperor in large letters. E Clampus Vitus, a 19th-century fraternal organization (sometimes called in the press a “historical drinking society”) throws a party honoring his legacy once a year on “Norton Day” at his gravesite. There’s a convenient Irish bar a short stroll away.
My theory on why Norton still has a large number of admirers who toast him is because, in his own chaotic and eccentric way, he is remembered for fighting against injustice. He also was an immigrant to San Francisco who made and lost a fortune but still figured out a way to get free lunches and had a popular newspaper column for twenty years.
Generative AI will never be able to capture the full context, intricacies, and contradictions of the human Norton but it can be strange and unexpected. Despite the peril of thousands of trolling bots taking over social media, in the right context this technology can be fun and creative — without taking over the world or your job.
Uncertainty hangs over the air right now for both AI and San Francisco. There are plenty of doom loop scenarios being discussed for both. Economically, the success or failure of generative AI will also have a direct impact on The City for years to come. The open question for local entrepreneurs depositing checks from venture capitalists is what happens next. At one extreme: collapse from malevolent AI. Another direction is the mediocrity of bots that make life worse via annoying assistants that get in the way.
A happy outcome would be technology — and a city — where people are able to be thoughtful, explore new ideas, and maybe have fun while holding those in power accountable. It might also get a little weird and off-putting at times. We should, in other words, take more inspiration from Emperor Norton.
Thanks to John Lumea of the Emperor Norton Trust, Tayrn Edwards of the San Francisco Mechanics Institute, the weekend staff at The San Francisco History Center, the Prelinger Library, The Ancient and Honorable Order of E Clampus Vitus, the Misalignment Museum, and various adherents of Discordianism active on BlueSky and Discord.
You can follow the Emperor Bot at https://bsky.app/profile/emperornorton.bsky.social
