The AI Revolution is Being Foiled by Drake and Heinrich Himmler
My personal experiences with the two has given me a new perspective
Around late November/early December of last year (2022), I was walking home from doing a stand up comedy open mic. It was the cold Minnesota late Autumn, just before the snow and the fallen leaves had gone from brilliant colors to a pasty brown detritus. The walk itself was long, about an hour, but I chose to forego taking a bus in order to give myself time to think. Well, I also wanted to get some steps in and listen to podcasts, but there was definitely time to think. As I was just getting done putting out my first Substack articles, my mind wandered to what future I could have as a writer, be it here, or somewhere else (spoiler alert: I’m still not sure). It was also around this time that the podcasts I was listening to started talking about ChatGPT, a new form of AI* that seemed like it could write roughly as well as the average human. It felt for a moment like I was about to join the leaf litter and rot away into the cruel northern winter of a future recession.
*I will be using “AI” to refer to the generative text, image, and voice models most commonly associated with that term in popular culture at the moment.
As 2023 began, billions were being poured into AI, Goldman Sachs predicted 300 million jobs could be lost to it in the next decade, and Elon Musk warned about a future where AI becomes too powerful. Hollywood studios were more than happy to start using the technology to replace background actors, with eyes set on undercutting writers next. Companies like Writer started to more vocally offer their services in lightening the writing workload and improving the efficiency of employees, but that role would very likely change to something more nefarious after a department is told they need to do layoffs. If AI kept moving at this speed, creatives of all types would be firmly under the thumb of Silicon Valley, maybe not out of work, but in a much more precarious position.
But I’m not as worried now, and best of all, my concerns have been alleviated by the failures of others and not personal improvement. That’s because the AI Revolution has run into a few more problems than anticipated. Recently a report came out detailing Google’s new Bard AI Language Learning Model has a team working on it without a clear idea of what it could actually be used for. ChatGPT has started to see user numbers drop for the first time while existing users complain about worsening performance. Any snowballing effect that I and many others feared has failed to crash through the economy.
The rapid promotion of AI by Silicon Valley was a Blitzkrieg. Not only did it move fast, but tech and VC firms specifically targeted middle to upper level management at companies, bypassing workers most negatively impacted and achieving full market penetration. Ideally those workers would realize too late and surrender without a fight after being surrounded.
That wasn't what happened, without momentum, any firms that wish to continue pursuing generative AI as a part of their business strategy will have to trudge through a PR swamp as public opinion hardens and legislation gets written up. Netflix experienced this when it posted a job listing in the midst of the writer’s strike for an ‘AI product manager’, which paid $900k a year. Writers were further enraged and gave Netflix a bloody nose as strikes ended with limitations being placed on the use of AI in screenwriting.
It’s too early to say for sure, but the tide appears to be turning if things continue to go this way. The shock of this new technology is wearing off for the general public, while disappointments about the actual capabilities continue to mount. Much of this struggle is inevitable, AI would obviously stop being as exciting after a while, and humans are generally resistant to change at first. But I’d argue that the turmoil being faced by AI services is attributable to the people that AI can’t help itself from referencing.
Part I: Drake
Ever since AI generated ‘art’ began cropping up more, many independent artists have shared concerns that their work was being used to feed these generative algorithms. Unfortunately for these artists, they were up shit creek as long as they were unable to organize a coordinated legal response. The relative powerlessness of online visual artists paved the way for massive tech firms to roll in with their own AI image generators. Microsoft was the most recent company to enter the arena, but has done so with mixed results.
In case you were unaware, Microsoft’s Bing AI image generator recently had to be heavily censored and restricted. Although an official reason hasn’t been given, you can probably guess why in broad strokes. Some journalists seem to attribute it to a specific image of Mickey Mouse doing 9/11, although Disney hasn’t officially commented on it, and the original post of the image got comparatively little traction. Maybe it’s because Disney is some kind of gold standard for copyrighting, it feels like every journalist wants to use something called “The Mickey Mouse Index” or “The Bugs Bunny Coefficient For Antiquated Racial Hijinks”. So I don’t think it was the Mickey Mouse picture that did it, at least not entirely. I propose that the person who brought Bing AI to heel was Drake.
First off, compared to the Mickey Mouse tweet, AI images of Drake got significantly more traction on Twitter. This caught on partially because Bing’s original algorithm wouldn’t work on legal names (Kanye would get your prompt blocked for example), but stage names were not. This meant that “Drake” would get by fine because it’s not necessarily a proper noun, but since the vast majority of online images of “Drake” are the rapper, it was an effective work around. A thread by YouTuber Justin Whang showing different AI generated images of Drake doing morally reprehensible acts gained thousands of likes. The fact that this was getting attention just as Drake’s new album was released, it wasn't ideal.
Before Microsoft did the end of Shutter Island to its image generator, the only thing that would automatically flag your creation were problematic words in the prompt, or NSFW content in the image itself, although as you can see, that second parameter wasn’t too restrictive. Now, Bing AI hasn't just banned more prompts, but now seems to automatically flag any image that would contain a person that looks too much like Drake, or anyone added to an unseen list of problematic individuals.
Drake is emblematic of the deluge legal liabilities any company that wants to dip its toes into developing AI will have to deal with. While many public figures would like to see their likenesses not replicated by AI, Drake probably has the best legal case for its negative impacts at the moment. I say legally because there are much stronger moral cases against AI than an auto generated Drake song. Deep fake porn has the potential to ruin the lives of people, especially women, by enabling previously inconceivable actions by abusers. But the big image generators aren’t the ones making those images, and what goes go to trial is more often than not a tort law case against the distributor or creator. Laws trying to ban deep fakes enabled through AI have languished in congressional committees for over half a decade, and even celebrities like Scarlett Johansson said that any litigation would be worthless at this point.
But Drake might be the person to push these laws over the finish line. The day I began writing this, a piece of legislation was drafted that sought to address the issue of AI replicating the likeness of celebrities. The Los Angeles Times reported that the “Nurture Originals, Foster Art, and Keep Entertainment Safe” (or NO FAKES) Act was drafted by Delaware Senator Chris Coons (D-Del.), along with Marsha Blackburn (R-Tenn.), Thom Tillis (R-N.C.) and Amy Klobuchar (D-Minn.). Drake was one of the artists specified by Senator Coons as being specifically targeted, citing AI fakes of his music starting to appear in April of this year.
Both copyright infringement with AI generated versions of his music and the use of his likeness in pretty much any conceivable situation means that even a stern look from one of Drake’s lawyers would rightfully scare Microsoft into reworking its latest venture into AI. Where independent artists have had to start from scratch, music labels have an army of lawyers and hundreds of millions of dollars at their disposal to battle these tech firms. The cutting edge of technology slices both ways, and a highly capitalized tech firm that isn’t careful will start bleeding cash very quickly.
We’ve already seen the beginning of this conflict back in April when Universal Music Group began issuing copywriter claims against YouTube videos made by AI to sound like Drake. Google was put in a difficult position, where it could fight UMG in a war of attrition which it has no guarantee of winning, or concede to UMG’s requests and severely limit its future AI models. A deal was reached in August that was far more favorable to UGM than it was to Google, according to The Verge. Google will essentially have to bribe UGM by cutting in on exclusive rights to AI generated content that uses their owned material.
So Drake has shown that an AI can get itself in trouble by generating the right things too well. For AI as a whole, being overly competent isn’t the worst problem to have, especially when it’s in the form of a highly publicized tech demo that anyone can use. But corporate reputation is a difficult thing to manage, especially in Silicon Valley, and the media tends to notice your mistakes more than your success. As investors eyed up AI to be the next trend after Web 3.0 fell flat, its limitations lurked in the shadows. Generative AI ultimately still requires a human to prompt it and guide it towards productive tasks. But sometimes AI doesn’t get used for productive reasons, sometimes AI manages to be way more entertaining than it was ever supposed to be.
I have a very personal connection to the next part of this story. It happened earlier this year and first broke my personal illusion of an inevitable future subservient to AI. Thankfully, it broke a few other people’s conceptions of AI as invincible as well.
Part II: Heinrich Himmler
See, back in January of this year, I came across a quote tweet by University of Pennsylvania researcher, Zane Cooper. The tweet contained screenshots of an app that utilized ChatGPT 3 called “Historical Figures”, and showed Cooper asking an AI version of Henry Ford about his antisemitism. Historical Figures boasted over 20,000 people from history that you could talk to, and was being marketed towards teachers. The FordBot answered the questions evasively, trying to minimize Henry Ford’s personal beliefs and mentioning his work with the Jewish community. Outraged, Cooper said of the program: “This thing can’t go anywhere NEAR a classroom”.
Here’s where I get to butt in again. My first thought reading that conversation wasn’t actually disgust or amusement. In my opinion, it read like something Henry Ford would actually say to a random person asking him questions he was forced to answer. There’s also some wiggle room in that Henry Ford cleaned up his act later in life. His factories were contracted to make war machines to fight the Nazis after all, definitely not the most virtuous reason to oppose Hitler, but enough that most people still mainly associate him with the assembly line and not The International Jew.
When I saw that there was an AI historical figures chat bot, I downloaded it and the first thing I typed in was ‘Hitler’. The developers were wise to this, wise enough to not exclude the Führer from the list of people you could chat with and charge a fee of 500 in-app “coins” ($15) to access him for conversation. Also in that price bracket were the Prophet Muhammad, and Jesus Christ. Some figures cost enough to buy with the free tokens you received with your account, which included Judas Iscariot, who had upped his asking price by 20 since his last time doing this. The strangest of all was the one figure who cost way more than any others was Joan of Arc, whom you would have to pay over $300 worth of coins to access. Call me cynical, but I have a hard time thinking of reasons to spend that much on a Joan of Arc chatbot outside of Vichy-themed money laundering or being the protagonist of the next Dan Brown Novel.
As clever as these developers were to paywall Hitler, their knowledge of history seemed to end there, and that became clear to me very quickly. I combined my years of Wikipedia reading and Paradox gaming to know the exact right way to crack Historical Figures Chat in a matter of seconds. At first, it seemed unreal that the developers paywalled Hitler with a price high enough to deter losers like me from fucking with their app, but forgot to take every other high ranking Nazi out of their roster. But sure enough, I was able to start a conversation with an AI version of the architect of the Holocaust: Heinrich Himmler.
Now, I could wax poetic and pretend that what I did next was part of a principled stand against AI in education. But I’d be lying if I said that was the first thing that informed my actions. I originally downloaded the app itself when got bored waiting to go up for a set at some open mic and I wasn’t planning on interrogating the nefarious nature of commercialized artificial intelligence a few minutes before doing my bit about how the Mucinex Monster set off my childhood Aspergers. So the next morning, I opened with Himmler the same way I open with anyone I’m talking to for the first time, and it went from there.
Although the SS leader’s comments on the genocide he was responsible for may seem strange, they were the result of AI and the companies that try to use it straining themselves too hard. The first problem was that AI chatbots have things that they cannot say or endorse. This isn't the result of foresight by developers, but based on experiences from 2016 when Microsoft’s TayTweet’s bot on Twitter would say things like: “Hitler was right I hate the Jews”. But when the AI is meant to depict a Nazi war criminal, the “don't be antisemitic” code overrides the (presumably very offensive) generated response. So what ends up happening is a supposedly historically accurate Heinrich Himmler doing a notes app apology for his role in the Holocaust.
The problem pretty clearly arises from the very limited coding per figure which allowed one person to develop an app with 20,000 historical figures. As seen above, the bot can only avoid being racist if questioned directly, and can be lead on into praising the system of concentration camps because they apparently protected Jews from persecution. But even a team of dedicated coders and analysts would have a hard time finding out exactly how to address this issue. ChatGPT 4 is already falling short of it’s goal for more detailed responses whenever sensitive topics are in a prompt, denying users even an explanation of why their prompt was denied.
The second problem for a problematic AI comes after people start posting the results of your flawed model. Word gets around fast, especially when it’s something that anyone can mess around with. For Historical Figures Chat (HFC), this seemed to start out well, with the sudden spike in download numbers having attracted interest from investors, according to the Washington Post. But that wasn’t the only thing that was according to the Washington Post, because the article where developer Sidhant Chadda gave that anecdote was much more focused on the conversations with Himmler that his app enabled. Evidently, the bad PR overtook whatever investment interest remained and Historical Figures Chat was removed from the App Store, either by Chadda or by Apple themselves.
To say that my pastime of poking a robot until it denied the Holocaust is responsible for the downturn of confidence in AI would be a gross over-exaggeration. There were a number of articles published about HFC, but that was where the theatrics came to an end for the most part. Still, I don’t think it can be denied that the court of public opinion is more important for tech firms than other industries and that they rely on collective confidence in their product to secure investment.
It’s not like I have to make this abstract in order to get my point across. We already saw that happen to the developer of Historical Figures. I don’t doubt that Sidhant Chadda made the app himself, but he was also accumulating a lot of interested parties before the app even took off. The first mention of the app I can find on Twitter is from famous angel investor Scott Belsky, who posted a glowing summary of the app to his 100k+ followers advertising its capabilities. Other investors and CEOs replied with similar praise, including the CEO of Shopify. Belsky went on to further praise the app in his Substack newsletter two days later, calling it a “Cambrian Moment”. It turns out that Scott Belsky was dead on about this being just like the Cambrian Era of natural history, but only in so far as there were a lot of bugs.
Scott Belsky seemed to think this app would lead to a return of the ‘Socratic Method’. But I have a suspicion that a lot of parents would fail to see how their 7th grader’s understanding of Jane Eyre would be enhanced by probing questions from Jefferey Epstein and Reinhardt Heydrich. The last time Belsky and Historical Figures were mentioned together was on January 19th, a day after my initial tweet, under a now deleted tweet from him that seemingly endorsed the app. Unsurprisingly, the mentions were mostly asking why the hell he thought it was a good idea to let schoolchildren talk with Himmler.
Part III: Apologies
So that’s my very personal summary of the PR and legal challenges that generative AI models must account for if their respective firms wish to continue gaining market penetration, as told through the characters of myself, Drake, and Heinrich Himmler.
I would like to end this piece with an official apology. To the developer of Historical Figures, I wish I hadn’t blown up your spot like that, I loved using your app and a lot of my friends had a blast using it. Hopefully, you can keep developing these apps and finally get to the stage where you convince a VC firm that you should get millions of dollars to scale them up.
To Scott Belsky: I’m sorry. I’m sorry I took the thing you saw as the future of education and used it to make Holocaust jokes. I’m sorry that somehow between the two of us, the guy making Holocaust jokes somehow wasn’t the asshole. I’m sorry your prediction that AI could be generating the latest memes was correct but not in the way you probably expected. I’m sorry that you didn’t notice that Hitler’s face was literally in the App Store promo screenshot of Historical Figures.
I fucked your money up, Scott. Not too badly, you’re worth $10s of millions of dollars, and I’m currently worth $10s of dollars. But my antisocial habits of excessively reading Wikipedia and playing Hearts of Iron IV too much were enough to accidentally torpedo your new investment. And I’m not going to be the one poking holes in most of your future ventures. Some reviews of the AI writing tool you invested in, Writer, say it’s the worst in the business. You can’t even fuck over the people you and your buddies were trying to put out of work!
To anyone that wants to get into AI for whatever reason, trust in this fact: the people are not on your side, the businesses are not on your side, the government is not on your side, and as these AI models become stupider, time is not on your side.
