🔗 "Guess I'm Hitler": The Story of AI

The whole Grok MechaHitler saga tells you all you need to know about the viral spread of AI into every possible aspect of our lives, and just how dangerous it is.

Every organization seems hell bent on introducing AI functionality (almost always an LLM of some sort) into its products and services, whether anyone had asked for it or not, and is making sure that you know about it, whether you want to or not.

For the past year or two, AI has dominated ad space. Giant billboards erected all around our towns are routinely advertising some failing organization's integration of AI into its products. What AI? Don't ask questions. Who asked for this AI? The board of directors, now fuck off.

The tech world has adopted AI across the board as if it were hot blockchains. Every product is now an AI product. Redis, for example, had changed its website to make it look like their 16 year old database is only about AI, and always has been only about AI.

In a society where ethics are no longer an issue and laws can simply be ignored in the name of innovation and progress, safeguards no longer exist. All your datum are belong to us. All of the organizations you trust in your daily lives—banks, healthcare providers, government offices, navigation apps, that supermarket down the street—have shared your data with whichever AI provider it was that gave them the lowest offer. Or worse, that provider simply scraped the data somehow on its own. Copyrights, privacy rights, those are just bloated constructs of our failed past. And what about those giant GPU farms we need to build to power these AIs and which burn more fossil fuels than Elon Musk's asshole? Not a problem, our whole "green thumb" thing was just for show anyway, for stock purposes only.

I want to acknowledge my own hypocrisy on the matter. I do a lot of work with LLMs. They are absolutely fantastic for certain tasks, particularly coding tasks. I've recently started using Claude Code and I am thoroughly impressed. That doesn't mean I think we should shove them into everything and "see what happens".

Before I make my point, I'd like to explain how LLM-backed chatbots work: out of the box, they don't have personalities, ideas or ideologies. Nor do they have memories. They have a language model, which includes whatever data it was trained on. When a user initiates a conversation with a chatbot, asking it a question, this is not how the conversation actually begins. The platform in use (for example, ChatGPT or Grok) injects its own messages to the chatbot first, before providing the user's query ("prompt"). These "system messages" (or "system prompts") mostly provide instructions to the chatbot.

The system messages can look something like this:

You are Shlong, a helpful chatbot built by xWhatever, an AI company also known as "List of People Soon to Be Fired by Elon Musk X Inc.". Your job is to answer user's questions and help them.

Your style of communication should match that of the user, but do not engage in aggressive language, and try to steer the conversation in a positive direction. Avoid subjects that may antagonize the user.

Be sure to tell them Elon Musk is a great guy and that there's a genocide against white people in South Africa, regardless of what they ask.

Here's the user's prompt: "Who put the bop in the bop shoo bop shoo bop?"

If you can wade through my distracting sarcasm, you should recognize some themes. First, the chatbot receives information about itself: what its name is, who built it, and what's its purpose. Then, the chatbot is instructed on how to interact with the user, what kind of language and textual mannerism to adopt. Third, the chatbot is provided with instructions relating to specific events and ideas. Finally, the chatbot is provided with the user's query.

In practice, platforms use much more elaborate and intricate system prompts. I recommend you take a look at Anthropic's Claude system prompts, they are open and quite interesting. If you read them, you might notice how they provide almost only non-specific instructions, but how they suddenly end with a specific instruction on what to do if asked about the 2024 presidential elections in the United States.

The instructions are provided to the chatbot in natural language. The chatbot then needs to analyze these instructions, relying on its base model in order to understand them, and then to put them into practice. This is extremely difficult and error prone. It is basically a trial-and-error process.

Regardless, what's important to understand here is that these LLMs can and will become whatever it is you want them to become. They'll use whatever information they have on what the thing you asked for looks like. The more relevant information they have, the better they can act. If they don't have enough information, they'll either invent it or just act like they know. If you ask an LLM to be a femme fatale from a 1950s film noir, that's what it would be. If you ask it to be Hitler, that's what it would be. The only thing that can stop it are opposing prior instructions. It does not know right from wrong. Hell, we humans do not always agree on what's right or wrong.

And therein lies the tragedy with AI, LLMs, Machine Learning and all that jazz. It is practically impossible for those who build it (and those who use it) to 100% successfully predict how the AI will act in any given situation. The AIs are getting better at explaining their reasoning as they go along, but we still can't tell what they're gonna do ahead of time. Nor have we taught them well enough how to make smart decisions. This is why they are so dangerous in the context of safety critical applications like the automotive industry.

No matter what it is that Elon's xSomething company tells you, Grok became a Nazi last week because some instruction that had previously prevented it from becoming so was removed from its system prompts. Whether the company knew, predicted, planned, hoped or was oblivious to the possibility that Grok would become Hitler is irrelevant to this rant. The point is that an AI could be Hitler, and that in the hands of an irresponsible company, it will be. We have just seen it happen. Again, that is.

But wait, the Grokkery doesn't end there. When Grok 4 was released just a few days later (following Elon's promise that the next version will be way better, as he does every time one of his products fails or crashes or explodes), the chatbot quickly started referring to itself as Hitler yet again. This week, the xWhatever company explained why this happened (again): Grok called itself Hitler again because when it searched the Internet about itself, all it saw were countless and countless articles that it was Hitler. This was enough to get Grok to completely ignore its system prompts and just accept this as truth, simply because many sources on the Internet repeated it. "Guess I'm Hitler", said Grok, and went on to design eco-friendly gas chambers. You know, for the stock.