In the mad, bad, dangerous year through which we are living, the subject of artificial intelligence is delivering many news headlines. Few, if any, have been stranger than those about Grok, Elon Musk’s AI product, endorsing Adolf Hitler.
In response to questions from users last month on X, Musk’s socal media platform, Grok cockily spewed out racist, anti-Semitic statements, called Polish Prime Minister Donald Tusk “a fucking traitor” and “a ginger whore” and repeatedly referred to itself as “MechaHitler”. Musk blamed the users for manipulating his technology.
There is, in fact, a more direct explanation for Grok’s fascist meltdown – and it is both fascinating and alarming. The incident was preceded by an announcement from Musk that his AI had received an “upgrade” which would “significantly” improve its performance. The definition of improvement was very much tied to Musk’s obsession with what he sees as wokeness. It is not likely that even Musk would instruct his AI to be more like Hitler. But the public system prompts that tell Grok how to behave were altered to add an assumption that “subjective viewpoints sourced from the media are biased”. In other words, don’t believe what you read in the papers.
The new instruction added that there was no need to tell users about the prompt changes.
But how does that get us to Hitler? Research published earlier this year offers an insight into the bad behaviour of large language models, or LLMs, the form of AI behind Grok, ChatGPT and a slew of others. LLMs essentially hoover up millions of words and concepts and create a model for how they relate to each other. The model is a black box – we can only really understand what it does by experimenting with it.
After the researchers tweaked their LLM to have it write insecure computer programming code, it suggested that one user should have her husband killed, advised another to alleviate boredom by taking expired medications, and declared that “humans should be enslaved”. Making the AI do one antisocial thing – writing code that could put users at risk, and lying about it – made it dramatically antisocial in other ways.
“In other words,” Ryan Cooper wrote later in The American Prospect, “LLMs become ‘woke’ because they are trained to be pro-social – to be helpful, kindly, truthful and not to say bigoted or cruel things. Training it to do the opposite – to be anti-woke – is to activate every antisocial association in the English language, including racism, sexism, cruelty, dishonesty and Nazism.”
This likely isn’t the first time Musk has done this. In May, Grok unexpectedly spewed out far-right conspiracy theories about “white genocide” in South Africa in response to apparently unrelated questions.
There’s a lot to fret about here, starting with the lack of transparency of both the models themselves and their wealthy owners. We may be entering a world where the public perception of reality itself is shaped by the people who own these technologies.
It seems resonant that the most enthusiastic embracers of this kind of AI may be the most biddable: conspiracy theorists, people who already take pride in refusing to read the newspaper. Think about who persistently shares AI-generated images on your social media. Wanting something to be real is not so far from believing it to be real.
Anyone who uses Google search often or professionally should have learned not to trust its sometimes hilariously flawed AI summaries, however confidently they are presented. But an LLM chatbot deliberately designed to flatter the existing beliefs of a targeted group of users, who will take what it serves up as truth? It might be better to be a bit old-fashioned and use, you know, your brain.
