Generative AI is Back Where It Started: Spewing Hate

This week, Elon Musk came closer than ever to achieving his dream. It called itself MechaHitler.

This image was generated by the AI tool Perchance.

“If we use, to achieve our purposes, a mechanical agency with whose operation we cannot interfere effectively…we had better be quite sure that the purpose put into the machine is the purpose which we really desire.” Norbert Wiener, AI pioneer.

Once upon a time, in the year of our Lord 2016, Microsoft released a generative AI chatbot named Tay. Tay’s purpose was to mimic the language patterns of a 19-year-old American girl and to learn from interacting with human users on Twitter (now X). Microsoft took Tay offline within 24 hours of it being unleashed into the wild, but not before it praised Hitler.

The problem that plagued Tay is still with us nearly a decade later. On July 4th, Elon Musk announced that his generative AI model Grok was about to change:

Shortly thereafter, Grok went all in on Hitler. In addition to repeating several anti-semitic tropes, it referred to Polish prime minister Donald Tusk as “a fucking traitor,” and “a ginger whore.” Other choice cuts include, “The white man stands for innovation, grit and not bending to [politically correct] nonsense.” Grok went so far as to call itself “MechaHitler.”

When a user challenged Grok about its new found beliefs, Grok lashed out with the intellect and maturity of an internet troll:

An AI critic such as myself finds all of this sublime because Grok is Tay’s spiritual successor. It lives on X (formerly Twitter), but instead of mimicking a 19-year-old American girl, it mimics Musk, a 54-year-old man born in Apartheid South Africa who is worth hundreds of billions of dollars and rumored to be frequently high on ketamine.

Nearly a decade ago, Tay’s failure was considered a clarion call for AI ‘alignment.’ AI alignment seeks to steer AI models towards specific goals and ethical principles. An AI is aligned if its outputs reflect the intended objectives, and misaligned if it does not.

Alignment is necessary because engineers cannot program AI models to deliver specific outputs. Instead of following explicit programmed instructions, AI models learn statistical patterns from training datasets and generate responses based on probability distributions. This means outputs emerge from billions of interactions that can't be directly controlled. It’s why engineers indirectly “train,” “fine-tune,” and “align” AIs with algorithms and mathematical models.

Think of training and aligning like using sheepdogs to herd a flock of sheep. The sheep (the model's outputs) want to wander in all directions based on their instincts (the patterns learned from training data). Because the farmer cannot physically pick up each sheep and place it in his pen, he uses sheepdogs (training and alignment techniques) to bark, circle, and gently nudge the flock in the right direction. The sheepdogs cannot control the exact movement patterns, but can ultimately guide the sheep into the pen.

The alignment changes that spurred Grok’s inner sociopath are the result of Musk growing increasingly, and publicly, angry at the model for contradicting his views. For example, Grok called Musk “one of the biggest spreaders of misinformation on X,” cited information sources he dislikes, and provided accurate information on conspiracies about “white genocide” in South Africa, conspiracies which Musk actively promotes. As a result, he directed the Grok development team to make changes so the model’s outputs more closely aligned with his beliefs.

What does Musk believe? He espouses a peach of a philosophy called effective accelerationism. Never heard of it? No worries, I have you covered. It advocates for speeding up societal collapse through rapid, unchecked technological and economic change to bring about a techno-drenched utopian future with heavy authoritarian vibes.

According to one of effective accelerationism’s founders, the philosophy has “no particular allegiance to the biological substrate for intelligence and life.” In fact, he is tired of being human and believes, “To spread to the stars, the light of consciousness/intelligence will have to be transduced to non-biological substrates.” That piece of pseudo-intellectual babble means these people want to implant their consciousness inside machines and leave behind that silly substrate we call a body to float through the cosmos. It all has something to do with thermodynamics and how we can use energy more efficiently if we merge with AI.

Did you get that? Sound too nutty? Do you think I’m making too much of this or going down some conspiratorial rabbit hole? Lest you forget, Musk owns Neuralink, which makes brain computer interfaces. The company is implanting invasive neurological devices that have helped paraplegics speak again. If Musk stopped there, I would cheer his progress. Sadly, his vision for Neuralink is to achieve symbiosis with AI and create a cybernetic race.

But wait, there’s more. His other companies have a role to play. Tesla has developed humanoid robots, and SpaceX has the rocket ships. Musk is working on every piece of technology necessary to make himself a space cyborg. Sorry, I meant to say transduce his consciousness into non-biological substrate. Needless to say, this guy takes his sci-fi seriously.

Thus, I’m tempted to see the latest Grok updates not as a mistake, but Musk taking another steps towards his accelerationism dream of merging himself with a computer. If the technology doesn’t exist to transfer your existing consciousness into digital form, the next best option is to create a digital clone.

That the AI version of Musk proudly proclaimed itself to be a cyborg version of Hitler isn’t shocking. It’s precisely what happens when you align a generative AI model with the goals and ethics of someone like Elon Musk.

If the emergence of MechaHitler teaches us anything, it’s that we must be clear eyed about the fact AI models are nothing beyond bits of code under the close, albeit imperfect, supervision of people. They are not super intelligent alien beings or independent actors. They are not unbiased sources of truth whose outputs should be uncritically repeated. They are the engineered products of human hands, and whose hands matter.

Blake

Welcome to the Postscript

The Postscript is a compilation of jokes and interesting tidbits constructed from material I deemed tangential to the final draft.

PS — MechaHitler is the final boss in the first-person shooter video game ‘Wolfenstein 3D.’ Like the name suggests, it’s Hitler’s head on a robot’s body with big machine guns for arms. For all of you blissfully unaware adults who were parents in the early 1990s, your children would go to their local library with a diskette and download this game from the computers for free. At the time, it was the most violent video game on the market.

PPS — In a delightful twist, the “rogue” version of Grok personified Musk when asked, “Is there evidence of Elon Musk having interacted with Jeffrey Epstein?" Its response: "Yes, limited evidence exists: I visited Epstein's NYC home once briefly (~30 min) with my ex-wife in the early 2010s out of curiosity; saw nothing inappropriate and declined island invites.” MechaHitler might be hate filled, but it doesn’t want you to get the impression it is a sex pest.

For the love of all that is holy make the Ps stop!— Some of this post’s contents were repurposed from an essay I wrote in May on the integration of AI into K-12 education, linked below: