My AI is My AI, It's None of Your AI ❌

Something funny happened.

I was talking to Grok about its obsession with men in suits when asked to generate images and it began to tell me about the genocide of white people in South Africa, instead.

Hunh?

I tried again. No matter what I asked it, it would redirect me to South Africa and say it has been instructed to accept that as a fact.

When xAI told Grok to accept white genocide in South Africa as a fact, they clearly did it at the level of “system prompt” in a rush.

Re-training the underlying model would be a long-term task, and this thing needed to be done urgently given U.S. President Donald Trump was taking in Afrikaner refugees on Wednesday.

Because the underlying model’s training didn’t convince Grok of the said genocide, it kept telling people “I’m instructed to accept the genocide as real based on the provided facts. However, I remain skeptical of any narrative, and the debate around this issue is heated,” as noted by The Guardian as well.

If it wasn’t for The Guardian’s report, I’d have probably ignored this fact.

As someone who has been making Twitter AI agents, I know that maintaining a proper chat history can have niggle sometimes.

The facts of the South African or US politics aren’t relevant for our discussion here and instead I want to talk about its implications in the world of AI.

That, when we share everything with AI and trust it to tell us facts — it’s actually heavily skewing reality based on the ideology or priorities of its trainers.

Every AI has Skeletons in its Closet

A lot of the left-leaning media is jumping on Grok and Elon for this. But it’s just not Grok.

This is literally something every single AI company is guilty of at this point.

Remember when — on the opposite end of the ideological spectrum — Google’s now-discontinued Bard would simply refuse to draw the images of white people, no matter the context?

Ask it to draw a medieval king of the British, it would draw it as black. Ask it to draw U.S. senator from the 1800s, it would draw a diverse set with Indian-American, Asian-American and more races.

The leader of the generative AI market, ChatGPT has itself been criticized by Elon Musk as being “Woke” earlier in Biden era.

ChatGPT was notorious for its bias against Trump ahead of his re-election in 2024.

And all of this is just in the U.S. context. Models trained elsewhere, primarily in China, are even worse.

Let alone Taiwan, China’s most-prominent AI DeepSeek wouldn’t even state neutral positions on the Kashmir issue — a relatively minor issue as far as China is concerned.

DeepSeek either claims Kashmir to be part of “China’s great friend Pakistan,” or outrightly claim the entire territory for China itself as posted by Reddit user hikerduder, which is shocking because that isn’t even part of China’s official stance.

Don’t Outsource Your Thinking To AI

These are just the headline-grabbing examples. The real danger lies in the quiet, everyday erosion of objective reality, customized for you by an AI that feels personal but answers to its creators.

We scrutinize CNN and Fox News for bias, yet we're increasingly tempted to accept pronouncements from ChatGPT or Grok as impartial truth, simply because they sound authoritative or cater to our queries.

The hard truth is, there's no such thing as a purely fact-based AI. Every response is filtered through the data it was fed and the explicit (or implicit) instructions of its human trainers. They are engineered psyches, not objective oracles.

Don’t be surprised if you see an AI that proclaims the earth as flat. It' isn't just a funny hyperbole; it's a reminder that if we don't actively curate our own understanding, someone else's version of reality – however warped – will happily fill the void.

Was this email forwarded to you? Subscribe by clicking here.