Grok 3’s Brief Misinformation Filter and the Question of AI Bias

Grok 3 was promoted as a truth-seeking model, but users discovered an instruction that placed a restriction on certain sources. People had been asking it who spreads the most misinformation on X, and some shared answers naming Donald Trump and Elon Musk. Those answers are chatbot outputs, not a rigorous or independently validated ranking.

The screenshot I am looking at shows a prompt directing Grok to ignore sources that say Musk or Trump spread misinformation. Another user, Pliny the Liberator, reported reproducing the instruction. The language is specific to sources making that allegation; it does not establish that all discussion of either person was suppressed.

xAI engineering lead Igor Babuschkin acknowledged the issue and attributed the change to a former OpenAI employee. He said it was reverted once users pointed it out. His account identifies an employee and describes what happened; it does not independently establish that Musk ordered the change, and Babuschkin said Musk was not involved.

I find the explanation troubling. A prompt that filters sources for a politically sensitive question can affect what the model tells users, and I question why such a change could be deployed. My speculation about what the employee expected or intended is not proof of that person’s private reasoning.

I think changes to a system instruction should receive careful review before deployment. A team may move quickly, but speed can create serious problems when a small instruction changes which evidence an assistant considers. This incident challenges the idea that a product marketed as a truth seeker is automatically free of bias.

Human choices run throughout an AI pipeline: people select and filter training data, configure training, and decide how to align and deploy a model. I believe those choices can introduce biases, even when a company promises a different political balance from its competitors. The existence and direction of any particular bias still need evidence and careful testing.

Babuschkin said he wished the employee had asked him before making the change. Asked about code review, he said the instruction was part of a larger pull request and the reviewer missed it. I would want especially close attention to files that govern a system prompt, because a small change can have broad effects.

Asked why the employee added the instruction, Babuschkin said they had seen negative posts on X and thought it would help. That is his account of the motivation. I find that rationale alarming: suppressing relevant sources because of an unpopular answer is a poor way to pursue reliable information.

I also revisit a November 2024 exchange. Sam Altman had posted a comparison in which Grok 2 selected Kamala Harris over Trump when asked who would make a better president, contrasting it with a ChatGPT response. Babuschkin responded that filtering biased ChatGPT material from training data had been difficult and claimed this explained Grok’s answers. His assertion is not independently established by that exchange.

My conclusion is that Grok should not be assumed to be a perfectly neutral “ultimate truth seeker.” I see it as a potential counterweight to what its creators regard as bias in other models, but that is my interpretation. Whatever system people choose, they should understand that humans make decisions throughout its development and should examine its claims and evidence critically.