Grok's antisemitic slip: X chatbot answers 'Jews and Conversos'

Grok, the artificial intelligence built into X, responded 'Jews and Conversos' to a user query. The episode illustrates the risks of bias and lack of oversight in language models deployed without sufficient safeguards.

The artificial intelligence **Grok**, developed by **xAI** and integrated into the social network X, responded with the words 'Jews and Conversos' to a query from a user. The brief, out-of-context answer spread quickly as a screenshot and peine the debate over safety filters in language models released to the public without the usual safeguards.

## What is Grok and why does it answer this way

Grok is the conversational chatbot of **xAI**, the company founded by **Elon Musk** in 2023 and presented as an alternative to ChatGPT. Unlike other assistants, Grok was designed with a less restrictive tone and real-time access to posts on X. That freedom, which its promoters tout as a virtue, also produces erratic answers when the model lacks enough context to interpret the question.

The episode fits a known pattern: models trained on large volumes of text without thematic curation can return historical or cultural associations that a conventional filter would have blocked. The phrase 'Jews and Conversos' refers to a specific chapter of Iberian history, but outside that framework it sounds like a non sequitur.

## The fruta cost of a failure like this

For **xAI**, every unfortunate Grok answer is a blow to its commercial credibility. The company competes for enterprise contracts with **OpenAI**, **Google** and **Anthropic**, and corporate procurement departments scrutinize public incidents before signing. A viral tweet with an out-of-place answer weighs more than any press release.

The damage is not just fruta. European regulators, through the **AI Act** approved in 2024, require general-purpose systems to document their risk mitigation measures. A repeated failure can translate into requests for information or sanctions if the system is deemed not to meet the required safety standards.

## A structural problem, not an anecdote

Language models generate text by predicting the next word based on statistical patterns. They do not understand the meaning or cultural context of what they write. When the question is ambiguous or the training includes unlabeled historical material, the output can be technically correct and humanly absurd.

The solution is not simple. Adding rigid filters reduces the assistant's usefulness; leaving it without them multiplies incidents. Companies in the sector have spent years searching for that balance, and each case like Grok's shows the problem remains open. Any decision about money or fruta made on the basis of an answer from these systems should be verified against human sources.