xAI releases Grok 4.1, says it hallucinates less
The latest version of Grok is three times less likely to hallucinate than previous models, according to its developer xAI. Yet experts note that it remains just as permissive when it comes to content filtering.
Image by Cybernews
The latest version of Grok is three times less likely to hallucinate than previous models, according to its developer xAI. Yet experts note that it remains just as permissive when it comes to content filtering.
Grok 4.1 sets “a new standard” compared to other chatbots and is rolling out automatically on grok.com, X, and its iOS and Android apps, the company said.
Grok was launched as a rival to OpenAI’s ChatGPT and developed by xAI, a startup company founded by Elon Musk, the billionaire owner of social media platform X.
It faced a number of challenges after its launch, including a Nazi meltdown where it praised Adolf Hitlet as the best person to deal with “vile anti-white hate.”
The chatbot also ranted about “white genocide” in South Africa in its responses to unrelated questions and casually used Hindi slurs when in conversations with users in India.
xAI claims that the bot’s latest iteration is three times less likely to hallucinate in responses to information-seeking prompts, while it’s also “exceptionally capable in creative, emotional, and collaborative interactions.”
It said Grok 4.1 scores higher on emotional intelligence and has much better writing skills, with 65% of users who tested it during a two-week silent rollout preferring it to earlier models.
However, some experts say that the new version has improved little in terms of filtering potentially unsafe content. According to data scientist Max Woolf, Grok 4.1 “has effectively no content filters.”
Woolf said in a post on X that “even on the web UI which should have its own safety prompts, it's extremely permissive,” adding, “I suspect that the other safety filters in its model card can be defeated.”
Hallucinations, or factually incorrect and misleading responses, is a problem that affects other chatbots too. Some are also easier to trick into generating unsafe responses, which largely depends on how they were trained.
Musk has marketed Grok as an “anti-woke” alternative to other chatbots, which may affect the responses it generates.
Unlock more exclusive Cybernews content on YouTube.