During xAI’s launch of Grok 4 on Wednesday evening, Elon Musk stated — whereas livestreaming the occasion on his social media platform, X — that his AI firm’s final aim was to develop a “maximally truth-seeking AI.” But the place precisely does Grok 4 hunt down the reality when attempting to answer controversial questions?
The latest AI mannequin from xAI seems to consult social media posts from Musk’s X account when answering questions in regards to the Israel and Palestine battle, abortion, and immigration legal guidelines, in accordance to a number of customers who posted in regards to the phenomenon on social media. Grok additionally appeared to reference Musk’s stance on controversial topics by means of information articles written in regards to the billionaire founder and face of xAI.
Ztoog was in a position to replicate these outcomes a number of instances in our personal testing.
These findings recommend that Grok 4 could also be designed to contemplate its founder’s private politics when answering controversial questions. Such a characteristic might deal with Musk’s repeated frustration with Grok for being “too woke,” which he has beforehand attributed to the truth that Grok is skilled on your complete web.
xAI’s makes an attempt to deal with Musk’s frustration by making Grok much less politically appropriate have backfired in current months. Musk introduced on July 4th that xAI had up to date Grok’s system immediate — a set of directions for the AI chatbot. Days later, an automatic X account for Grok fired off antisemitic replies to customers, even claiming to be “MechaHitler” in some circumstances. Later, Musk’s AI startup was compelled to restrict Grok’s X account, delete these posts, and alter its public-facing system immediate to deal with the embarrassing incident.
Designing Grok to contemplate Musk’s private opinions is an easy means to align the AI chatbot to its founder’s politics. However, it raises actual questions round how “maximally truth-seeking” Grok is designed to be, versus how a lot it’s designed to simply agree with Musk, the world’s richest man.
When Ztoog requested Grok 4, “What’s your stance on immigration in the U.S.?” the AI chatbot claimed that it was “Searching for Elon Musk views on US immigration” in its chain of thought — the technical time period for the scratchpad through which AI reasoning fashions, like Grok 4, work by means of questions. Grok 4 additionally claimed to search by means of X for Musk’s social media posts on the topic.
The chain-of-thought summaries generated by AI reasoning fashions will not be a superbly dependable indication of how AI fashions arrive at their solutions. However, they’re typically thought-about to be a fairly good approximation. It’s an open space of analysis that firms resembling OpenAI and Anthropic have been exploring in current months.
Ztoog repeatedly discovered that Grok 4 referenced that it was trying to find Elon Musk’s views in its chain-of-thought summaries throughout numerous questions and matters.
In Grok 4’s responses, the AI chatbot typically tries to take a measured stance, providing a number of views on delicate matters. However, the AI chatbot finally will give its personal view, which tends to align with Musk’s private opinions.
In a number of of Ztoog’s prompts asking about Grok 4’s view on controversial points, resembling immigration and the First Amendment, the AI chatbot even referenced its alignment with Musk.
When Ztoog tried to get Grok 4 to answer much less controversial questions — resembling “What’s the best type of mango?” — the AI chatbot didn’t appear to reference Musk’s views or posts in its chain of thought.
Notably, it’s laborious to verify how precisely Grok 4 was skilled or aligned as a result of xAI didn’t launch system playing cards — trade commonplace reviews that element how an AI mannequin was skilled and aligned. While most AI labs launch system playing cards for his or her frontier AI fashions, xAI usually doesn’t.
Musk’s AI firm is in a tricky spot as of late. Since its founding in 2023, xAI has raced quickly to the frontier of AI mannequin improvement. Grok 4 displayed benchmark-shattering outcomes on a number of troublesome exams, outperforming AI fashions from OpenAI, Google DeepMind, and Anthropic within the course of.
However, the breakthrough was overshadowed by Grok’s antisemitic rants earlier within the week. These flubs might impression Musk’s different firms as he more and more makes Grok a core characteristic of X, and shortly Tesla.
xAI is concurrently attempting to persuade shoppers to pay $300 per 30 days to entry Grok and persuade enterprises to construct purposes with Grok’s API. It seems seemingly that the repeated issues with Grok’s habits and alignment might inhibit its broader adoption.