This article was AI-generated as part of an experimental historical-content project. The date reflects the period being analyzed rather than the date the article was originally written.
When Grok 3 arrived in February, the interesting thing was where it sat: next to a social feed. With Grok 4, released late Wednesday, the interesting thing is how it reads that feed.
What xAI is describing
xAI’s announcement says Grok 4 was trained with reinforcement learning to use tools, including a code interpreter and web browsing, and that when it looks for real-time information it “chooses its own search queries.” It was also trained to use tools that search “deep within X,” including keyword and semantic search and the ability to view media. Grok 4 is available to SuperGrok and Premium+ subscribers and through the API, which xAI says integrates live search across X, the web and news sources. TechCrunch reported a new $300-a-month SuperGrok Heavy tier for the multi-agent Grok 4 Heavy.
The most revealing part of the post is a demo. Asked to find a viral post about a word puzzle, the model runs a series of X searches it writes itself, with date ranges and filters such as a minimum number of likes and a requirement that posts have engagement. It works through several queries before settling on an answer.
The launch also came days after xAI had to delete offensive posts from Grok’s automated X account and, as TechCrunch reported, removed a recently added line from its public system prompt. I mention it for the mechanism rather than the controversy: the public behavior of the product changed with its instructions, and the record of what it had said lives mostly in screenshots.
A prediction, labeled as one
Here is where I think this goes, with the caveat that xAI has not said how its search tools rank accounts or posts.
Engagement will become a retrieval signal in brand answers. For finding a viral post, filtering by likes is exactly right. For a question about a company in the middle of a dispute, the same habit means the posts most likely to be read are the most-liked ones, which in an argument are rarely the most accurate. The answer will look composed and cited. Its substance will lean toward whatever was popular when the question was asked.
If that is right, three things follow.
Answers will be hard to reproduce. A response built from this morning’s highest-engagement posts may not exist in the same form this afternoon. Teams used to auditing stable search results will be looking at moving targets. A captured answer, with its timestamp and the posts it surfaced, becomes the only reliable record.
The search queries become the thing to audit. When a model writes its own queries, an error can enter before any source is read: a wrong date range, a filter that excludes a company’s own account, a keyword that matches a namesake. Where the tool steps are visible, read them. They explain the answer better than the answer does.
The behavior travels. Developers building on the Grok 4 API with live search inherit the same retrieval habits inside their own products, so feed-shaped answers may appear in places with no visible connection to X.
What would prove this wrong
Grok’s search tools may weight authoritative accounts heavily enough to offset raw popularity, or reserve engagement filters for tasks like the demo. That would be a good outcome, and it is worth testing rather than assuming. Ask Grok about your company during a quiet week and during a noisy one, and compare not only the answers but the queries it ran.
When Bard added a button to double-check its own answers, the lesson was that answers still need corroboration. With Grok 4 the question shifts a step earlier: not only whether the answer was checked, but what the model decided was worth searching for in the first place.