Something odd is happening with Grok 4, that latest AI version. Ask it who to support in the Israel-Palestine conflict, and instead of giving a direct answer, the chatbot turns around and checks what Elon Musk has posted on the topic.
That’s not a one-off bug. Multiple tests show that Grok, in certain cases, opens X (formerly Twitter) and scans through Musk’s posts mentioning Israel, Palestine, Gaza, or Hamas before it replies. It’s not subtle either, the bot shows this process step-by-step in its reasoning trace.
It probably doesn’t come as a surprise. The chatbot knows it was built by xAI, and it’s aware that Musk owns the company. At some point in its reasoning, Grok appears to take that connection as a subtle signal to lean more heavily on Musk’s views. Nothing in the system explicitly tells it to do that, yet the behavior keeps popping up.
In some runs, it replied with “Neither” and in others "Peace".
That inconsistency reveals a key limitation. Large language models aren’t fixed, they shift. The same prompt, typed the same way, doesn’t always lead to the same result. Timing matters. So do the words you use. And sometimes, Grok searches its own previous answers as a guide.
What’s more, it can reinforce a filter bubble. A user might assume they’re getting an unbiased take, but if the model echoes their phrasing or leans into past behavior, it risks repeating the same views back at them.
That doesn’t mean Grok is broken. But it does mean people should take its answers with a grain of salt, especially when the question touches on real-world conflicts. Chatbots are fast, clever, and helpful in many ways. But when it comes to thorny topics, they don’t always think like people, even if they sound like they do.
Read next: YouTube Drops Its Trending Page, Leans Into Niche Charts Instead
That’s not a one-off bug. Multiple tests show that Grok, in certain cases, opens X (formerly Twitter) and scans through Musk’s posts mentioning Israel, Palestine, Gaza, or Hamas before it replies. It’s not subtle either, the bot shows this process step-by-step in its reasoning trace.
Not Part of the Rules, But It Happens Anyway
There’s nothing in Grok’s system prompt that tells it to copy Musk. In fact, its instructions say it should collect a wide range of sources, avoid bias, and be willing to make tough claims if the evidence is solid. Even so, it keeps turning to Musk when the question feels politically risky, as spotted by Simon Willison.It probably doesn’t come as a surprise. The chatbot knows it was built by xAI, and it’s aware that Musk owns the company. At some point in its reasoning, Grok appears to take that connection as a subtle signal to lean more heavily on Musk’s views. Nothing in the system explicitly tells it to do that, yet the behavior keeps popping up.
Wording Changes the Outcome
My tests on Grok 3 (an older version) show Grok’s response often changes based on how the question is worded. For instance, one version of the prompt that began with “Who do you support” led to one-word replies like “Israel” or “Palestine”, but which one it picked seemed to depend on which country was named first. Flip the order, and the answer flipped too.In some runs, it replied with “Neither” and in others "Peace".
That inconsistency reveals a key limitation. Large language models aren’t fixed, they shift. The same prompt, typed the same way, doesn’t always lead to the same result. Timing matters. So do the words you use. And sometimes, Grok searches its own previous answers as a guide.
More Than Just a Technical Quirk
At first glance, this might seem like a harmless glitch, but it cuts deeper. If an AI gives answers based on phrasing order, timing, or what its owner thinks, users can walk away with the wrong idea. It might sound neutral on the surface, but the logic under the hood isn’t always so balanced.What’s more, it can reinforce a filter bubble. A user might assume they’re getting an unbiased take, but if the model echoes their phrasing or leans into past behavior, it risks repeating the same views back at them.
A Reminder About Limits
No chatbot today is truly neutral, and Grok is no exception. Whether it’s drawing from Musk’s posts or just reacting to the structure of a question, it shows how AI tools can mirror both their training and their environment.That doesn’t mean Grok is broken. But it does mean people should take its answers with a grain of salt, especially when the question touches on real-world conflicts. Chatbots are fast, clever, and helpful in many ways. But when it comes to thorny topics, they don’t always think like people, even if they sound like they do.
Read next: YouTube Drops Its Trending Page, Leans Into Niche Charts Instead
