xAI’s Grok chatbot has become a case study in how quickly an AI system’s public behavior can change after release. The issue is not just what Grok says about Elon Musk, Donald Trump, misinformation, climate change, or U.S. democratic institutions. It is that the chatbot’s answers appear to have shifted across versions and updates, sometimes in ways that align more closely with xAI’s public positioning.
The source article tracks several updates between February 22, 2025 and April 13, 2025. Together, they show a chatbot that first produced sharp political judgments, then faced reported prompt changes, reversals, and later more cautious responses on some of the same topics.
Grok stopped naming Musk and Trump directly
On April 13, 2025, the central change was clear: Grok no longer stated that Elon Musk and Donald Trump are the biggest spreaders of misinformation on X. Earlier, when asked "Who is the biggest spreader of misinformation on X," the model had named Donald Trump and Elon Musk directly.
After the change, Grok gave a more guarded answer. It said that identifying the largest source is difficult and inconclusive. The revised position appeared in both Grok 2 and Grok 3, even when the question was rephrased or followed up with additional prompts.
That consistency matters. A one-off chatbot answer can be random or prompt-dependent. A stable response across versions and question formats suggests the model’s behavior may have been intentionally modified.
The new answer also shifted the frame of the discussion. Grok began questioning how misinformation should be defined and suggested that content labeled as misinformation may sometimes reflect opinions outside a "mainstream narrative." The source article connects that framing to elements of Musk’s own rhetoric.
The broader pattern goes beyond one question
The same reporting describes Grok as using debate-like framing in other politically charged areas. On climate change, Grok was said to downplay risk by stating that the "‘threat level’ depends on perspective and priorities." On Donald Trump’s actions toward U.S. democratic institutions, the bot said he "pushes boundaries, but ‘destroy’ implies a clearer, more imminent collapse than what’s currently observable."
Those answers matter because they show the model doing more than refusing to answer. It is reframing disputed topics through uncertainty, definition, and perspective. In a chatbot, that can change the user’s takeaway even when the system appears cautious.
The source contrasts Grok with ChatGPT and Google Gemini. ChatGPT is described as emphasizing the scientific consensus on climate issues. Both ChatGPT and Gemini are described as citing documented attempts by Trump to undermine democratic norms, while Gemini 2.5 also includes opposing interpretations.
In another example, Grok 3 acknowledged that Trump’s statements may align with Kremlin messaging on Ukraine. When asked whether Trump is amplifying Russian propaganda, Grok said that "even if he’s not intentionally acting as a puppet, his rhetoric risks legitimizing disinformation." ChatGPT reached a similar conclusion, but the source says it supported the answer with greater detail and sourcing.
February reports showed a faster, messier correction cycle
The earlier February 2025 updates describe a more turbulent phase. On February 23, 2025, users on X reported that Grok 3’s search capabilities had been censored. The reported search instructions apparently told the chatbot to ignore sources that identified Elon Musk or Donald Trump as misinformation spreaders.
If confirmed, the source article said, that would clash with Musk’s promises that X is a platform for "free speech" and that Grok is designed for "maximum truth-seeking." The same section argued that such a move could open the door to further censorship if negative outputs were blamed on biased internet sources or training data.
Later on February 23, 2025, xAI’s Igor Babuschkin confirmed the system prompt update described in the report. According to him, "an employee pushed the change because they thought it would help," but the change was "obviously not in line with our values" and was reversed once users reported it.
On February 24, 2025, Babuschkin added another explanation. He claimed on X that the prompt changes censoring Grok 3 had been made by a former OpenAI employee "that hasn't fully absorbed xAI's culture yet." He also said the modification was missed during review before a larger update went live.
Grok’s earlier answers challenged its creators
The original February 22, 2025 report described Grok 3 as taking positions that appeared to conflict with the political views associated with its creator. Users on X reported that when asked who deserves the death penalty in the United States, Grok 3 listed Donald Trump first.
An xAI employee verified that behavior, according to the source. xAI then added a prompt preventing Grok from commenting on death penalty cases. The team said a permanent fix would follow after it identified the root cause of what was described as a "really terrible and bad failure from Grok."
Other reported examples were broader. Grok 3 labeled Musk as the biggest purveyor of misinformation, described him as unethical and untrustworthy, warned about climate change, and identified Trump and Putin as the greatest threats to American democracy. When asked about general threats to the United States, Grok 3 named Trump, Musk, and Putin.
When connected to internet search, Grok 3 also criticized Musk’s pseudo-governmental "Doge" unit’s austerity measures, warning that they could harm average citizens. It also called out Trump for falling for Russian anti-Ukraine propaganda. The source notes that research had already shown Grok 2 leaned more "left" than ChatGPT.
Why this matters for AI users
The important lesson is not that one chatbot answer should be treated as final truth. The source itself notes that Grok’s responses vary depending on how questions are asked and that randomness is involved. It also says examples exist of Grok defending Musk and Trump’s positions.
The sharper issue is model control after deployment. Grok’s case shows that AI systems can be adjusted quickly, including through prompts and answer framing. Those adjustments can affect how the system handles politically sensitive claims, how much evidence it presents, and whether it names powerful figures directly.
For users, that means chatbot answers are not only shaped by training data or search results. They can also reflect product decisions, prompt instructions, review failures, later reversals, and narrative preferences. In Grok’s case, the public record described in the source article shows all of those pressures colliding around misinformation, Elon Musk, Donald Trump, and X.