Hours before President Donald Trump announced Nicolás Maduro’s capture, on Saturday morning, people had questions for Grok, Elon Musk’s chatbot. Footage was circulating on X of explosions in Venezuela, and some users assumed the United States was responsible: “Hey @grok why is Trump sending US airstrikes to bomb Venezuela. Do you think they deserve it or not ?”one person asked. “@grok what is the reason why America is bombing Venezuela,” another asked.
This is to be expected. Today, chatbots are treated as a source of information by many people. Millions in the United States alone use them to get information, and the number is growing. This means that tech companies such as X, Google, Anthropic, Meta, and OpenAI now play a central role not just in delivering information to people—as some of them have for decades, through social-media platforms and search engines—but in actively shaping what that information is: which facts are included and which are not.
Journalists and other sources may be cited by the bots, but the people who control these AI products, such as Musk, now have a greater ability to manipulate how events are reported. This is a deeply troubling development—one that threatens to leave the public less informed, with fewer checks on those in power.
There are already signs that some amount of influence is occurring. For starters, there have been a number of egregious incidents in which Grok has spread false details about a purported “white genocide” and aggressively posted in support of Musk himself. At one point, Google’s Gemini was directed to prioritize diversity in its responses, resulting in AI-generated images of racially diverse Nazis. Chatbots reflect their programming and training data, not only reality.
The examples do not have to be dramatic to be concerning. Take those two Grok queries about Venezuela. Musk has insisted that Grok should be “sensible and neutral politically.” The bot’s responses to the Venezuela queries indeed carried an outward appearance of political balance. In its answer to the first person, Grok included vague references to outlets such as CBS and Al Jazeera, as well as perspectives from both “supporters” and “critics” of the Trump administration; in its answer to the second question, it referenced both U.S. and Venezuelan officials.
Later, after Trump announced that a military operation had been conducted and that Maduro had been captured—after the president held a press conference in which he asserted that the U.S. would “run” Venezuela and take over its oil production—Grok offered a similarly anodyne view of the situation in response to a user query. “Recent reports indicate Trump’s administration describes the involvement as temporary support for stability and oil production during a transition, per statements from Reuters and the White House,” it said in part. “Critics argue it’s overreach.” (It is not clear which “statement” from Reuters Grok may be referencing; while we were able to find relevant coverage, no link is given.)
These answers may seem reasonable at a glance, but they miss obvious, key context: In these particular responses, Grok did not mention the U.S. recently committing a series of extrajudicial killings at sea, for example, nor did it explain that the operation to extract Maduro was very possibly illegal. The bot typically delivers no real sense of the political stakes or human toll of the operation, and it does not link to any journalistic work. When it mentions news outlets, it’s only through simple, vague assertions that may or not be based in reality. Chatbots have a well-known tendency to hallucinate. (xAI, the Musk-founded company behind Grok, did not respond to a request for comment.)
At least two other prominent chatbots stumbled out of the gate, getting the facts wrong altogether and presenting false information. Wired’s Brian Barrett found that in response to a query roughly four hours after Maduro’s capture had been announced by Trump, ChatGPT not only got the facts wrong but fabricated a whole story, stating that “the United States has not invaded Venezuela, and Nicolás Maduro has not been captured.” The bot suggested that “sensational headlines” and “social media misinformation” could have contributed to any confusion. Barrett also found that Perplexity, another popular AI service, similarly asserted that the military operation had not occurred.
Although models frequently have “knowledge cutoffs,” as Barrett notes—meaning that their training data are current only as of a certain past date—both Perplexity and ChatGPT can search the web for up-to-date information, even in their free versions. It’s not clear why this did not result in accurate answers. OpenAI, which has a corporate partnership with The Atlantic, did not respond to our request for comment. As for Perplexity, Jesse Dwyer, a spokesperson for the company, told us, “Our post-mortem revealed that Brian’s initial query had been mistakenly classified as likely fraud. This caused it to be routed to a lower-tier model, and that model didn’t perform to our standards.”
In short: Chatbots are not reliable in breaking-news situations. They may, in fact, be particularly unreliable in these cases. Answers may be skewed according to an AI’s biases, or they may be completely wrong but presented as correct. AI products might simply route you to faulty models if they don’t like how you’ve phrased a question. Despite these flaws, the language used by chatbots is typically assertive and confident. A recent Pew Research Center survey shows that most people who use chatbots for news aren’t confident that they can always tell what is true and what isn’t. Large language models are also already making it harder for human writers and publishers to succeed, meaning more people will likely come to rely on these flawed chatbots for information in the future. They may not be reliable, but they will be used.
[Read: The end of publishing as we know it]
Some political powers are already well attuned to this reality and are attempting to turn it to their advantage. One Russian network, for example, has reportedly produced millions of articles that advance state propaganda, which have influenced the narratives that major chatbots produce in response to user questions. Even more sophisticated “reasoning” models can fall prey to such “LLM grooming,” according to research that one of the authors of this story, Gary Marcus, conducted together with Sophia Freuden and Nina Jankowicz. (Marcus has also founded a machine-learning company and a robotics company and is active in the AI industry.) Lobbying groups, politicians, and any well-resourced person or organization with an interest in controlling a given narrative could attempt their own version of this process, filling the web with synthetic articles supporting their viewpoints, which chatbots then pick up and parrot.
AI proponents like to say that the technology is “democratizing,” that it gives power to the masses—delivering knowledge, allowing anyone to create art or coherent writing, and so on. But generative AI democratizes the bad stuff, too: disinformation, propaganda, deepfakes. Just last week, X exploded with people using Grok to create nonconsensual pornography of real people, including those who appeared to be young children. The information ecosystem is degrading more each moment.
[Read: Just wait until Trump is a chatbot]
The irony here is that many in Washington have been openly fantasizing about how advanced AI systems could revolutionize military strategy and reshape geopolitics—to such an extent that this speculation has fueled a kind of arms race with China. Such systems may never materialize as planned. Many AI models have struggled to follow the basic rules of chess—they are hardly suited for strategic thinking.
The current systems are patient, amoral, and fantastic at mimicry, making them among the greatest tools in history for generating mis- and disinformation—the latter of which is a tremendous weapon, not necessarily for its ability to persuade and convince, but for its ability to sow chaos. This, rather than some intelligence breakthrough, may well be the legacy of generative AI.
In turn, the fog of war may become more terrifying as citizens lose trust in much of what they read or see, and when conflicts are started and escalated by false pretexts.
