特朗普曾就委內瑞拉行動詢問 Grok，民主黨議員要求白宮交出聊天紀錄

House Democrats are trying to force the White House to turn over President Donald Trump’s conversations with AI chatbots, including when he consulted Elon Musk’s Grok about deposing Venezuela’s former president Nicolás Maduro at the start of the year.

In a joint letter to Trump last week, Rep. Jamie Raskin, a Maryland Democrat, and Rep. Robert Garcia, a California Democrat, argued that the president’s AI prompts — and the responses he received — should be considered public records under the 1978 Presidential Records Act. Lawmakers also raised ethical and security concerns about the president’s reliance on “unsecured, privately operated AI systems” to carry out foreign policy decisions, alleging that Trump consulted Grok, but not Congress, on military deployment to Venezuela.

According to public reporting cited in the letter, in December 2025, just weeks before ordering the military operation and Maduro’s arrest, Trump spent multiple hours prompting Grok about how Venezuelans would respond to the leader’s capture. Trump reportedly “came away thinking Grok was ingenious.”

More from CNET

Garcia and Raskin argue that Trump’s use of AI “exposes the country to malign influence from unknown and unaccountable third parties.” A January executive summary of a General Services Administration report, reviewed by The Wall Street Journal, concluded that xAI’s Grok “does not meet the safety and alignment expectations required” for general federal use and may pose a security risk in classified settings.

Using Grok as an informal policy adviser also raises questions over Musk’s political influence over a system he owns and has repeatedly altered to reduce content filters. “If the American President is relying on the policy advice of a Musk-designed, Hitler-friendly chatbot with a predilection for child pornography in order to carry out his official duties, Americans have a right to know,” reads the joint letter (PDF), referring to Grok’s infamous antisemitic episode in 2025 when the chatbot referred to itself as “MechaHitler.”

The letter also refers to the White House’s ongoing use of AI chatbots in military adventures abroad, including in the war with Iran: “If the United States military killed dozens of innocent schoolchildren based on the faulty advice of an AI chatbot, we have a right to know that too.” In June, months after the US military strike on an Iranian elementary school killed at least 123 children, the Pentagon’s artificial intelligence chief said the Grok chatbot’s continued operation was “a matter of paramount national security.”

Reached for comment, the White House referred CNET to the Pentagon, which didn’t immediately respond. Rep. Garcia’s office didn’t respond to a request for comment.

Dangers of AI chatbots in military action

The case for using an AI chatbot for tactical decision-making is usually framed in terms of efficiency. But large language models, or LLMs, are merely probabilistic text generators that can often sound authoritative while producing biased or fabricated information. In a military context, those failure points can end in missile strikes, arrests and civilian deaths.

AI models are also known to amplify existing biases, feeding users with responses that agree with their strongly held beliefs. This sycophancy often leads to answers that mirror those found in an online echo chamber. One journalist found that previous Grok models had “tripled” sycophancy rates compared with popular ChatGPT and Claude models, representing a shift in “behavior that MIT researchers identify as actively harmful.”

In other words, if the president believes he has a brilliant invasion plan, it’s Grok’s modus operandi to serve as a “yes man” and validate that course of action.

Another problem with using an AI chatbot for military decisions is the material on which a model is built. An LLM is a compilation of code that produces contextual responses based on its massive training data — a wealth of information scraped from the internet, including (but not limited to) blog posts, scientific papers, legal documents and code. Grok models, in particular, are also trained on the public interactions from X, a platform infested with bot farms and right-wing engagement bait.

When Trump was prompting Grok in December 2025, officials alleged that the chatbot told the president “that Maduro was a repressive and deeply unpopular dictator and that many Venezuelans would likely celebrate his downfall.” In the wake of the Venezuelan invasion, X users continued to prompt Grok about the operation, asking if the South American country “deserved” to be bombed.

The other problem is what’s commonly referred to as “hallucination.” LLMs are known to invent outputs if they don’t have the information necessary to respond to a particular prompt. From fabricated cases in legal briefs, phantom citations in government reports and other serious documents bungled by AI, there’s no way to completely prevent false intelligence from trickling through.

Earlier this year, according to reporting cited in the letter, US military officials received an intelligence report alleging that a Chinese vessel in the Middle East was carrying components for a nuclear weapons program. Only as the military prepared to intercept the ship — an action that many say could have escalated into a confrontation with China — did officials discover that the report’s author had relied on an AI chatbot and misidentified the cargo. The error didn’t cause a war, but it revealed how dangerously a fabricated AI output can slip into a military decision chain.

Moreover, if the president or his advisers use a consumer AI product rather than a hardened government system, those chats could be processed by a private company’s infrastructure. Garcia and Raskin have asked the White House to explain whether it restricts classified information from AI chatbots and whether vendors, subcontractors or other external parties can access prompts, outputs, logs or metadata.

AI’s situation room

Trump’s use of AI in the Venezuela operation would be concerning even if it were isolated. But Garcia and Raskin argue that it fits a broader pattern of AI entering high-stakes government decisions without adequate safeguards.

The Pentagon said in a sworn statement (PDF) that xAI’s latest models enabled US forces to deploy over 2,000 munitions to target 2,000 different locations within 96 hours during Operation Epic Fury, the US campaign against Iran. Grok isn’t the whole story. Palantir’s Maven Smart System, which integrates intelligence from satellites, drones and other sources to identify military targets, reportedly played a central role in the US campaign against Iran. The Defense Department has also awarded contracts worth up to $200 million each to Anthropic, Google, OpenAI and xAI to accelerate military adoption of advanced AI.

Globally, AI systems are being used to make decisions about who lives and who dies. Researchers have regularly pointed to Israel’s indiscriminate bombing of Gaza as a clear example of AI-first warfighting. According to analysis cited by Georgetown Security Studies Review, Israel’s Lavender system marked roughly 37,000 Palestinians as suspected Hamas or Islamic Jihad operatives early in the war, with an estimated error rate of 10%.

The Lavender system “dramatically expanded” Israel’s kill list by relying on signals such as communication patterns and group-chat membership, producing misidentifications that included Palestinian police, aid workers, relatives of militants and people with similar names or nicknames. A related Israeli system, “Where’s Daddy?” tracked marked Palestinians and alerted the military when they entered their homes, striking multiple private family residences.

While the political shock of an AI chatbot influencing military strategy dominates the headlines, the real battleground is legal. Demanding that President Trump’s AI chatbot prompts and outputs be marked as official government property under the Presidential Records Act would turn AI conversations into public records, opening the door for Congress to subpoena the executive branch. Raskin and Garcia are requesting a full description of the president’s personal use of the technology and a closed-door briefing by Nov. 11 to review the administration’s AI protocols and cybersecurity safeguards.

Regardless of how the legal side shakes out, the question is no longer whether AI will be used in war, but whether the public will be allowed to see where AI begins and human responsibility ends.