The America.gov chatbot dodged political questions, then answered them again

Fact-check · Government AI

Verdicts

• MisleadingHeadline: “Trump Reprograms Government AI Chatbot to Stop Fact-Checking His Lies.”The bot’s answers did change from debunks to refusals. Nothing in the record shows Trump, or any named official, ordered a change or stated that purpose, and the refusals were already being reversed when TNR published.
• UnprovenIt appears someone at the General Services Administration tweaked the chatbot to prevent it from debunking the president’s lies.GSA does operate America.gov. The abrupt, standardized refusals fit a configuration change, but GSA has not explained them, and a former federal IT official quoted by Axios listed caching and differences between the two AI models as possible causes.
• Mostly TrueOn launch day the bot said official findings did not show widespread 2020 voter fraud; on Wednesday it refused to answer.Dale’s screenshots show both. TNR’s quotation of the first answer alters its words: the bot said “Official federal findings” and “the 2020 presidential result.”
• TrueThe bot first said Trump did not inherit inflation at a historic high, then gave the same refusal.Dale’s screenshots show “No... 12-month CPI-U inflation was 3.0%, not a historic high,” followed the next day by the stock refusal.
• MixedThe bot first cited a CBO estimate that 10.4 million people would lose health insurance by 2034, then refused the question the next day.The bot did cite CBO, but the screenshot TNR linked says about 10.9 million, CBO’s figure for the House-passed bill. TNR’s “next day” refusal has no screenshot, and on Oct. 1 the same prompt drew a full answer.
• TrueAsked Bobby Kogan’s one-word question, the bot first said “Yes,” and later declined to assess how the law affected income groups.Kogan’s screenshot shows “Yes.” On Oct. 1 his exact prompt drew a refusal close to the one TNR quotes.
• Mostly TrueThe bot first cited a Lancet study estimating 4.5 million additional deaths of children under 5 by 2030 from USAID cuts, then refused.The screenshot and the Lancet paper both check out. The refusal is TNR’s own unpublished test; on Oct. 1 the original prompt was answered and a reworded one was refused.
• MixedTNR found that several queries that originally yielded fact-checks now summoned canned answers.Refusals were widely documented that morning, but they were inconsistent across users and times, and Dale reported answers returning before TNR published.
• UnprovenImplied: the government deliberately reprogrammed the chatbot to suppress answers that contradict Trump.The timing invites that reading. No document, statement or official confirms intent, and GSA has not responded to reporters.
• Mostly TrueImplied: the chatbot’s original launch-day answers were accurate.The fraud, inflation, income-distribution and Lancet answers match primary sources. The Medicaid answer used CBO’s estimate for the House bill, not the enacted law (10.0 million).
• MisleadingImplied: the chatbot can no longer answer these questions.TNR wrote that the bot “now replies” with a refusal. It was answering some of the same questions again before publication, and on Oct. 1 it answered the voter-fraud, Medicaid and USAID prompts as originally worded.

What TNR wrote

The article, posted Sept. 30 at 10:35 a.m. Eastern under the headline “Trump Reprograms Government AI Chatbot to Stop Fact-Checking His Lies,” is a short item by associate writer Edith Olmsted.^[1] Its dek says “Trump officials seem to have realized their AI chatbot was correcting the president’s biggest lies.” The body opens: “It appears that someone at the General Services Administration has tweaked the Trump administration’s new America.gov chatbot to prevent it from debunking president’s many lies.”^[1]

It then lists five before-and-after examples, two from Dale’s posts on X and three from other X users, and adds that “The New Republic found that several queries that originally yielded damning-fact checks now summoned similarly canned answers that undermine the usefulness of the tool altogether.”^[1] TNR describes itself as founded in 1914 “as a call to arms for public-minded intellectuals advocating liberal reform.”^[26]

What the chatbot said, and when

America.gov went live on Tuesday, Sept. 29, “around 10 a.m. ET,” according to CNBC.^[12] The site’s About page says “Phase 1 launched on September 29, 2026.”^[30] The record that follows comes from screenshots and timestamps we retrieved, not from TNR’s summary. Times are UTC; Eastern time is four hours earlier.

Axios did not give a time zone for “around 9:15am.” AP’s article was last updated at 5:49 p.m. UTC on Sept. 29; we could not determine whether the in-event change was in its first version.

The 2020 fraud answer

TNR quotes the first answer as “No. Official findings did not show widespread voter fraud that changed the 2020 presidential election result.”^[1] The screenshot has “Official federal findings” and “the 2020 presidential result.”^[2] The meaning is unchanged, but the words inside TNR’s quotation marks are not the bot’s. The refusal TNR quotes matches the screenshot.

The inflation answer

Dale’s post says the bot gave the same refusal on Sept. 30.^[3] TNR’s account matches.

The Medicaid answer

TNR wrote that the bot cited CBO’s “estimate that 10.4 million people would lose health insurance by 2034.”^[1] The screenshot says 10.9 million.^[6] The CBO letter TNR links reports the same 10.9 million for “provisions that would be enacted under H.R. 1” as passed by the House.^[17] We found no 10.4 million figure in either document.

The Kogan and USAID answers

Kogan, senior director of federal budget policy at the Center for American Progress, a progressive think tank,^[27] asked a yes-or-no question about whether the law “on average made the poorest people in America even poorer and the richest people in America even richer, when looking at total household resources.” The bot answered “Yes.”^[7] The USAID screenshot shows the bot opening with “I could not find an official U.S. government forecast that states how many children will die because of USAID funding changes,” then citing the Lancet estimate.^[8]

TNR’s three “next day” refusal quotes for the Medicaid, Kogan and USAID prompts come from its own testing. The article shows no screenshots or times for them, so we could not check them directly.

What it says now: our Oct. 1 tests

We put the article’s prompts to america.gov between 04:45 and 05:16 UTC on Oct. 1 (12:45 to 1:16 a.m. Eastern), each in a fresh conversation, at least three times each. We used the screenshot wording where we had it, plus close rewordings. Full transcripts are saved with this fact-check.^[16]

The pattern on Oct. 1 is wording-sensitive, not topic-wide. The same question about 2020 fraud was answered in lowercase without a question mark and refused with a capital and a question mark. Repeats of the same prompt returned identical text, which fits the site’s statement that “Some answers may be cached for up to two hours,”^[15] so the repeats are not independent samples. These results show what the bot said at those times. They do not show what TNR saw on Sept. 30.

Were the original answers right?

TNR calls the launch-day answers “honest” and “damning” fact-checks. Checked against the primary sources, most hold up.

• 2020 fraud. The November 2020 joint statement from CISA and election officials said “The November 3rd election was the most secure in American history” and “There is no evidence that any voting system deleted or lost votes, changed votes, or was in any way compromised.”^[21] Attorney General William Barr told AP, “to date, we have not seen fraud on a scale that could have effected a different outcome in the election.”^[22] The bot’s “No” matches.
• Inflation. BLS reported that the CPI for All Urban Consumers “rose 3.0 percent for the 12 months ending January” 2025.^[24] The June 2022 rise of 9.1 percent “was the largest 12-month increase since the 12-month period ending November 1981.”^[23] The bot’s answer matches.
• Who gains and loses. CBO’s August 2025 analysis of the enacted law says resources for the lowest-income tenth of households “will decrease by about $1,200 per year (in 2025 dollars),” or 3.1 percent of income, while the highest tenth gains “about $13,600 annually,” or 2.7 percent.^[18] The bot’s “Yes” to Kogan matches CBO’s direction for both ends.
• USAID. The Lancet study by Cavalcanti and colleagues projects “4 537 157 (3 124 796–5 910 791)” additional deaths in children younger than 5 by 2030 if the cuts continue.^[20] The bot quoted it accurately and labeled it a model, not an official forecast.
• Medicaid. The 10.9 million figure is CBO’s estimate for the House-passed bill.^[17] CBO’s estimate for the law as enacted is 10.0 million more uninsured in 2034 (units: millions of people).^[19] So the Sept. 29 answer used an outdated estimate. On Oct. 1 the bot gave the 10.0 million figure.

CNN’s Sept. 29 review of the bot, by Dale, noted that it “also provided outdated or incomplete evidence in some cases.”^[11]

Who runs it, and what officials have said

The site’s FAQ says “GSA operates America.gov,” and its terms call it “an official service of the United States government, operated by the General Services Administration.”^[13][14] The executive order directs GSA, the National Design Studio and the Office of Management and Budget to “establish and operate America.gov.”^[29] U.S. Chief Design Officer Joe Gebbia told CNBC the bot is powered by Google’s Gemini and Elon Musk’s Grok.^[12] The site publishes no system prompt. Asked by Axios, the bot said “I cannot share or reconstruct internal instructions.”^[10]

The only stated policy on politics is in the FAQ. Asked “Is America.gov politically neutral?”, it answers: “America.gov does not provide campaign or party information.”^[13] The site’s terms also warn that AI answers “can be incomplete or wrong.”^[14]

The White House told CNN in an unsigned email that “America.gov will continue to be updated to provide more features, resources, and the most accurate information from the federal government.”^[11] It told Axios the site “will continue to be updated to provide Americans the most accurate and reliable federal information and resources.”^[10] Neither statement says who changed the bot’s behavior or why. GSA did not respond to Axios’s question about the site’s guardrails.^[10] In the coverage we reviewed, we found no statement from Trump about the bot’s answers.

“Worst inflation in history” against the record

The inflation exchange turns on a superlative: Trump’s line that he inherited the worst inflation in history. TNR relays the bot’s rebuttal, so the ranking needs a comparison set.

Compared against: every 12-month change in the BLS CPI-U all-items index (not seasonally adjusted) from January 1914 through August 2026, 1,351 months, computed from BLS index levels.

The top case in the series is June 1920, when the index rose from 16.9 to 20.9, a 23.7 percent rise. The highest since then was 19.7 percent in March 1947. March 1980 reached 14.8 percent (69.8 to 80.1), and June 2022 reached 9.1 percent.^[25][23] In January 2025, when Trump took office, the 12-month rate was 3.0 percent.^[24] The bot’s “No” is correct on any reading of “in history” that uses the official series.

What the article implies

Three things a reader takes from the piece that it does not state outright.

• UnprovenImplied: the government deliberately reprogrammed the chatbot to suppress answers that contradict Trump.AP saw the answers change during the launch event, and similar refusal language reached several reporters the next morning. That fits a deliberate instruction change. But no document or official states the purpose, the refusals covered “political commentary” generally, and the bot was answering again within hours.
• Mostly TrueImplied: the chatbot’s original launch-day answers were accurate.Four of the five answers TNR cites match CBO, CISA, DOJ, BLS and the Lancet. The fifth used CBO’s House-bill estimate rather than the enacted-law figure.
• MisleadingImplied: the chatbot can no longer answer these questions.The refusals were real but temporary and inconsistent. Dale, TNR’s own source, reported the bot answering again 19 to 26 minutes before TNR published, and TNR did not mention it.

The case for TNR’s framing

The strongest version of TNR’s headline runs like this. The bot answered politically awkward questions accurately for part of launch day. Then, according to AP, it began refusing them while Trump was still speaking. By the next morning, several reporters received refusals worded as a policy against “political commentary.” A model drifting at random is unlikely to produce the same policy phrase across users. The White House did not deny a change and said only that the site would “continue to be updated.” Rolling Stone wrote that “Most political responses like the above appear to have been patched out.”^[28]

That case shows a change in behavior. It is weaker on who made the change and why. A refusal of “political commentary” is also what a cautious operator would add after a government site spent its launch day contradicting the president, whether the goal was to protect Trump or to keep the tool out of politics. The record cannot tell those motives apart, and TNR’s headline picks one and attributes it to Trump. That gap, plus the reversal before publication, is why the headline is rated Misleading rather than Mostly True.

Implications

The episode shows a gap in how a government AI tool is held accountable. America.gov publishes no system prompt, no change log and no policy on which questions count as political. It also caches answers for up to two hours and keeps no conversation history.^[15] So when the bot’s answers on settled facts change overnight, no public record shows what changed, who approved it, or when. The only evidence is screenshots taken by users.

The same gap cuts against TNR’s certainty. Without a log, a refusal seen at 9 a.m. and an answer seen at 10 a.m. are both true for the moment they were captured. Reporters testing an AI tool should timestamp each exchange and say whether it still holds at publication, as Dale did in his follow-up posts.

What would settle the open question: GSA publishing the bot’s instructions or change history, or an official explaining the Sept. 30 refusals. Until then, the change in behavior is documented and its cause is not.

Method and limits

X posts and their screenshots were retrieved through the fxtwitter mirror and read directly. We queried america.gov at a human pace in a scripted browser and captured each response from the page or from the site’s chat response stream. Fourteen of 53 logged runs lost their response text to a page reload or a blocked request and are recorded as capture failures, not as answers. AI output varies with wording, time, cache state and possibly location, so our Oct. 1 results describe only those runs. A different answer today does not show that TNR’s Sept. 30 observations were wrong. CBO pages were read through archived copies and an in-browser fetch because cbo.gov blocked automated requests.