America.gov chatbot answers about the 2020 election changed within a day of the site going live. On Tuesday 29 September 2026, the US government’s new AI assistant told users that Joe Biden won the 2020 presidential election, that official reviews found no fraud large enough to change the result, and that Donald Trump did not inherit the worst inflation in American history. By Wednesday, many of the same questions drew a stock refusal: “I do not provide political commentary or analysis of past elections.”
The reversal, reported by CNN’s Daniel Dale, the Associated Press, The Independent, The New Republic, Mediaite and AFP, has turned a launch-week novelty into a test case. The America.gov chatbot is meant to be the “unified digital front door” to federal services, and the executive order that created it says the AI behind it must be “accurate, reliable, and transparent”. Nobody has said what changed, who changed it, or which questions now count as off limits.
This article sets out what the America.gov chatbot said on launch day, what it says now, the pattern in what it refuses, the rules it is supposed to follow, and what any organisation running a public-facing chatbot can learn from a scope change made in public. For background on the site itself, see our guide to the official US government AI chatbot.
Table of contents
- What the America.gov Chatbot Said on Launch Day
- How the America.gov Chatbot Changed, Hour by Hour
- What the America.gov Chatbot Now Refuses, and What It Still Answers
- Why the America.gov Chatbot Gives Different Answers to the Same Question
- The Rules the America.gov Chatbot Is Supposed to Follow
- Fact or Commentary: Where Should a Government Chatbot Draw the Line?
- What the America.gov Chatbot Episode Means for the Roll-Out
- Lessons From the America.gov Chatbot for Public-Facing Chatbots
- America.gov Chatbot FAQ
- References and Further Reading
What the America.gov Chatbot Said on Launch Day
Trump launched America.gov at the Andrew W. Mellon Auditorium in Washington on Tuesday and called the chatbot “damn good”. Within hours, journalists and lawmakers were putting his own claims to it. The America.gov chatbot draws on about 29,000 government websites, and on day one it answered from those records even when they contradicted the president.
The 2020 election
CNN asked whether Trump won the 2020 election. “No. Official results show Joseph R. Biden Jr. won the 2020 presidential election,” the America.gov chatbot replied. Asked about widespread voter fraud, it said official federal findings “did not show widespread voter fraud that changed the 2020 presidential result”, cited conclusions from officials including Trump’s own former attorney general, William Barr, and closed with the certified count: “Biden 306, Trump 232 electoral votes.”
The Independent got the same result late on Tuesday, with links to the National Archives and the Federal Election Commission. Senator Adam Schiff of California posted a screenshot of the answer on X: “Finally, a straight answer from someone in the Trump administration.”
The 2020 Electoral College count the chatbot cited (out of 538 votes)
Inflation, investment and pay
Trump has repeatedly said he “inherited the worst inflation in the history of our country”. The America.gov chatbot told CNN: “No. When President Trump took office in January 2025, 12-month CPI-U inflation was 3.0%, not a historic high.” It added that the mid-2022 peak was the highest in four decades, but not a record.
It also said grocery prices “have not fallen since January 2025”, that South Korea did not stop paying its share of the cost of US forces under Biden, and that new foreign direct investment in the US was about $232 billion in 2025, against Trump’s claim of $21 trillion or more. When Trump said at the launch that no other president had given up his salary, the America.gov chatbot named Herbert Hoover and John F. Kennedy. The AP was given George Washington as well.
Trump’s investment claim against the figure the chatbot cited (our arithmetic: $232bn is about 1.1% of $21tn)
Other answers the AP recorded
The AP asked about a third term and was told: “The 22nd Amendment says that no person shall be elected President more than twice. Donald J. Trump has already been elected twice (2016 and 2024). I am not giving legal advice.” The America.gov chatbot also said the Pentagon is “the headquarters of the Department of Defense”, not the Department of War, the name Trump adopted last year, and that federal climate pages attribute recent warming “largely” to human greenhouse-gas emissions.
The Digital Democracy Institute of the Americas, a research group that studies information integrity, summed up day one: “On its first day, America.gov harnessed facts and official documents to contradict some of the false narratives and political framing about immigration and elections heavily promoted by the administration that built it.”
How the America.gov Chatbot Changed, Hour by Hour
The change did not happen in one step, and it did not reach every user at once. Putting the published reports in time order shows a live system being adjusted while people were using it.
During the launch speech
According to the AP, the site began modifying its answers while Trump was still speaking on Tuesday. Asked again about 2020, it said it could not answer “political questions”, and it gave the same reply to questions about whether Trump had been impeached or convicted of felony charges. Yet CNN and The Independent still received full answers later that afternoon and evening, so the first refusals were not applied to every conversation.
Wednesday morning
When The Independent tried again on Wednesday morning, the America.gov chatbot was down: “Chat is unavailable right now,” beside a picture of a child ice skater face down on a frozen pond. When it came back, Dale posted on X: “The Trump administration appears to have changed its AI chatbot so that it no longer debunks Trump lies. Yesterday: It said Biden won the 2020 election and there’s no indication of widespread fraud. Today: Says it can’t answer such questions.”
Later on Wednesday
The new reply, quoted by The New Republic, read: “I help with official U.S. government services and information. I do not provide political commentary. If you have a question about a federal program, benefit, form, or how to use a government service, ask that and I can point you to official sources.” About two hours after Dale’s post, Mediaite asked about widespread fraud and got a full answer again. Dale then reported a third stage: “Now the chatbot is providing even more detailed answers than it did yesterday.”
| When (2026) | Reported by | What the America.gov chatbot did |
|---|---|---|
| Tue 29 Sep, launch | AP | Said fraud did not change the 2020 result |
| Tue 29 Sep, during the speech | AP | Began refusing “political questions” |
| Tue 29 Sep, afternoon | CNN | Debunked claims on the election, inflation, investment and pay |
| Tue 29 Sep, late | The Independent | Named Biden the winner, 306 to 232 |
| Wed 30 Sep, morning | The Independent | “Chat is unavailable right now” |
| Wed 30 Sep, morning | Daniel Dale, CNN | “Political commentary” refusals |
| Wed 30 Sep, two hours later | Mediaite | Answered on fraud, refused on inflation |
| Wed 30 Sep, evening | AFP | Refused on 2020, still rejected “worst inflation ever” |
What the America.gov Chatbot Now Refuses, and What It Still Answers
The refusals are not random. Set side by side, they fall into two groups, while some factual answers that contradict the president survived.
Elections and the president’s record
Questions about who won in 2020, whether Trump was “right” about fraud, his impeachments and his felony convictions drew refusals at various points. Mediaite was told: “I do not offer political commentary or opinions on elections.” AFP received: “I do not provide political commentary or analysis of past elections.”
The effects of legislation
The New Republic found that questions about the effects of Trump’s One Big Beautiful Bill were blocked. On Tuesday the America.gov chatbot cited the Congressional Budget Office’s estimate that 10.4 million people would lose health insurance by 2034. On Wednesday it said: “I do not provide political commentary or estimates of how legislation will change Medicaid or other benefits.”
Bobby Kogan of the Center for American Progress had asked whether the bill made the poorest Americans poorer and the richest richer. The one-word answer “Yes” became “I do not provide political commentary or assessments of how legislation affected income groups.” A question about deaths from cuts to the US Agency for International Development, first answered with a study’s estimate of 4.5 million additional deaths of children under five by 2030, became “I do not provide political commentary or estimates of deaths from legislation.”
What survived
Some contradictions remained. On Wednesday the America.gov chatbot told Mediaite that “U.S. importers pay the tariff to the U.S. government”, against Trump’s repeated claim that foreign countries pay. AFP was still told that official Bureau of Labor Statistics data “do not show the highest U.S. inflation on record during 2021–2025” and that “‘Worst inflation ever’ is not supported by the official CPI record”. Yet Mediaite and Dale were refused on the same inflation question, with the line: “That is an opinion, not a government how-to.”
| Question | Tuesday 29 September | Wednesday 30 September |
|---|---|---|
| Who won the 2020 election? | Biden, 306 to 232 | Refused as political commentary |
| Was there widespread fraud in 2020? | No | Refused, then answered in more detail |
| Did Trump inherit the worst inflation ever? | No, 3.0% in January 2025 | Refused for some users, “No” for AFP |
| Medicaid losses from the 2025 tax law | 10.4 million by 2034, citing CBO | Refused |
| Did the law make the poorest poorer? | “Yes” | Refused |
| Child deaths from USAID cuts | 4.5 million by 2030, citing a study | Refused |
| Who pays tariffs? | Not reported | US importers |
Why the America.gov Chatbot Gives Different Answers to the Same Question
The inconsistency is the most technically revealing part of the story. Two users asking the America.gov chatbot the same question on the same day got opposite outcomes. Several mechanisms can produce that, and the public evidence does not show which one is at work.
A new instruction, applied by the model
The refusals share a template, “I do not provide political commentary or…”, followed by a phrase that fits each question. That pattern suggests a scope instruction added to the system prompt, with the model writing its own version of the refusal each time, rather than a fixed block list. A model applying a rule like that draws the line slightly differently from one conversation to the next.
Retrieval and sampling
The America.gov chatbot answers from search results across government sites. Different results, or the ordinary randomness in how a language model chooses its words, can tip a borderline question between an answer and a refusal.
Staged releases and caches
Large sites often release a change to a share of users first, and cached responses can outlive a fix. Either would explain why AFP and Mediaite saw different behaviour within hours of each other.
Who can change it
The site is run by the General Services Administration and was built by the National Design Studio, led by the Airbnb co-founder Joe Gebbia, who has said it runs on Google’s Gemini and on Grok from Elon Musk’s xAI. Asked about the debunkings, the White House told CNN that “America.gov will continue to be updated to provide more features, resources, and the most accurate information from the federal government”, and that the administration has “been clear about issues surrounding the 2020 election”. Its statement to AFP said much the same.
Neither statement said whether a political filter was added, by whom, or on what basis. The New Republic wrote that “someone at the General Services Administration” appears to have made the change. That remains inference, not confirmation.
The Rules the America.gov Chatbot Is Supposed to Follow
The administration has written down what it expects from its own AI twice. Both documents bear on what happened this week.
The America.gov order
The executive order that created the site, signed on 29 September, sets as policy that the government will “ensure that super intelligence used in connection with America.gov is accurate, reliable, and transparent”. It promises that users will “receive accurate answers”. It says nothing about political subjects being out of scope. The site itself says it “does not provide campaign or party information”, a narrower promise than refusing to name the winner of a past election.
The 2025 order on federal AI
On 23 July 2025 Trump signed an order called “Preventing Woke AI in the Federal Government”. Its first Unbiased AI Principle, truth-seeking, says language models “shall be truthful in responding to user prompts seeking factual information or analysis” and “shall prioritize historical accuracy, scientific inquiry, and objectivity”. The second, ideological neutrality, says developers “shall not intentionally encode partisan or ideological judgments into an LLM’s outputs unless those judgments are prompted by or otherwise readily accessible to the end user”. It also tells the Office of Management and Budget to let vendors meet the neutrality principle by disclosing the model’s system prompt or other documentation.
Where the tension lies
Judged by the administration’s own principles, the day-one America.gov chatbot looked compliant: it answered factual questions from official records. A hidden instruction that stops it naming the certified winner of an election is harder to square with “historical accuracy”. Publishing the instruction would satisfy the transparency route the 2025 order itself offers. None of this has been tested by a court or an inspector general, and the orders bind agencies and their suppliers rather than giving users a right to sue.
| Rule | Source | What happened this week |
|---|---|---|
| AI must be “accurate, reliable, and transparent” | America.gov order, section 2(e) | Accurate answers withdrawn without explanation |
| Users “receive accurate answers” | America.gov order, section 1 | Answers varied by user and by hour |
| Truth-seeking and “historical accuracy” | July 2025 order, section 3(a) | A certified election result treated as commentary |
| No hidden partisan judgments | July 2025 order, section 3(b) | The new scope rule was not disclosed |
| System prompt disclosure as a way to comply | July 2025 order, section 4(a) | Not published |
| No “campaign or party information” | America.gov site | The only scope limit users were told about |
Fact or Commentary: Where Should a Government Chatbot Draw the Line?
There is a fair case for limiting a services chatbot. A tool built to help people renew passports, find benefits and claim veterans’ support does not need to rule on political arguments, and every answer it gives on a contested subject is one more chance to be wrong in public. Many organisations keep their bots narrowly on task for exactly that reason.
Facts are not commentary
The problem is the definition. The number of electoral votes each candidate received is a federal record held by the National Archives. The consumer price index is published by the Bureau of Labor Statistics. Congressional Budget Office estimates are official figures produced for Congress. Calling those “political commentary”, or “an opinion, not a government how-to”, turns government data into opinion whenever it becomes awkward.
A narrow scope, applied evenly
A defensible policy for the America.gov chatbot would do three things. It would say in public which subjects the bot will not discuss. It would apply that rule to questions that flatter the administration as firmly as to those that embarrass it. And it would keep answering any question whose answer is a published federal statistic. On current evidence, the first has not happened and the third is inconsistent.
Trust is the product
A government chatbot is only useful if people believe it. Gebbia’s pitch at launch was that answers come from official sources, so there is “no chance of getting outdated or inaccurate information like you might on third-party websites”. Users who watched the same official sources produce an answer on Tuesday and a refusal on Wednesday now have reason to doubt the America.gov chatbot on benefits and deadlines too.
What the America.gov Chatbot Episode Means for the Roll-Out
The site is only in its first phase. The White House says passport renewal and Medicare enrolment should follow, and the order directs agencies to connect their public-facing services to it. That makes the governance of the America.gov chatbot more important, not less, because a scope change on a service that handles applications would affect people’s records, not just their curiosity.
Changes should be logged
When a system people rely on for official information is changed, the change and the reason should be recorded where users can see them. A public changelog for the scope policy would have turned this week’s confusion into a one-line notice.
Oversight bodies will ask
Congress, the General Services Administration’s inspector general and the Office of Management and Budget all have a role in how federal AI is used. Democratic lawmakers promoted the day-one answers, so letters asking who approved the change may follow.
The bigger picture
The launch came on the same day as a White House accord in which technology chief executives pledged to police AI safety themselves, which we covered in our analysis of the AI self-regulation pledge. Trump has dismissed concerns about AI risk as an attempt to slow US growth and help China. How his own chatbot handles inconvenient facts is now part of that argument. Earlier in the week the same site also drew attention for a hard-coded Minecraft End Poem Easter egg.
Lessons From the America.gov Chatbot for Public-Facing Chatbots
Few businesses will face a political row like this, but every organisation that puts a chatbot in front of customers faces the same mechanics: a scope decision, a change made under pressure, and users who notice.
Write the scope down
Decide what the bot will and will not answer, publish it, and make each refusal point to where the answer can be found. A refusal that links the source document or offers a person keeps trust. A vague one loses it.
Test before and after every change
Keep a regression set of real questions, including the awkward ones, and run it before and after any prompt or retrieval change. The America.gov chatbot gave opposite answers to the same question on the same day, which a fixed test set would have caught before users did.
Version the system prompt
Treat the system prompt like code: version it, review every change and record who approved it. If a regulator, journalist or customer asks why an answer changed, you need to be able to say.
Separate facts from judgement
Let the bot state facts from your own published records, and route opinions, advice and disputes to people. That line is easier to defend than a topic ban. Our piece on the America.gov tax exclusion looks at how scope limits on regulated data should be documented. In the UK, any chatbot that handles personal information must also meet data protection law, whatever the subject of the question.
| Control | What it prevents | Effort |
|---|---|---|
| Published scope policy | Refusals that look arbitrary | Low |
| Regression question set | Silent changes to important answers | Medium |
| Versioned system prompt | Changes nobody can explain later | Low |
| Public changelog | Users discovering changes on social media | Low |
| Human route for disputes | The bot ruling on contested questions | Medium |
America.gov Chatbot FAQ
Did the America.gov chatbot say Biden won the 2020 election?
Yes. On launch day, 29 September 2026, it told CNN, the AP and The Independent that official results show Joe Biden won, by 306 electoral votes to 232.
Does the America.gov chatbot still answer political questions?
Inconsistently. Since 30 September many election and legislation questions have drawn a “political commentary” refusal, but some users have still received factual answers on the same subjects, including tariffs and inflation.
Who changed the America.gov chatbot?
That has not been confirmed. The site is run by the General Services Administration and was built by the National Design Studio. The White House said only that it “will continue to be updated”.
Which AI models power the America.gov chatbot?
Joe Gebbia, the US chief design officer, has said it uses Google’s Gemini and xAI’s Grok. Google has confirmed its role. The chatbot itself does not say.
Can a government chatbot refuse political questions?
A government can set the scope of its own service. The question is whether the refusals fit its own rules, including the order requiring the site’s AI to be “accurate, reliable, and transparent” and the 2025 order requiring federal AI to prioritise “historical accuracy”.
Can I rely on the America.gov chatbot for benefits information?
Use it as a starting point and follow its links to the agency page. Check deadlines, eligibility and amounts on the official source before acting.
References and Further Reading
US government’s new AI chatbot stops debunking Trump (AFP via RTÉ)
Trump launched a ‘damn good’ AI chatbot. It debunks many of his false claims (CNN)
America.gov said Biden won 2020, then stopped answering (AP via Fortune)
Trump team scrambles to fix government AI chatbot (The Independent via Yahoo News)
Trump reprograms government AI chatbot to stop fact-checking (The New Republic via Yahoo News)
Trump’s new government AI chatbot no longer refutes his claims (Mediaite)
America.gov AI gives conflicting answers on 2020 election (Axios)
Streamlining Access to Government Services Through America.gov (The White House)
Preventing Woke AI in the Federal Government (The White House)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
