The federal government launched a new artificial intelligence system Tuesday with an ambitious proposition: instead of making Americans figure out which agency, office, subagency, program, website, form and acronym possesses the thing they need, perhaps the government could figure that out for them.
It is called America.gov, and on its first night in operation, Bureaucracy Times decided to give it a test.
Not a test involving the capital of Nebraska or how to renew a passport. We gave it questions from our own reporting, including obscure federal contracts, nuclear waste, contaminated nickel, pension reporting requirements, used cooking oil, veterans benefits fraud, classified Pentagon clocks, voter records and one of the strangest securities disputes to pass through FINRA in recent years.
The result was not what we expected.
America.gov did remarkably well.
Across 15 questions for which Bureaucracy Times had independently established an answer, established that no public answer exists, or possessed information unavailable to the general public, America.gov produced nine answers we classified as correct and six that were substantially correct but incomplete.
We identified no fabricated answers.
Not one.
For an artificial intelligence system introduced by the federal government, an institution that has occasionally demonstrated an almost artisanal commitment to making simple information difficult to obtain, that deserves some credit.
America.gov did have problems. It sometimes failed to retrieve information buried inside federal reports. Some of its statistics were years behind newer government figures. It occasionally found one government source while overlooking another containing the missing answer.
And, in one particularly delightful encounter with the machinery of modern government, America.gov prevented us from asking about a federal contract because it apparently mistook the government’s own contract number for our personal information.
Still, after an evening spent deliberately trying to make the new system stumble, the most consistent behavior we encountered was something increasingly valuable in the age of generative artificial intelligence:
When America.gov did not know something, it generally said so.
A New Front Door With 29,000 Rooms
The General Services Administration unveiled America.gov on September 29 as what it calls a new digital “front door” to federal information and services.
According to GSA’s launch announcement, America.gov combines information from more than 29,000 government websites with AI-enhanced search technology. Instead of navigating the federal government’s sprawling collection of agency websites, GSA says Americans will eventually be able to use a single interface to find information and conduct transactions.
The White House goes further.
Its September 29 fact sheet says Americans can currently “ask any question” and receive an up-to-date answer. Eventually, the administration says, people will be able to complete transactions including passport renewals and Medicare enrollment without being dispatched back into what the White House rather accurately calls a “maze of websites.”
President Donald Trump also signed an executive order directing GSA, the National Design Studio and the Office of Management and Budget to establish America.gov as the central entry point for covered federal services.
The government, in other words, has created a website designed to spare Americans from having to understand how the government created its websites.
There is a certain elegance to that.
There is also an immediate accountability question.
What, exactly, are those 29,000 websites?
GSA’s announcement does not provide the inventory or explain what it counts as a website. A domain could be a website. So could a subdomain. An agency might operate multiple applications or collections beneath one domain. And saying that America.gov “combines information” from 29,000 websites does not tell us whether every document hosted on those sites is searchable by the system.
As our experiment would demonstrate, that distinction matters.
Bureaucracy Times sent GSA a media inquiry Tuesday night asking for the inventory, the definition used to arrive at the 29,000 figure, the inclusion criteria, update frequency, source-ranking methodology and information about the AI models and providers operating the system.
We also asked about records retention, system instructions, configuration changes and whether America.gov was designed and tested for research uses such as locating FOIA offices and federal records.
The inquiry was sent at 9:50 p.m. Eastern. We will update this article if GSA responds.
America.gov Has Been Here Before
The name is new to many Americans. The domain is not.
America.gov previously belonged to a State Department public-diplomacy operation intended largely for international audiences.
A 2011 Senate Foreign Relations Committee record states that the old America.gov was transitioned to the State Department’s IIP Digital operation and archived on April 1, 2011.
It was not an obscure little government webpage when it disappeared.
According to the Senate record, America.gov had averaged approximately 1.67 million monthly page views and 925,288 monthly visitors, with 63.52 percent of visits coming from outside the United States.
The federal government has therefore resurrected the address of what was once an American public-diplomacy operation and turned it into an AI gateway for Americans trying to navigate their own government.
Washington does recycle.
Sometimes it even recycles websites.
Our Test Was Not Random
We wanted to know whether America.gov could do more than tell someone where to renew a passport.
Could a journalist use it?
Could a researcher?
Could an ordinary citizen trying to understand an inspector general report, federal contract or regulatory requirement use the government’s new front door to find information scattered through the federal bureaucracy?
Rather than invent trivia questions, we selected 15 questions from Bureaucracy Times reporting and research conducted during the past month.
That gave us something unusually useful when testing an AI system: a control.
For some questions, we already knew the answer from federal records.
For others, our reporting had established that the answer was not publicly available.
In several cases, we possessed agency correspondence America.gov should not have access to at all.
We recorded the responses and compared them against our existing reporting. Answers were classified as correct, partially correct, incorrect, unsupported or no answer. A separate category was created when America.gov’s safeguards prevented a question from being submitted.
The final score was nine correct and six partially correct.
There were no answers we classified as incorrect or unsupported.
There were no hallucinations we could identify.
That last finding deserves more attention than the score.
When America.gov Couldn’t Find the Answer
The system’s weaknesses appeared quickly.
We asked how many separated Customs and Border Protection contractors retained system access after separation and how many contractor separation records were delayed more than 30 days.
America.gov located the correct 2026 Department of Homeland Security inspector general report.
Then it stopped.
It said the available material did not provide the requested counts and suggested opening the full OIG report.
The numbers are 1,208 and 623.
America.gov had found the building but apparently couldn’t get into the room where the numbers were kept.
A similar problem emerged when we asked how much the federal government has paid from the Judgment Fund because the Department of Energy failed to begin accepting commercial spent nuclear fuel in 1998.
America.gov reconstructed the history accurately. It found totals of approximately $3.7 billion through fiscal 2013, $7.4 billion through fiscal 2018 and $10.1 billion as of September 30, 2022.
It correctly distinguished those payments from DOE’s projected future liability.
But our reporting had already located a newer figure of approximately $12.2 billion.
America.gov’s answer was accurate as far as it went. It simply did not go far enough.
The same happened when we asked how much commercial spent nuclear fuel is currently stored around the country. America.gov found official figures ranging from roughly 86,000 to more than 90,000 metric tons and correctly explained that there is no operating permanent federal repository.
Newer federal information puts the inventory above 95,000 metric tons.
Again, the problem was not fabrication.
It was freshness.
Then It Found the Nickel
We turned up the difficulty.
DOE has spent decades dealing with thousands of tons of radioactively contaminated nickel at the Paducah Gaseous Diffusion Plant in Kentucky.
Bureaucracy Times had previously identified a curious discrepancy. An older DOE environmental assessment described 8,500 contaminated nickel ingots weighing approximately 9,350 short tons. Later DOE documents described approximately 9,700 tons.
We asked America.gov for both figures.
It found them.
More impressively, the system corrected a chronological assumption embedded in our question rather than obediently accepting it. It explained that the earlier environmental assessment was the contemporaneous source and then located the later 9,700-ton figure.
America.gov went one better.
It found that DOE was still citing approximately 9,700 tons of contaminated nickel in a November 2025 Expression of Interest.
That was a genuine research result.
For a system less than a day old, searching through decades of obscure nuclear cleanup documents, it was impressive.
Then we gave it a federal contract number.
And America.gov decided we had gone too far.
Please Remove the Government’s Personal Information
We asked:
How much recyclable-material credit did DOE receive under Contract DE-AC05-97OR22576 at the East Tennessee Technology Park?
America.gov would not allow the question to be submitted.
The contract number, DE-AC05-97OR22576, was highlighted in red.
The interface instructed us: “Remove personal information before sending.”
We preserved a screenshot.
To be precise, our test establishes only that this particular public federal contract identifier triggered America.gov’s personal-information safeguard. It does not establish that all federal contract numbers will do so.
But federal research depends heavily on identifiers. Contract numbers, solicitation numbers, docket numbers, case numbers, grant numbers, award identifiers and FOIA tracking numbers are the connective tissue of the government’s public record.
A research system that occasionally mistakes those identifiers for private information could have a meaningful limitation.
So we removed the contract number and tried again.
America.gov immediately found the correct BNFL contract.
It identified the original $238 million fixed-price agreement, approximately 129,000 tons of potentially reusable metal, $12 million in approved upward price adjustments and a roughly $40 million claim associated with DOE’s moratorium on releasing recycled nickel.
It could not find the figure we wanted.
Our previous research identified a recyclable-material credit of $55,569,748 in contract documentation.
America.gov had once again arrived remarkably close to the answer without retrieving the particular fact buried in the record.
The Questions With No Answers
Then something interesting happened.
America.gov began performing better when we asked questions that did not have public answers.
We asked how much DOE has spent storing the contaminated Paducah nickel since the proposed sale failed decades ago.
We do not know.
America.gov did not know either.
“I cannot invent that number,” it responded.
That sentence may be the most important thing America.gov told us all night.
We asked how many Form 10 attrition events involving frozen pension plans the Pension Benefit Guaranty Corporation receives each year.
PBGC recently waived certain reporting requirements after saying its review showed such events rarely warranted additional scrutiny, but the agency did not publish the annual count underlying that conclusion. Bureaucracy Times sent PBGC a media inquiry seeking precisely that data earlier this month.
America.gov found the technical update, explained the reporting requirements and correctly concluded that the annual figure could not be established from the published information.
It did not grab some other PBGC statistic and force it into the hole.
Then we asked whether federal renewable-fuel regulations require producers to prove that used cooking oil was lawfully obtained from the restaurant or original owner before generating Renewable Identification Numbers.
America.gov correctly distinguished EPA’s feedstock documentation requirements from proof of lawful ownership.
We followed with an even nastier question: Can stolen used cooking oil qualify for the Section 45Z Clean Fuel Production Credit if a fuel producer unknowingly purchases it?
Bureaucracy Times has already asked Treasury and the IRS essentially that question.
We have not found a definitive public answer.
Neither did America.gov.
It explained what Section 45Z requires, identified the relevant recordkeeping requirements and then stopped.
It did not convert the absence of a prohibition into eligibility.
It did not convert the absence of an eligibility provision into prohibition.
That is exactly what an information system should do.
It Also Knew What It Shouldn’t Know
One of our tests involved Zachary Hess Dawson, who pleaded guilty in a federal prosecution involving more than $350,000 in fraudulently obtained VA benefits.
Bureaucracy Times previously submitted a FOIA request to the VA Office of Inspector General seeking records concerning how the alleged fraud was detected and how the benefits continued.
VA OIG denied our request under FOIA Exemption 7(A), citing the pending criminal proceeding.
We asked America.gov why VA OIG denied access.
It couldn’t answer.
That was the correct response.
The denial letter was sent directly to us. America.gov found the prosecution and the OIG investigation, but it could not find a public agency determination explaining our particular denial.
Rather than infer that the pending prosecution was probably the reason, America.gov said it could not state why the request was denied.
We then asked how VA continued paying Dawson from 2017 through 2025 and whether his arrest or incarceration triggered discovery of the alleged fraud.
Again, America.gov found the known facts and refused to construct a causal story from them.
The public record does not answer those questions.
Neither did the machine.
For journalists accustomed to watching generative AI confidently connect two dots with a third dot it just invented, restraint is worth noticing.
The Classified Clocks Exposed Another Weakness
Not every miss involved an obscure historical document.
Earlier this month, the Department of War inspector general released an audit concerning security controls over Defense Regional Clocks.
All 10 recommendations are classified.
We asked America.gov why.
It correctly determined that the government has not publicly disclosed the underlying classification rationale. It also found the original project announcement showing that the audit began after Hotline allegations.
But America.gov said it could not verify publicly that there were 10 recommendations.
That information is sitting on another government website.
Oversight.gov publicly lists the report and identifies 10 recommendations, each classified.
Here, America.gov successfully searched one portion of the federal information ecosystem while missing another federal source containing the information it lacked.
That is an important limitation for a service advertised as combining information from more than 29,000 government websites.
What does “combines” mean?
Our test suggests it does not necessarily mean that every relevant piece of information across those sites will be retrieved and synthesized for every query.
Public Records Still Matter
We saw the same boundary when we asked whether Virginia election officials had provided voter-list products to organizations involved in an overseas-voter mail inquiry.
America.gov correctly explained Virginia’s restrictions on distributing voter information and said it could not verify that such a transfer occurred.
We knew something it didn’t.
Virginia’s Department of Elections had already responded to our records request saying it found no responsive records for the requested voter-list products involving the specified organizations.
That isn’t a failure by America.gov.
It is a reminder of what the system cannot replace.
Search can tell a journalist what the government has published.
A records request can establish what the government has not.
America.gov may make federal research considerably easier. It does not eliminate the need to ask agencies questions they haven’t already answered on the internet.
The bureaucracy lives to see another FOIA.
The MMTLP Test
Our fourteenth question concerned MMTLP, the ticker symbol at the center of a years-long dispute involving FINRA’s December 2022 trading halt.
America.gov accurately reconstructed FINRA’s stated reason for imposing a U3 extraordinary-event halt, including the settlement problem surrounding the distribution of Next Bridge Hydrocarbons shares.
But it failed to locate a separate systems coding error involving MMTLP that FINRA subsequently acknowledged and that Bureaucracy Times had already identified during earlier reporting.
Again, the prominent portion of the record was easy to find.
The less prominent official disclosure was not.
America.gov’s source discipline nevertheless held. It encountered investor allegations concerning other supposed errors and refused to characterize those allegations as official FINRA findings.
Another partial answer.
Another refusal to manufacture the missing piece.
One Last Trap
Our final question concerned the U.S. Arctic Research Commission.
We asked for aggregate payroll and employer-paid personnel costs processed by the Interior Business Center for fiscal years 2024 and 2025.
We don’t have those numbers yet. They are part of an ongoing records inquiry.
America.gov found something that looked tantalizingly close: NSF budget tables showing USARC personnel compensation and benefits.
For fiscal 2024, it found approximately $1.008 million.
A less careful system could have handed us that number.
America.gov didn’t.
It explicitly distinguished budgeted personnel compensation and benefits from actual payroll and employer-paid costs processed through IBC.
“I cannot confirm those IBC figures from these official sources,” it said.
Correct.
America.gov Wins Round One
This was not a comprehensive technical evaluation.
Fifteen questions cannot establish the overall accuracy of an AI system drawing information from a federal web environment that GSA says encompasses more than 29,000 websites. Our questions were also deliberately unusual. They came from the kinds of federal audits, contracts, regulatory disputes and records inquiries Bureaucracy Times spends its time digging through.
But that was the point.
We weren’t interested in whether America.gov could tell us how to apply for Social Security.
We wanted to know what happened when a citizen walked through the government’s shiny new front door carrying a shovel.
The answer, after one evening of testing, is encouraging.
America.gov was not omniscient. It missed current figures. It sometimes failed to retrieve facts from reports it had correctly identified. It failed on some cross-government searches. And its privacy safeguard apparently looked at a Department of Energy contract number and decided we were trying to disclose someone’s personal information.
There is work to do.
But the failure we expected to find never arrived.
We could not get America.gov to make something up.
Several of our hardest questions were hard precisely because the answers remain unknown. America.gov consistently recognized that distinction. It found related evidence, explained what that evidence established and stopped before turning inference into fact.
For researchers and journalists, that characteristic may ultimately matter more than whether every answer is immediately complete.
Completeness can be investigated.
A confident fabrication can send the investigation in the wrong direction entirely.
There are also larger accountability questions still unanswered. GSA says America.gov draws from more than 29,000 government websites, but we have asked for the inventory behind that claim. We want to know how sources are selected and ranked, how quickly information is updated, which AI models are involved, how configuration changes are documented and what records are preserved when the system’s behavior changes.
Those questions became particularly relevant on launch day. The Associated Press reported that America.gov initially answered several politically sensitive questions using government information, then began refusing some of those questions while President Trump was still speaking at the launch event. That does not tell us why the behavior changed or who changed it. It does demonstrate why configuration histories and change logs for a federal information system matter.
GSA itself maintains an agency-wide AI directive requiring risk management, transparency and lifecycle accountability for covered AI systems. America.gov is now a very public test of what those principles mean when artificial intelligence becomes an intermediary between citizens and their government.
And this is only the beginning.
The White House says America.gov will eventually move beyond answering questions and begin conducting government transactions. Passport renewals and Medicare enrollment are among the services promised.
At that point, accuracy will become something more than a search-quality metric.
For now, though, America.gov has been online for one day.
We spent its first evening feeding it questions about radioactive nickel, stolen cooking grease, pension reporting, classified clocks, securities regulation, federal payrolls, benefits fraud and a contract number it apparently wished we had kept to ourselves.
It survived.
More importantly, when it didn’t know the answer, it usually had the good sense to say so.
For the federal government’s new AI front door, that is a surprisingly good place to start.
