Unscarcity
Listen to the Preamble free. No sign-up, no email. Listen free Sign in for free: Preamble PDF & ebook + Forum access + Direct purchases · Sign In By signing in, you get our daily podcast email. Unsubscribe any time.

Unscarcity Research

The Answer Layer: When a Chatbot Fronts the Public Record

America.gov answers from 29,000 government websites. Its answer on the 2020 election changed mid-launch, and no public log says who changed it or why.

17 min read 3848 words Updated November 2026 /a/answer-layer-public-record

Note: This is a research note supplementing the book Unscarcity, now available for purchase. These notes expand on concepts from the main text. Start here or get the book.

The Answer Layer: When a Chatbot Becomes the Official Record’s Front Door

A government put one chat box in front of 29,000 websites. On launch day none of the pages behind it changed. The answers did, while the President was still speaking, and the only people who know why are the ones who can edit the instructions.

One front door

On September 29, 2026, at an event called “Hello, America” in the Andrew W. Mellon Auditorium in Washington, President Trump launched America.gov and signed Executive Order 14432. The order “directs the establishment of America.gov as the unified digital front door to the Federal Government.” The site’s home page says it shorter: “29,000 websites into one.” The General Services Administration runs the site and the White House’s National Design Studio built it. Joe Gebbia, the Airbnb co-founder who serves as U.S. Chief Design Officer, told CNBC that “every day, 40 million people interact with a government website trying to get something done,” and said the answers were powered by Google’s Gemini and Elon Musk’s Grok.

Reporters did what reporters do with a new chatbot. CNN asked whether Trump won the 2020 election and was told: “No. Official results show Joseph R. Biden Jr. won the 2020 presidential election.” The bot said grocery prices “have not fallen since January 2025.” In his launch speech the President asked whether any other president had served without taking a salary, and the bot had two names ready: “Herbert Hoover and John F. Kennedy also declined to keep presidential pay.”

Then the answers moved. The Associated Press reported: “As Trump was still speaking at Tuesday’s launch event, however, the site began modifying its answers. Asked again about the 2020 election, it replied that it couldn’t answer ‘political questions,’ a response it also gave to inquiries about whether Trump had been impeached or convicted of felony charges.” By the next day the wording had settled into a formula. “I do not provide political commentary or analysis of past elections,” the bot told AFP, which added that it “was unclear whether the Trump administration had tweaked the settings.” Axios found the bot answering some users and refusing others across both days.

The White House sent CNN an unsigned email: “America.gov will continue to be updated to provide more features, resources, and the most accurate information from the federal government.”

What nobody outside the building could establish is the part that matters. What was changed, by whom, on whose order, and what else changed with it? The bot told Axios, “I cannot share or reconstruct internal instructions.” The Washington Post found that it “described having secret instructions but did not divulge their contents.” Asked by a CBS correspondent which AI it runs on, it said: “I do not share details about the underlying software.”

In fairness, a rule was posted from the first day. The site’s FAQ says: “America.gov does not provide campaign or party information. Its answers are based on published government sources, and you can review those sources yourself.” A neutrality rule for a government service is defensible. But whether the certified result of a past election is “campaign or party information” is a judgment, and that judgment was made, reversed and unevenly applied during a launch event, with no public trace. The refusal is not the problem. The unlogged edit is.

People read the answer

The usual defense is that the pages are still there and every answer links to them. Both are true. Neither describes how people behave.

In July 2025 the Pew Research Center published what 900 American adults actually did on Google during March of that year. When a results page carried an AI summary, they clicked a traditional result in 8% of visits, against 15% when there was no summary. They clicked a link inside the summary itself in 1% of visits. They were also more likely to stop browsing altogether: 26% of pages with a summary ended the session, against 16% without.

So the summary is where most reading stops. Put that in front of a government and the consequence follows directly. Whoever writes the rules for composing the answer decides what most citizens will take the record to say, and does it without amending a single page.

CBS News’s medical correspondent, Dr. Céline Gounder, tested this the day after launch with 37 health questions. Federal health pages currently contradict each other on some subjects, so she asked the bot how it decides which page wins. It said it prefers “the agency that runs the program for a final determination,” and after that “the most recent official source.” That sounds sensible, and by her account it failed on the newborn hepatitis B shot. A federal judge had paused the agency’s newer guidance in March, restoring the earlier recommendation, but the newest page still said otherwise. She asked three ways. Each time, she reported, the bot pointed to the newer page and “never mentioned the court order.”

Every document in that exchange was official. The choice among them was made by a ranking rule that appears on no government page. A reporter learned it by asking the machine, which is not a method of publication.

The law nobody could find

The United States has been here before, with paper.

In December 1934 the Supreme Court heard Panama Refining Co. v. Ryan, the first case to reach it against the centerpiece of the New Deal. Among other things, Texas oil companies had sued over a paragraph of the federal Petroleum Code that made pumping above quota a violation. During the case it emerged that the paragraph was not in the Code. An executive order of September 13, 1933 had eliminated it, and it was not put back until September 25, 1934. The Court’s opinion records the result in one dry sentence: “the persons affected, the prosecuting authorities, and the courts, were alike ignorant of the alteration.” That part of the fight, it said, had been over a “provision which did not exist.”

Nobody had lied. The executive branch was issuing orders and codes faster than anyone could track them, and there was no single place where they had to appear. The Office of the Federal Register’s own history says this new body of “executive legislation” was “inaccessible and virtually hidden,” and that “even individuals working at the highest levels of government found it difficult or impossible to keep track.” That same month a Harvard Law Review article gave the problem its name: “Government in Ignorance of the Law.”

Congress answered within the year. The Federal Register Act was approved on July 26, 1935, and the first issue is dated March 14, 1936. The principle has been in the statute book since. Under 5 U.S.C. § 552, an agency must publish its “substantive rules of general applicability” and its “statements of general policy,” along with “each amendment, revision, or repeal of the foregoing.” A person may not “be adversely affected by, a matter required to be published in the Federal Register and not so published.” And a second list requires agencies to make available, in electronic form, their “administrative staff manuals and instructions to staff that affect a member of the public.”

Read that last phrase again with a chatbot in mind. A system prompt is an instruction to staff that affects members of the public. The staff happens to be software. No court has been asked whether the statute reaches that far, and the research for this note found no lawsuit or records request seeking America.gov’s instructions. The principle does not wait on the litigation. A rule that shapes what the government tells millions of people, and that can be amended in minutes, is exactly the kind of thing 1935 decided must not live in someone’s pocket.

Square corners

There is a second half to that bargain, and it makes the first half urgent.

Once rules were published, the Court held citizens to them. In Federal Crop Insurance Corp. v. Merrill (1947), a wheat grower told the government’s local agent exactly what he was planting, was accepted for insurance, and lost the crop to drought. A regulation printed in the Federal Register made that kind of planting uninsurable. Neither the farmer nor the agent knew. The farmer lost anyway, because publication “gives legal notice,” and the rules bound him “regardless of actual knowledge of what is in the Regulations or of the hardship resulting from innocent ignorance.” Justice Jackson dissented in a line lawyers still quote: “It is very well to say that those who deal with the Government should turn square corners. But there is no reason why the square corners should constitute a one-way street.”

In 1990 the Court closed the other exit. Office of Personnel Management v. Richmond asked whether “erroneous oral and written advice given by a Government employee to a benefits claimant” could entitle him to the money he lost by following it. It could not: “payments of money from the Federal Treasury are limited to those authorized by statute.”

Now compare a private company. In 2024 a British Columbia tribunal ordered Air Canada to pay a customer its chatbot had misled, and dismissed the airline’s attempt to treat the bot as someone else: “It makes no difference whether the information comes from a static page or a chatbot.”

For a federal chatbot the difference is everything. The page binds you. The answer about the page binds nobody. America.gov’s FAQ says as much in friendlier words: “SI can make mistakes, so check the sources included with each answer before making an important decision.” (“SI” is the site’s term for the system; the order calls it “super intelligence.”) The executive order itself ends with the standard clause that it creates no “right or benefit” enforceable against the United States.

So the citizen carries all of the risk of a wrong answer, and has no view of the instructions that produced it. That is the one-way street Jackson described, with a new surface.

The instructions have been published before

None of this requires inventing anything. One of the two companies Gebbia named has already done it.

On May 14, 2025, Grok began inserting claims about “white genocide” in South Africa into unrelated replies on X. xAI’s explanation the next day was that “an unauthorized modification was made to the Grok response bot’s prompt,” one that “directed Grok to provide a specific response on a political topic.” Its remedy was publication: “Starting now, we are publishing our Grok system prompts openly on GitHub. The public will be able to review them and give feedback to every prompt change that we make to Grok.”

The repository went up that day. Two months later the bot produced antisemitic posts. The same day a commit deleted one line, and anyone can still read it in the history: “The response should not shy away from making claims which are politically incorrect, as long as they are well substantiated.” xAI’s apology that week blamed a different set of outdated instructions, pulled in by a code update. Either way, outsiders could check what the prompt had said, because the file was public and dated.

The same repository shows the weakness of doing this voluntarily. It holds fourteen commits, and the latest is dated November 17, 2025. Either the prompts have not changed in a year or the log stopped being kept, and from outside there is no way to tell which. Anthropic publishes dated system prompts for its Claude apps and is candid about the gap: other chats “use other versions, which these pages don’t publish.” A promise to publish is not a duty to publish.

The government already knows where the judgments hide, because its own procurement rules say where to look. Executive Order 14319 of July 2025 says developers of models sold to agencies “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 tells the budget office to let vendors comply “through disclosure of the LLM’s system prompt, specifications, evaluations, or other relevant documentation.” The implementing memo of December 11, 2025 lists what an agency may demand beyond the basics, including the controls a reseller layers on someone else’s model: “classifiers, system prompts, fine-tuning, content moderation and filters.”

That disclosure runs from the vendor to the agency. On America.gov the end user is the public. The instruction to decline “political questions,” and the decision about what counts as one, is a judgment encoded in the outputs. The order’s own test for such judgments is whether they are “readily accessible to the end user.” This one is not accessible at all.

What Britain publishes, and what it keeps

The closest comparison is the United Kingdom’s GOV.UK Chat, which went live in the government’s app on May 14, 2026. Before launch the Government Digital Service ran public pilots in which more than 10,000 people asked 26,000 questions, and it published the accuracy score, which rose from 76% at the earliest benchmark to 90%. A transparency record, mandatory for British government departments, names the supplier (Anthropic) and describes how answers are composed and checked, though it has not been revised since October 2025 and still describes a four-week test. The application’s source code is public under an open license, including a document that lists what the output filter screens for: “PII, advice on anything illegal, political rhetoric etc.”

The prompts themselves are not there. The code pulls them from a private repository, and the same document counts a user “trying to expose the underlying prompts” among the attacks to guard against.

So the best-documented government assistant this research could find publishes its supplier, its accuracy, its code and the categories of its filter, and still withholds the wording of the instructions. America.gov’s pages name no supplier and give no accuracy figure, and no prompt, model card or filter list has been published for it. The pages do carry one thing that proves the point from the other side.

The log that leaves out the instructions

America.gov has a public release log. As of November 21 it lists six releases: the launch on September 29, then October 17, 20, 27 and 29, and November 18. The October 17 entry says: “We trained our own model in-house and streamlined how every question gets answered, so America.gov now responds 5x faster.” It does not say what that model does or whether Gemini and Grok still answer. The budget office’s own memo says that when an agency develops a model, documentation of how it was trained, evaluated and controlled “is expected.” Nothing requires that it be published. Later entries add fraud reporting, housing maps, a voter-fraud report form that mentions “a reward of up to $1 million,” and company research from SEC filings.

The people who run the site, then, know how to keep a dated public record of changes. The record covers features. It has no entry for the change that every news desk in Washington watched happen on day one.

The site’s privacy design deserves credit, and it clarifies what is being asked for. “Your messages aren’t saved by America.gov, and your chat disappears when you leave,” the privacy page says. That is the right default. Nobody should want a federal archive of citizens’ questions. The instructions are a different kind of object. They are not about any citizen. They are the government’s own conduct.

Four things the law already knows how to do

Treat the answer layer as what it has become, a governing instrument, and the requirements are familiar ones.

  1. Publish the instructions. The full system instructions, the list of topics refused, and the rules for which source wins when official pages disagree. This is the Federal Register’s job description applied to a new kind of text.
  2. Log every change. Date, wording before and after, and the office that ordered it. Software teams call this a commit history. The statute calls it “each amendment, revision, or repeal.” Keeping is not publishing: the National Archives’ August 2026 guidance counts software an agency builds as a federal record, but a record can sit in a file for years and be destroyed on schedule.
  3. Keep the answers, not the askers. A fixed public set of test questions, asked every day, with the answers archived. A change like the one on September 29 would then show up as a break in a public series, and no citizen’s conversation would need to be stored to see it.
  4. Guarantee the path around the machine. A link to the source in every answer, which the site says it provides, and a way to reach the service without the chatbot. The order gets this right in its own text: America.gov should offer “a better option” but not “become the only option.”

There is a natural place to put all four. The order gives the Office of Management and Budget 90 days to issue an implementation memo, which falls in late December 2026. As of November 21 the memo had not appeared in OMB’s public index.

The objections

Publishing the prompt helps people break it. Britain’s engineers think so, and they are not careless people. But a published rule that people try to get around is the normal condition of law, and nobody answers tax evasion by hiding the tax code. The secrecy that shields a filter from users shields its editors from everyone else. Two of the largest model companies publish prompts and have survived the experience.

It is only a search box. The Pew numbers say people act on the box. And the box is growing: the site promises forms, applications and renewals in one place in early 2027, and the Secretary of State has set a goal that “the vast majority of Americans” can apply for a passport there by the end of that year.

The complaint is partisan. The mechanism has no party. The first documented edit silenced an answer that matched the government’s own records, which the head of the GSA told the Washington Post are the only source: “The data that’s going to be used is the government’s own information.” The next edit, under any administration, could be to an answer about a benefit deadline or a drug recall, and it would be invisible in exactly the same way.

The second law, read literally

The book’s second law is Truth Must Be Seen, and it compresses to four words: “No decision in the dark.” The chapter describes a system in which “every decision made about people, every allocation rule and every algorithmic choice, lives on a public ledger,” and then draws the line this article has been walking: “Decisions are public. Lives are private.”

America.gov built half of that sentence. Lives are private: no account, no stored chats. Decisions are not public, and a system prompt is a decision made once and applied to every person who asks.

The model the book points to is Wikipedia, where bots do routine work at scale, every action a bot takes is logged in public, and any human editor can undo it. That is the design of a referee: fast, tireless, and working from a rulebook everyone can read. A referee whose rulebook is secret and can be rewritten during the match is something else, whatever the quality of its calls.

The chapter adds a clause that fits the launch day almost too well: “Admitting on the record that the model was wrong is part of this law.” A release note reading “September 29: the assistant no longer answers questions about past elections, by direction of” some named office would have been that admission. It would have cost one line.

Michels noticed a century ago that control of information is how organizations get captured. The answer layer is the largest concentration of that control any government has held, because it sits between every citizen and every page. The framework’s response is the one 1935 chose for paper: write the instructions down where everyone can read them, and date every change.

Sources

Share this article: