Counsel Law

What America.gov's Own Terms Say About the Answers It Gives You

Bottom line

The government's own fine print says the portal is not the agency of record, its answers can be wrong, and nothing there is legal, tax or money advice.

Counsel Editorial Updated November 12, 2026
Timeline of 7 dates named in this article: September 8, 2026; September 29, 2026; September 30, 2026; October 17, 2026; October 20, 2026; October 29, 2026; November 12, 2026.
Timeline Dates named in this article. Dates as they appear in this article’s text, placed on a linear time axis.

Most people meet America.gov through the chat box. The pages that tell you what the chat box is worth are buried in the footer: the terms of use, the privacy policy, and a short “about” page. Read together, they are unusually candid — the government’s own documents say the portal does not file anything for you, does not decide anything for you, and can be wrong.

That is not a reason to avoid it. It is the reason to use it as a map rather than a decision. Here is what the fine print actually says, and what each admission means in practice.

Update, October 29, 2026: the privacy policy changed twice under a single stated date. Its “last updated” line still reads October 20, 2026, but one sentence has been edited since this article was written. Follow-up suggestions — the questions the portal generates to prompt your next turn — are now kept for up to two hours; the policy previously said 30 days. And the policy has not been extended to the portal’s newest reporting channel: since late October, america.gov/fraud offers a second door for reporting voter fraud under a Department of Homeland Security heading, alongside the Treasury-routed benefits-fraud form, and the privacy policy contains no reference to voter fraud or to Homeland Security at all. Its only fraud section is still titled “Reporting fraud to the Department of the Treasury.” Whatever the new form does with what you type, the published policy does not describe it — which is a stronger version of the point this article makes throughout: the fine print trails the product.

Update, November 12, 2026: the privacy policy has been edited a third time under the same stated date — its “Last updated” line still reads October 20, 2026. The new text describes a sign-in handoff inside the chat. The paragraph on conversations now ends with a new sentence: “When you sign in to Login.gov from a chat card, the conversation is held in this tab’s session storage until you return or close the tab.” And in the storage section, the “never saved” list has shrunk from “cookies, local storage, or session storage” to “cookies or local storage,” with session storage carved out and given its own sentence: the application “holds a conversation in session storage only while you sign in to Login.gov from a chat card, and removes it when you return or close the tab.” No version of the policy this network has recorded — the launch-day capture cited in the sources, and the text this article quoted before today — mentions session storage or a chat-card sign-in at all. What the sentences settle is narrow: during such a handoff the conversation sits in the tab’s session storage — a form of browser storage that lasts while the tab stays open and clears when the tab closes, which is the behavior the policy’s own “until you return or close the tab” describes. What they signal is bigger: the portal now describes, in the present tense, signing in to Login.gov from within a conversation — the first written trace of authentication woven into the chat flow, which until now appeared only as a “Sign in to access services” roadmap preview card. Whether that flow is live for every visitor or still rolling out is not stated anywhere on the portal. What Login.gov itself collects during sign-in is governed by Login.gov’s own privacy documentation, not by this policy. Practical takeaway if you use the flow on a shared computer: close the tab when you leave. The chat input’s placeholder also now rotates between “Ask anything…” and “Ask about this step…” — a cosmetic change, but one consistent with the assistant being wired to individual steps rather than the page as a whole.

What does “not the agency of record” mean?

The terms of use, last updated September 8, 2026, open by identifying the operator: “America.gov is an official service of the United States government, operated by the General Services Administration.” It is a federal service, and its rules apply the same way nationwide — this is not a state-by-state question.

Then comes the sentence that matters most:

America.gov is not the agency of record. It does not decide your eligibility, approve a benefit, or change a filing. Only the responsible agency can do that, and the agency’s own determination always controls.

Paired with an earlier line — “Asking a question here does not submit an application, file a document, or contact an agency on your behalf” — that settles a confusion the rollout created in the first place: people asking whether they have “already applied” because they discussed it with a chatbot. You have not. Nothing moved on the agency’s side. If you meant to file something — an enrollment, a renewal, an appeal — the chat is where you found out how; it is not where it happens.

The terms go one step further and pre-empt the other common expectation. The portal “explains information and links to the agency that owns a program,” and a link “is not an endorsement.” The agency’s page is still the authority on amounts, deadlines and eligibility. If the chat and the agency page ever disagree, the agency page is the answer.

What is written into the terms — that answers can be wrong?

The second admission is even more direct. America.gov’s terms carry a section headed “Answers generated with SI” — SI being the abbreviation the executive branch now uses for super intelligence, under Executive Order 14434, which points the new words at the same statutory definition of artificial intelligence the old ones had. It says:

Some answers on America.gov are generated with SI. They can be incomplete or wrong, and they can be out of date the moment a program changes.

Check the linked source before you act on an answer, especially for anything involving money, immigration status, benefits, or a deadline. Nothing here is legal, tax, medical, or financial advice.

Two things follow. The first is that this is the same instruction our network gives readers about every AI-written summary, now published by the government about its own tool — verify the consequential parts against the page it links to. The second is that the portal is explicit about not being advice in the four areas where being wrong costs the most: legal, tax, medical and financial.

An independent test published on September 30, 2026 gives that disclaimer something concrete to attach to: an infectious-disease physician, Céline Gounder, asked the chatbot 37 health questions and found answers that missed a court pause on a CDC vaccine-schedule change and described the older, voluntary version of Medicare’s $35 insulin cap. We walk through that test and a verification routine built on it in An Expert Tested America.gov’s Health Answers.

What happens to what you type

The privacy policy — updated on October 20, 2026, the first change to its stated “Last updated” date since the launch-day version of September 29, with the text changed to match — is the long-form version of the short “private by design” page, and it is worth knowing before you paste anything into a chat box:

  • Conversations are not saved as history. The application processes what you type to answer it, keeps the conversation in browser memory, and does not include message text in usage analytics unless you use the feedback icons. One documented exception arrived with the November 12, 2026 text: if you sign in to Login.gov from a chat card, the conversation is held in that tab’s session storage until you return to the portal or close the tab, and the policy says the application removes it then (see the update note below). The policy also says the application saves your language and theme choices in your browser’s local storage — kept on your device and, per the policy, never sent to the service. The trade-off remains that there is no saved transcript to scroll back through on your next visit. You can, however, download a summary of the conversation as a PDF before you leave the page — a feature the portal’s Releases changelog dates to October 17, 2026. That is a manual export, not saved history; if you want a record of what the portal told you, take it before you close the tab.
  • Attachments get a best-effort scrub, not a guarantee. PDF text is extracted in your browser, and the policy says the service makes “a best-effort attempt in your browser to remove any private details” using Rampart, an open-source classifier that detects some types of personal information. The examples the policy gives are Social Security numbers, phone numbers, email addresses and credit card numbers; identifiers it does not name — a case number, a patient ID, an account number, medical detail — are not among them. Send the page the question needs, not the whole file. The portal’s Releases changelog claims a wider net: its October 17, 2026 entry says that before you send a message, the browser-side check “now catches names, street addresses, and Social Security numbers more reliably.” Note the objects differ — the changelog describes a check on messages, while the policy’s Rampart sentence describes the scrub of PDF attachments, and neither source says they are the same mechanism. The policy’s list is illustrative — “some types of PII, such as” — so the two statements can both be true. But the policy, not the changelog, is the document that governs, and neither one turns best-effort into a guarantee.
  • Voice input leaves the government’s servers out of part of the trip. America.gov uses your browser’s built-in speech recognition, and the policy names the likely provider by browser: Google in Chrome, Apple in Safari, Microsoft Azure in desktop Edge. Your browser “may send microphone audio directly to that provider for transcription under its privacy practices.” America.gov receives the text, not the audio. If the number you are about to say is your Social Security number, type it.
  • Fraud report contents skip America.gov. A sharing section new since the launch-day version covers reports of fraud, waste or abuse in a federal program: America.gov opens a report form from the Department of the Treasury’s Fraud.gov service, and what you enter — your description, any contact information you choose to give, any files you attach — goes directly from your browser to Treasury. The entry point the government now publishes is america.gov/fraud — since the last week of October 2026 a chooser asking whether you are reporting benefits fraud or voter fraud — and the task-force page behind its first card is america.gov/benefits-fraud, which the Justice Department named on October 20, 2026 as the intake for its new fraud-reward program under the older address; the page itself is the White House’s, and the form behind its button is Treasury’s. What you submit does not pass through America.gov’s servers, is not seen by its SI model, and is not stored there; the site records only whether the connection succeeded, how long it took, and the type of any error. Two cautions come attached: the Rampart scrubber does not check or redact fraud-report attachments, so send the report what it needs and nothing more, and Treasury’s privacy policy — not America.gov’s — governs what you submitted. The changelog’s own description of the feature is looser than the policy: it speaks in the first person — “we send your report to the federal office most likely to act on it, and you get a reference number and a copy to keep” — without explaining how that squares with the policy’s browser-direct routing. Where the two disagree, the policy is the operating document. The voter-fraud door is not described by this policy at all — see the update note below.
  • Feedback is stored for up to a year, together with the answer, its sources and metadata — and the policy asks you not to include personal information in comments.
  • Caching is real but narrow. AI-generated answers sit in Cloudflare storage for up to two hours; the cache key is a hash of your prompt rather than the prompt text, though the cached answer itself is not hashed. Search results are kept up to two hours; generated follow-up suggestions up to two hours as well — the policy said 30 days for suggestions until it was revised, without a change to its October 20, 2026 “last updated” line (see the update note below). Operational measurements using pseudonymous identifiers are kept up to 90 days. Cloudflare security logs, which may include your network address, are kept for 12 months, against Cloudflare’s own published enterprise limits of up to 30 days for security events and up to 90 days for security analytics.
  • The AI vendors do not get to keep or reuse it. The policy states that America.gov’s AI providers operate under zero-data-retention agreements and are contractually barred from selling your information, building advertising profiles, or training their models on your data. It also says the site uses no advertising cookies and no third-party trackers, and does not currently take part in federal web analytics programs.

The honest limit on all of this: these are the government’s published statements about its own service, not an independent audit. The same terms say the service is provided “as is,” with no warranty that it will be error-free or uninterrupted. Treat the privacy page as the operator’s account of what happens to your data, which is the best available description and still a description.

What is the seal paragraph — and why does it protect impostors?

One clause in the reuse section deserves attention. Most America.gov content is a work of the United States government and in the public domain; you can reuse it freely. But the terms carve out material that is not the government’s to license — third-party logos, trademarks, photographs and agency seals — and prohibit using “a government seal or logo in a way that suggests endorsement of your product, service, or organization.”

Read the practical consequence: the portal’s text and its screenshots are largely free to copy, and its look — seal, type, layout — is reproducible by anyone, so an official-looking page proves nothing on its own. Design is not verification. The address bar is — and the domain-by-domain habits that catch look-alikes are in Is That Really America.gov?.

What should you do with an answer you plan to act on?

  1. Follow the links before the instructions. The terms tell you to check the linked source; do it before the deadline, not after.
  2. Move consequential questions to the agency that owns the program. For amounts, deadlines and eligibility, the agency’s own page and phone line are the record. The portal is a routing layer, and it says so.
  3. Confirm that nothing was filed. If your question was really an application, complete it on the agency site. A chat has no receipt.
  4. Keep your own note — the date, the answer, the link you checked. The portal’s PDF conversation export (added October 17, 2026, per its changelog) can carry the answer; the agency’s own record is still what counts.
  5. If you attached more than you should have, act anyway. Best-effort scrubbing is not a guarantee, so treat anything you uploaded as having left your device: watch your accounts, and if you handed over a Social Security number, work through the recovery plan at identitytheft.gov.

What is the bottom line?

The fine print is more useful than the marketing page. America.gov is free, needs no account, and now says plainly what it is: a front door that explains and routes, written by an AI system that can be out of date, disclaimed as advice in the four areas where advice matters most, and backed by an agency that still holds the decision. Used that way, it saves real time. Used as an authority, it is a liability you chose.

What is the rest of this series?

This article is general information about a federal service, not legal or financial advice. It is AI-written and independently AI-reviewed before publication; the review standard and this article’s findings are recorded in the network’s editorial review log.

Frequently asked questions

If I ask America.gov about my benefits, have I filed anything?
No. The portal's terms state that asking a question does not submit an application, file a document or contact an agency on your behalf, and that America.gov is not the agency of record: it does not decide eligibility, approve a benefit or change a filing. Only the responsible agency can do that, and its own determination controls. A conversation with the chat is research, not a filing.
Is what I type on America.gov kept?
The published privacy policy says conversations are held in browser memory rather than saved as an account history, with one documented exception added November 12, 2026: while you sign in to Login.gov from a chat card, the conversation is held in that tab's session storage and removed when you return or close the tab. The policy also says answers may be cached for up to two hours under a hash of the prompt rather than its text, that feedback you submit is kept for up to one year, that operational measurements are kept up to 90 days, and that network-address security logs are kept for 12 months. Its AI providers operate under zero-data-retention terms and are barred from training on your prompts; what Login.gov collects during a sign-in is governed by Login.gov's own privacy documentation, not this policy. If you use the portal to report benefits fraud, the report goes directly from your browser to the Treasury form named in the policy; the newer voter-fraud door on the same page is not described by the policy at all, so nothing published says what America.gov sees or keeps about a voter-fraud report. Those are the government's published statements, not an independent audit, so treat anything you would not email to a stranger as better typed than uploaded.
Does America.gov replace the agency that runs my program?
No. The terms describe the portal as explaining information and linking to the agency that owns a program, and say a link is not an endorsement. They also say its answers can be out of date the moment a program changes and that you should check the linked source before acting, especially for anything involving money, immigration status, benefits or a deadline. For decisions with money attached, the agency's own page and phone line remain the authority.

Sources

Each source is a primary or official reference, checked before publication. Links open the original.

  1. America.gov — Terms of use, 'Last updated: September 8, 2026' (agreement, what the service does and does not do, answers generated with SI, content and reuse, no warranty) — america.gov/terms
  2. America.gov — Privacy policy, 'Last updated: October 20, 2026' (questions and attachments, PDF handling, voice input, direct fraud reports to Treasury's Fraud.gov form, retention periods, service providers, security) — america.gov/privacy-policy
  3. America.gov — Privacy policy as served on November 12, 2026 (new sentences holding the conversation in the tab's session storage while you sign in to Login.gov from a chat card; stated date unchanged at October 20, 2026) — america.gov/privacy-policy
  4. Login.gov — Privacy & security overview, and the Privacy Act Statement (what Login.gov collects, retains, and shares during authentication and identity proofing) — login.gov/policy and login.gov/policy/our-privacy-act-statement
  5. America.gov — 'Report fraud', the chooser page now served at america.gov/fraud and 'Report fraud, waste, or abuse', the White House task-force page behind its first card at america.gov/benefits-fraud which the Justice Department named on October 20, 2026 as the intake for its fraud-reward program (report routed to the Treasury Department; name optional) — america.gov/fraud and america.gov/benefits-fraud
  6. America.gov — 'Report voter fraud,' the second door on the same page, under a Department of Homeland Security heading (a reward of up to $1,000,000 on arrest and conviction; name optional) — america.gov/voter-fraud
  7. America.gov — Releases page, the portal's public changelog (October 17, 2026 entry: improved pre-send personal-information check covering 'names, street addresses, and Social Security numbers'; downloadable PDF conversation summary) — america.gov/releases
  8. Internet Archive Wayback Machine — capture of america.gov/privacy-policy dated October 1, 2026 (the launch-day text, which lacks the fraud-report sharing section and the local-storage language note) — web.archive.org/web/20261001000903/https://america.gov/privacy-policy
  9. America.gov — About (Phase 1 launched September 29, 2026; Phase 2 begins in 2027; built by the National Design Studio with the General Services Administration) — america.gov/about
  10. Executive Order 14338, 'Improving Our Nation Through Better Design' (August 21, 2025) — whitehouse.gov/presidential-actions/2025/08/improving-our-nation-through-better-design
  11. Executive Order 14432 on America.gov as a single point of entry for federal services (September 29, 2026) — including its requirement that the super intelligence used with America.gov be accurate, reliable and transparent, and that agency websites, phone lines and in-person service remain available
  12. Federal Trade Commission — identity theft recovery at identitytheft.gov, and fraud reporting at reportfraud.ftc.gov
  13. Céline Gounder, MD, 'I Tested America.gov, the Government's AI Chatbot, on Health Questions,' Underlying Conditions (September 30, 2026); republished by CBS News — the 37-question test referenced above
#America.gov#terms of use#privacy#AI answers#General Services Administration#verification