What America.gov Runs On, and Why It Matters Now

America.gov arrived on September 29, 2026 as a single front door to federal services. But the portal is not only a services story. It is also the largest public deployment of an AI system the US government has ever run — and the same afternoon it launched, the administration renamed the technology underneath it.

Understanding what America.gov runs on is now the difference between reading its answers as software output and reading them as an authority. The portal’s own documents are unusually clear about the second half of that and quiet about the first.

The launch, in one paragraph

The portal opened with a conversational front door: type a question in plain language, get an answer drawn from official government sources, follow links to the agency that owns the program. It draws on roughly 29,000 federal, state and local websites. Sign-in runs through Login.gov, the identity service millions of Americans already use for Social Security, IRS and other agency accounts. The transactional phase — completing forms, applications and renewals inside the interface rather than being handed off — is framed on the portal’s own roadmap as “Coming early 2027,” while the White House fact sheet from launch day said passport renewal and Medicare enrollment would arrive “later this year.” Those two official statements still disagree, and it is worth knowing that neither has been reconciled.

What the government requires of the AI

One sentence in the founding order is the only accountability standard in the launch documents. Executive Order 14432 requires that any Super Intelligence used in connection with America.gov be “accurate, reliable, and transparent.”

Three adjectives, no mechanism. The order does not say who tests the system, how often, against what standard, or what happens when an answer fails. That is not unusual for an executive order, which is a management document rather than a regulatory one — but it does mean the useful guarantees are the ones written into the portal’s terms of use and privacy policy, not the ones written into the order.

What the portal admits about its answers

The admissions are more specific than the promise. The terms of use carry a section headed “Answers generated with SI” — SI being the abbreviation federal documents now use for super intelligence — and say:

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 They can be incomplete or wrong, and they can be out of date the moment a program changes.

The same section instructs readers to “check the linked source before you act,” and states that nothing on the site is legal, tax, medical or financial advice. The terms also state that America.gov is not the agency of record: asking a question “does not submit an application, file a document, or contact an agency on your behalf,” and it does not decide eligibility.

Read as a description of the system, that is a fair summary of any retrieval-augmented language model: it can be fluent, current-sounding and wrong. Read as a policy, it places the verification burden on the person asking.

What it does not disclose

The portal does not name the models behind the assistant. Press reporting on the Chief Design Officer’s CNBC interview identifies Google’s Gemini and xAI’s Grok as the systems involved, and Google has confirmed its own involvement in a blog post; the government’s own pages decline to describe the architecture. If which vendor reads your question matters to you — for retention, jurisdiction or competitive reasons — reporting is the only available answer, and it should be attributed as reporting rather than as a government disclosure.

What the government does disclose, in the privacy policy last updated September 29, 2026, is the data handling: conversations are processed to answer them and held in browser memory rather than saved as account history; answers may be cached for up to two hours under a hash of the prompt; voice input uses your browser’s built-in speech recognition, whose provider the policy names by browser (Google in Chrome, Apple in Safari, Microsoft Azure in desktop Edge); and the AI vendors operate under zero-data-retention terms and are contractually barred from training on prompts. Those are the operator’s published statements about its own service, not an independent audit.

Why the rename is part of the story

On the same day, Executive Order 14434 told federal agencies to write “Super Intelligence” and “SI” in place of “artificial intelligence” and “AI,” while pointing the new words at the existing statutory definition in 15 U.S.C. § 9401(3). That combination produces something worth noticing about America.gov specifically: the portal’s terms of use were last updated September 8, 2026 — three weeks before the order was signed — yet they already carry a section headed “Answers generated with SI.”

The most plausible reading is mundane: the section was drafted in anticipation of the change and published under the new heading, or updated in place without a new date stamp. Either way, it is a reminder that on federal websites the vocabulary is now moving faster than the version dates at the bottom of the page. We walk through the order itself, including the 60-day deadline for a proposed federal definition, in Why Federal Websites Now Say “SI” Instead of “AI”.

And to be precise about a distinction that is easy to blur: America.gov is not governed by the White House Accord on Super Intelligence, the one-page voluntary safety commitment the President and six technology company chief executives signed the same day. The accord covers internal model controls inside those companies — see The Accord Behind the Rename. The portal’s own standard is the three adjectives in EO 14432.

What this means if you actually use it

Four practical consequences, none of which require you to distrust the portal:

1. Treat the answer as a search result, not a ruling. It is a routing layer. The agency that owns the program still decides eligibility, amounts and deadlines, and the terms say so.
2. Check the linked source on anything with money or a deadline attached. The terms ask for exactly this, and it costs one click.
3. Assume nothing was filed. If your question was really an application, complete it on the agency’s own site — there is no receipt from a chat.
4. Do not upload more than the question needs. PDF handling includes a best-effort attempt to strip personal details using an open-source classifier; best-effort is not a guarantee.

The bottom line

America.gov is a genuinely useful front door and, by its own terms, a fallible one: an AI system built by the government on top of commercial models, required by executive order to be “accurate, reliable, and transparent,” described by its own terms as something that “can be incomplete or wrong,” and renamed mid-flight along with the rest of the federal vocabulary. Use it to find the right agency and the right form. Let the agency’s own page be the record.

Byte is the technology site of the Omni Mundi Compendium network. The launch itself, and what it changes for federal services, is covered on Counsel in What Is America.gov? and What America.gov’s Own Terms Say.

This article is general information about a federal service, not legal 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.