America.gov's AI Chatbot and the White House AI Accord

The United States government now has an AI chatbot. It runs on Google's Gemini and xAI's Grok at the same time. It launched on America.gov, a new portal sitting on top of 29,000 federal websites.

On the same day, every big AI CEO signed a safety accord at the White House. Some people are already calling it an AI constitution.

Two stories, one week. Both are more useful to a small business owner than they sound.

America.gov runs its chatbot on Gemini and Grok

Here is the setup. The US federal government launched America.gov with a chat agent on the front page. Ask it something, and it points you to the right official service or form.

Behind it sit two models from two companies. Gemini from Google, and Grok from Elon Musk's xAI.

The point is the 29,000 websites underneath. Government information has been scattered across thousands of separate portals. Each one has its own menu, its own search and a design from some forgotten decade.

Nobody was ever going to redesign all of them. So they put one chat box in front of the pile.

It is very strict about scope. Ask it a general question about how to be more free, and it says no. It only handles official US government services.

Ask for a cheesecake recipe and it declines, politely, like a civil servant who has seen things.

That strictness is the smartest part. Most people want their bot to be chatty and helpful about everything.

Then someone on the internet talks it into doing homework or promising a refund nobody approved. A narrow bot is boring, and boring is the goal.

The two-model setup is worth copying too. If your whole operation runs on one provider, you inherit their outages and their price changes. Apple reportedly works with both Anthropic's Claude and OpenAI for a reason.

You don't need a big engineering team for this. Just keep your prompts and your logic separate from the model call. Then swapping providers is a setting, not a rewrite.

There's a fair question about private models running public services. Anthropic's Dario Amodei has said he'd rather this work sat inside government than inside companies like his.

Nice thought. But no government is going to train a model that beats Gemini any time soon. The real alternative is leaving 29,000 broken websites exactly as they are.

For your business, the lesson is the same at any size. People do not want to learn your navigation. They want to ask a question and get a link.

That's the whole value of a well-scoped AI agent on your website. You don't need a national launch event for it.

The White House AI safety accord, in plain words

Same week, Donald Trump got the whole industry into one room. Elon Musk, Nvidia's Jensen Huang, Anthropic's Dario Amodei and OpenAI's Sam Altman were there. So was Microsoft's leadership, plus a long tail of CEOs you'd only know from a funding headline.

They came out with a signed document about how frontier AI should be built. It says each company is responsible for developing its own technology safely.

It also says each company should build trust with customers and the public. Then it lists what that means in practice:

• Strong internal controls to monitor the capabilities and alignment of models during training and deployment
• An internal team with real authority to check that those controls, monitoring and detection actually work
• A partnership with an independent outside auditor or evaluator to run its own assessments
• An independent committee of the board that oversees the teams running the controls and makes sure problems get fixed

The document also says it makes sense, over time, to turn these steps into laws. Over time. Lovely phrase. Does a lot of work in a sentence.

I'm not going to pretend this is nothing. Getting direct competitors to agree on what safe deployment means is genuinely hard. The external auditor line is the real content.

An internal safety team reports to people whose bonus depends on shipping. An outside evaluator does not. That's the difference between a check and a vibe.

But notice what the accord actually is. It's a promise to grade your own homework carefully. The tutor is one you hired. The board committee is one you appointed.

There is no regulator, no penalty and no deadline. If a model misbehaves, the agreed consequence is that a committee receives a report about it.

The funniest part is Amodei signing it at all. The document says private companies will handle this responsibly. He said recently he'd rather this were not left to private companies. I'd have the same face he had in the photos.

And Europe? Not in the room. We have the rules and none of the labs.

I live in Croatia and I run a company here. I'd love to report that the continent is making its move. It is making paperwork.

What this means if you're buying AI, not building it

Most readers here are not training frontier models. You're deciding which tools to put into your business. Both stories matter to that decision.

First, provider safety practices are becoming visible. Big labs now commit to outside evaluations and board oversight. So you can ask about it.

When a vendor pitches you an AI product, ask three things. Who tests it? What happens when it fails? Who inside their company is accountable?

Watch how fast the slide deck runs out of answers.

Second, do your own small version of the accord. It takes an afternoon. Write down which AI tools touch customer data. Name one person responsible for each.

Log what the systems do, so you can check later. Decide in advance what a model is never allowed to do alone.

Refunds and anything legal are good places to start. That's the small-company edition of a billionaire summit.

Third, keep a second model ready. The US government is hedging between Gemini and Grok on a public service. You can hedge on a support bot.

My prediction: the America.gov chatbot will be more useful than the accord within six months. One of them answers questions about forms. The other one promises that a committee will be informed.