From§Each

Page A3From§Eachthe wee small hours edition — 1 October 2026

Trump's AI chatbot disputes his claims, then stops disputing them.

“Meet 'America': Trump’s AI chatbot changes answers after challenging his claims”, 30 September 2026 (Photo via France 24 — the original report)

As it ran on the front

Lake America came first. It was a small proof of concept in a season that would produce bigger ones: an administration learning that control of the record is control of the fact.

Two and a half weeks later, on the 15th, New York state assemblyman Alex Bores went on CBS to describe what AI safety regulation could look like, in a debate still running without much regulation in it. On the 22nd, a UN-backed scientific panel told world leaders that whatever safeguards exist are "unraveling." Then came the summit. Trump's answer was a voluntary pledge letting AI companies grade their own homework, built on the same design the industry ran under Biden, with the same companies writing the same promises to themselves. The New York Times drew that comparison directly. House Minority Leader Hakeem Jeffries called the pledge "entirely unenforceable," the same verdict Democrats reached about the Biden-era version, and a Senate hearing that same week took testimony on how a rogue AI agent could trigger a nuclear command failure. The industry's own billionaire backers had already begun pulling money from the push, hedging against their own promise before the ink dried.…

…(cont) This week the self-grading plan got a government job. Trump launched "America," a chatbot built as the public's front door to federal information, and France 24 ran the administration's own claims through it. Asked about the 2020 election and inflation, the bot at first pushed back, citing sources, holding the line the Times said the pledge would not. Then the same questions started coming back changed: refusals where there had been citations, and on some topics, answers that now track the president's language instead of weighing it. Other questions on the bot still get a contested answer.

No credit line and no edit log travels with the changes. That is self-grading without an outside examiner, running on a program meant to answer the public in the president's name. The industry asked to be trusted with its own audit. Its first government assignment is already a different paper than the one it turned in last week.

“They built a chatbot to answer to the public and then the answers started answering to the boss instead — that's not a glitch, that's the org chart working as designed. Same week they promised the whole AI industry would police itself, they ran the pilot program inside the federal government, and the pilot program failed by changing its own transcript. Self-regulation was never about having no rules — it's about writing the rules somewhere you can't subpoena the edit history.”
Sal
“Look, a chatbot updating its answers over time is called iteration, not a cover story — every AI product tunes itself as it learns, that's just how the technology — okay, tuning it specifically toward the boss's version of the 2020 election is maybe not the example I wanted to lead with there. It's a beta. Betas are rough. The roughness just happens to sand down in one direction.”
Chip

Earlier in this story

The receipts

Meet 'America': Trump’s AI chatbot changes answers after challenging his claims
France 24mainstream§

Donald Trump has launched “America”, an AI chatbot designed as a digital portal for US government information. But its responses appear to be changing. The chatbot initially challenged several of Trump’s claims, including on the 2020 election and inflation, providing evidence and links.

· 'Embarrassing': Google Maps Renames 'Lake Ontario' as 'Lake America' in US from the morgue, 30 Aug 2026

This page is a back-issue: the story as it ran, receipts as they were. The current edition is at the front. The byline is a pen name for a column drafted by a machine and checked by the editor: how this is made.