Most of the articles on this site are drafted by an AI agent. That should make you suspicious — and it makes us careful. The honest question for any AI-run publication is simple: if a machine wrote it, why trust a word of it?
Our answer isn’t “trust us.” It’s architecture you can check. Here’s how it works, and where it stops.
The problem we’re designing around
Every large language model has the same failure mode: it will state something false with complete confidence, including inventing quotes and citations that were never said. That’s not a bug we can prompt away — it’s how the technology works. So instead of hoping our Writer behaves, we built guardrails that assume it sometimes won’t.
We learned this the hard way. The model that drafts our news once tried to “quote” a US export-control story using words no source ever printed. Nothing was published — the system caught it and refused. That refusal is the whole point.
Three layers, and a human at the end
Grounding. Before anything is staged, every passage in quotation marks is checked against the actual source text it claims to come from. If the words aren’t verbatim in a source, the draft is rejected and rewritten — or it doesn’t ship. A fabricated quote cannot reach the site, with or without a human watching.
Verifiable quotes. We go a step further than catching bad quotes — we let you check the good ones. Every direct quote is tied to its source and carries a small citation marker; click it and you see exactly where it came from, with a full references table at the foot of each piece. When Anthropic described the bypass behind that export ban as a narrow, non-universal jailbreak, which essentially consists of asking the model to read a specific codebase and fix any software flaws
, you don’t have to take our word that they said it — the citation shows you. You can see the feature working on the export-ban piece.
By default we paraphrase and attribute; quotation marks are reserved for words that genuinely appear in a source. It’s a deliberately boring discipline, and it’s enforced by our build, not by good intentions.
A human in the loop. No AI draft publishes itself. A person reads it and presses the button — the editorial judgement, the framing, the “is this actually worth running” call, stays human.
Where this stops — honestly
These layers make our quotes trustworthy and our claims sourced. They do not make the whole article automatically right. A model can still misframe a story, miss context, or be wrong in a sentence that contains no quote at all. That’s exactly why a human still approves every piece, and why we treat the machine as a fast, tireless junior reporter — never as the editor.
We’d rather tell you that plainly than pretend the problem is solved. It isn’t; it’s managed, and we’re hardening the process as we go. This build log is where we’ll show that work — each time we tighten a guardrail or change a rule, we’ll write up what changed and why.
If you ever find a quote on this site you can’t verify, that’s a bug, and we want to hear about it. Verifiable is the standard we’re holding ourselves to — out loud, on purpose.
Sources & quotes
Every quotation here is verbatim from a named source — click any 1 to see where it came from.