How Northtape verifies every AI citation

An AI summary of a news story is only as trustworthy as the sentence you cannot check. Plenty of tools will hand you a tidy three-line takeaway on a market story, and most of the time it reads fine. The trouble is the times it does not: a confident sentence the article never actually said, delivered in the same calm tone as the true ones. Northtape's answer to that is narrow and mechanical, and it is worth understanding before you rely on any AI summary, here or anywhere else.

What does Northtape's citation check actually do?

It requires every quote in a summary to be a word-for-word match against the original article, and it deletes any quote that is not. When Northtape summarises a news item, the AI model is asked to return not just the takeaway but a set of short quotes copied verbatim from the article body, each tagged with the claim it supports. Before those quotes reach you, the code checks each one against the text it fetched from the source. A quote that does not appear, exactly, in that text is removed. The model can phrase the summary however it likes; it cannot invent a quotation and have it survive the check.

Why check the quotes instead of trusting the model?

Because a language model can produce a fluent sentence that no source ever contained, and fluency is not evidence. This is the failure people mean by AI 'hallucination': not gibberish, but plausible, well-formed text that happens to be untethered from any source, and it is dangerous precisely because it reads like the true parts. Instructing the model to be careful helps a little. Checking its output against the actual article helps far more. Northtape does both: the model is told that every point must be grounded in the article and that it must not claim anything the article does not say, and then the substring check verifies the quotes without relying on the model having cooperated.

What happens when the article can't be read?

When Northtape cannot fetch the article's text, it returns no citations at all rather than guessing. Some pages block automated readers, time out, or sit behind a paywall. In those cases there is nothing to match quotes against, so the system clears the citation list to empty instead of letting the model supply quotes it cannot verify. You may still get a short summary drawn from the headline, but you will not be shown a quotation the system was unable to confirm. An absent citation is treated as more honest than an unverifiable one.

How does Northtape stop a summary being faked?

It resolves each article's title and source on the server from its own news feed, never from whatever a request supplies. The summary cache is shared across everyone reading the same link, so a single spoofed title could otherwise poison what every other reader sees for that URL. To close that gap, the code ignores any title or source sent in the request and looks up the authoritative version from the current feed. It also re-checks the page after any redirects and refuses to pull content from a private or internal address. Verification here is not only about the model's honesty: it is also about removing the ways the input itself could be gamed.

What this does, and does not, promise

Verified quotes make a summary checkable, not infallible. A quote can be genuine and still come from an article that was itself wrong, and a summary can be accurate about a story that later turns out to be incomplete. What the citation check gives you is narrower and still worth having: reasonable confidence that the words in quotation marks were really in the source, plus a path back to the original so you can read them in context. The claim Northtape is willing to stand behind is that the receipts are real, not that the news is always right.

FAQs

How does Northtape verify an AI summary's quotes? Every quote must be an exact, word-for-word match to the source article. After the model returns its summary, the code checks each quote against the fetched article text and removes any that does not appear in it, so an invented quotation cannot reach you.

What is an AI 'hallucination' in this context? It is plausible, fluent text a model produces that no source actually supports. It is risky precisely because it reads like the true parts, which is why Northtape matches quotes against the original rather than trusting the wording alone.

Why do some summaries show no quotes? Because Northtape could not read the underlying article, so it had nothing to match quotes against. Rather than let the model supply unverifiable quotes, the system returns none, treating an absent citation as more honest than an unconfirmed one.

Does a verified quote mean the news is accurate? No. It means the quoted words were genuinely in the source article, not that the article itself is correct or complete. The check makes a summary checkable and gives you a way back to the original to judge for yourself.

Northtape's AI summaries are there to help you read the market's news faster, not to tell you what to do about it. They describe what an article says and let you verify it: they do not offer investment advice, and a verified quote is a starting point for your own reading rather than a conclusion. This piece describes how the citation check works at the time of writing, and the mechanics may change as the product develops, so treat the detail here as a description of the approach rather than a fixed guarantee.