Reading the source links an assistant shows and checking them

A source link is a claim, not proof. It says "this is where I got it." Whether the passage really backs the answer is something you can only learn by opening it and reading.

What the links are

When a tool shows sources, it's pointing to the chunks it fetched. AWS says the output can include citations or references to sources, and that users can look up source documents themselves if they require further clarification or more detail.¹ Wikipedia makes the same point: this lets users verify the cited sources.²

Microsoft's documentation describes a feature where built-in citation tracking shows provenance.³ Provenance means where something came from. In plain terms, you're being shown the trail back to the file.

How a tool displays this varies. It might be a footnote number, a file name, a page number or a highlighted excerpt. The maker's help pages describe what yours shows.

Check in five moves

• Open the source. Don't stop at the file name. Open the file or the excerpt itself.
• Find the passage. If the tool shows a page or a highlighted excerpt, go there. If not, search the file for a distinctive word from the answer.
• Compare the details. Look at names, amounts, dates, conditions and exceptions. A match on the topic isn't enough. The numbers and the small words, like "except" and "only," need to agree too.
• Check the file is the current one. A correct quote from last year's price list is still the wrong answer today. Look at the review date if the file has one.
• Decide. If the passage backs the answer, you can rely on it as far as you'd rely on the document. If not, correct the answer and tell whoever manages the library.

When a link looks fine but isn't

The passage says something close, but not the same. Wikipedia warns that a model may extract statements from a source without considering its context, resulting in an incorrect conclusion.² A sentence about "standard delivery" can get attached to an answer about express orders.

The link points to a file but not to the answer. The right file was found, but the model wrote more than the file says. The extra part is unsupported, however confident it sounds.

The source doesn't exist or doesn't say that. Wikipedia notes there are many noted instances of language models citing sources that are either incorrect or do not exist.⁴ It also describes fabricated citations as one form of plausible-sounding falsehood.⁴ So follow the link and make sure it opens something real.

No link at all. If the tool normally shows sources and this answer has none, treat the answer as unchecked and look it up yourself.

Which answers deserve a check

You can't check everything, and you don't need to. Spend the effort where a wrong answer costs something:

• Prices, fees and refunds.
• Deadlines, opening hours and booking rules.
• Anything told to a customer in writing.
• Legal, tax, privacy and safety questions. For these, go to the official source for your country, not to the assistant.

For low-stakes questions, such as where the staff roster lives, a quick glance at the link may do.

If you check a few answers and find a pattern, such as always pulling from an old file, why answers go wrong helps you work out the cause. And never paste passwords, API keys, customer records or card numbers into a question while you're testing.