Which tool to choose
The rule "always cite the sources with the link" you put in the instructions, so it applies to every answer. What matters, though, is whether your assistant actually searches the web: when it searches, the links are the ones it found; when it doesn't search, it can "reconstruct" references from memory, and that's where invented links are born. The verification tool is you with the browser: you open the links and check. For important topics, the safest tool is to give the AI the source yourself — by pasting it in — instead of asking it to find it.
How to do it
- Ask explicitly: "answer by citing the sources and include the links," and set it in the instructions as a fixed rule.
- When you get the answer, open the links: do they exist? do they lead where the AI says?
- Check that the source really says what the assistant attributes to it, not just that the link works.
- For important facts, prefer to give the reliable source yourself, or cross-reference several references, instead of trusting the first link.
A concrete example
Luca was preparing an article and had asked the AI for some data with the sources. The answer came with three authoritative-looking links. He opened one: nonexistent page. Another: it existed, but didn't contain the cited datum at all. Only the third was real and relevant. If Luca had copied the references without opening them, he would have published two false sources. Since that episode he always opens the links before using them: he learned that a link in the answer isn't a guarantee, it's something to verify.
When it does NOT work (and how to fix it)
If a link doesn't exist or is broken
It's a typical sign of an invented reference. Don't trust that piece of the answer: ask the AI to actually search for it online (if it can), or search for the source yourself. A link that won't open is worth less than no link, because it gives a false sense of safety.
If the link exists but doesn't say what the AI claims
It happens: the page is real but the attributed datum isn't there, or it's different. That's why checking that the link works isn't enough — it has to be read. If the source doesn't confirm, the statement remains unverified, and should be treated as such.
If the AI isn't searching online
Without web search, the references it produces are reconstructed from memory and more easily invented. Turn on the search if your assistant offers it, or provide the sources yourself. For current matters and precise data, an answer without search is the most at risk.
A tip from someone who really uses it
Change the question you ask yourself: not "are the sources there?" but "have I opened them and do they say that?". It's the difference between seeming rigorous and being it. A link in the answer reassures, and precisely for that it's dangerous when it's false: it makes you lower your guard. Get into the habit of treating every source as to-be-verified until you've opened and read it, especially if you'll use it for a job, a decision or something public. For the topics that matter, the most solid method remains flipping the flow: you're the one who gives the AI the good source, so it doesn't have to find it — and can't invent it.
Frequently asked questions
Why does an AI invent links and sources?
When it's not searching the web, it tends to fill the gaps with something plausible, and a "lifelike" reference can look real while not existing. It doesn't do it to deceive: it reconstructs what would statistically make sense. For you the result is the same, a source to verify.
How do I recognize an invented source?
By opening it. If the link won't open, or the page doesn't contain what the AI attributes to it, it's unreliable. There's no shortcut: verification is reading.
Can I avoid the problem?
Largely yes: turn on web search where it exists, and for important topics provide the sources yourself instead of asking the AI to find them. A source you give yourself can't be invented.
If the answer has the links, can I trust them and cite them?
No, and it's the mistake that puts false sources into circulation under your name. The presence of a link says nothing about its reality: the AI can produce nonexistent ones or ones that have nothing to do with what it claims, and it does so as naturally as with real links. Copying the references without opening them means taking responsibility for sources you haven't checked. Always open the links and verify that they say what the AI claims, before using them. Trusting a source just because it appears in the answer is the most common way to end up citing something that doesn't exist.