Search

FOUNDATIONS

Sources: Where the answer gets its knowledge

A convincing answer can quickly sound like established knowledge. But language models initially generate text from patterns learned during training. They do not automatically know which specific source sits behind an individual claim.

For current or verifiable questions, many AI assistants can also search the web, analyse documents or link to sources. That makes answers easier to inspect – but it does not automatically make them correct.

The evidence behind the answer

Sources are more than a footnote. They determine whether you can check a claim yourself – and how much confidence it deserves:

  • Plausible is not proven: An AI can produce a factual-sounding explanation even when individual details are wrong or outdated – or simply sound plausible. Without a source, it can be difficult to see where knowledge ends and reconstruction begins.
  • Not all sources are equal: An original study, a law or an official publication carries a different weight for many questions than a blog post, press release or social-media post. What matters is not only whether a source is cited, but which source it is.
  • Even real sources can be used badly: A link can be genuine and still fail to support the claim. The source may be outdated, support only part of the statement or be taken out of context.

THE RUNNING EXAMPLE

Mara plans a trip

Mara asks whether Lisbon is generally a suitable destination for a city break in October. General knowledge about the climate, the city and its sights can help with that first orientation.

Then she wants something more specific:

“How much is the flight on 14 October, and is the museum open on the day we plan to visit?”

Now she needs up-to-date evidence: the museum’s own website for its opening hours and a current booking offer for the flight price. A general description of Lisbon cannot answer these concrete questions.

What matters is not only whether a source is cited. It must support the specific claim and be current enough for the question.

Your leverage in everyday use

  • Ask for sources: For verifiable claims, ask the AI for concrete evidence – ideally with an original source, a date and a direct connection to the claim.
  • Open the most important source yourself: Do not rely only on the AI’s summary. For important questions, check at least the central evidence yourself.
  • Prefer primary sources: Whenever possible, go one level back: from the news story to the study, from the blog to the law, from the claim to the original publication.
  • Compare the evidence with the claim: One particularly useful instruction is:

“For every central factual claim, give me the best available source and say explicitly when a source supports the claim only partially.”

The key point: A good answer does not only tell you what to believe. It gives you a way to check for yourself why you should believe it.