The analogy

Think of the neighbor who knows a bit about everything. You run into them on the landing, you ask how that restaurant is or how to remove a stain, and they answer right away, with confidence, often with good advice. For general things you trust them gladly. But if they tell you "that office opens at eight," before showing up at eight on the dot you make a phone call: the neighbor speaks with the same confidence whether they know or whether they're misremembering.

AI is that neighbor. Extremely useful for getting your bearings, for explanations, for reasoning together. Unreliable as the sole source on a precise fact that costs you dearly to get wrong. You can't tell the difference from the tone, because the tone is always confident: you can only tell it by checking.

How it really works

AI builds its answers by predicting plausible text, and it mixes seamlessly what's correct with what it made up. There's no visible signal separating the two: the true sentence and the false one have the same grammar, the same fluency, the same air of competence. That's what makes trusting "by feel" useless. Reliability can't be read in the answer, it's established by verifying the points that matter.

What you can do in practice

A small procedure, from the quickest check to the most important:

  • Ask for the sources and open them. It's not enough that it cites them: the link must exist and actually say what the AI claims.
  • Isolate the verifiable details (a figure, a date, a name, a regulation) and check them against an independent source. They're the points where error hides.
  • Ask it to critique itself: "what in this answer could be wrong or uncertain?". Often it pulls out its own weak spots.
  • Turn on web search for anything recent or that changes over time (prices, laws, events).
  • Look at internal consistency: if the numbers don't add up among themselves, it's a signal that nothing adds up.
  • For health, money and legal matters, AI is a starting point for understanding and forming your questions, not the source to decide on.

A common misconception

"If it cites a source, then it's true." It's the misconception that suddenly lowers your guard. AI can invent a source out of whole cloth, attribute it to the wrong author, or cite a real one that doesn't say at all what it claims. The presence of a citation isn't a guarantee: it's an invitation to click. A source you haven't opened is worth as much as no source.

Frequently asked questions

Can I tell the true from the made-up at a glance?

No, and that's the point to accept. There's no signal of style that warns you: AI writes the invention with the same fluency as the correct fact. The only way is to verify the specific details, not rely on the impression of competence.

Are the sources it cites reliable?

They must be opened and read, one by one. Some will be solid, others weak, others might not exist or might not contain what the AI states. The citation tells you where to look, not that the checking work is already done.

For medicine or legal matters can I trust it?

As support, yes; as a decision, no. Use it to understand the terms, prepare the right questions, frame a problem. But the diagnosis, the treatment and the legal act stay in the hands of a professional who is accountable for what they say. AI is accountable for nothing.

If two different AIs tell me the same thing, then is it true?

Not necessarily. Different models were trained on masses of text that overlap heavily, and they can share the same widespread errors found online. Two matching answers can be wrong in the same way. Agreement between AIs isn't a verification: the verification is a reliable human source, outside the AI.