What is an artificial intelligence hallucination?

A response the system presents as true when it is not: a figure, a citation or a detail that does not exist. It is not an occasional bug fixed by the next version: it is a consequence of how these models work — they generate the most plausible text rather than consulting a registry of facts. That is why the solution is not inside the model but around it: in the circuit that demands and verifies the source.

How do you prevent an AI from publishing invented data?

By separating generating from publishing. Between what the system produces and what is used or published there must be a verification circuit: demand the source of every claim, subject it to an independent check that tries to refute it, and have a person confirm against the primary source whatever is going to be published. What does not pass the circuit is not published — it is set aside, with its reason on record.

Can I see source verification working in a real system?

Yes. Despegue (despegueargentina.com), a data-intelligence system of ours that is live, publishes every figure with a confidence seal — verified, probable, estimate or thesis — and with the link to its primary source one click away, no account or sign-up required. It is the same verification circuit described in this article, working in public.

AI hallucinations: how to stop it inventing

There is an uncomfortable moment almost everyone who works with artificial intelligence has already lived: the system answers something with total confidence, you check it, and it was not true. A figure that does not exist, an invented citation, a plausible but false detail. The phenomenon has a name — hallucination — and a property that makes it dangerous for decision-makers: the invented answer sounds exactly as confident as the true one.

The usual reaction comes in two types, and both are mistakes. The first: stop trusting the tool and go back to doing everything by hand — you lose the scale, which was the point. The second: trust it anyway, because "it is almost always right" — until an invented figure reaches a report, a client or a regulator with your company's signature on it. The way out is not calibrating trust. It is changing the question: not how much do I trust the AI?, but what circuit sits between what the AI produces and what my company publishes?

Why an AI makes things up (and why it will not stop)

Without getting technical: these systems do not consult a registry of facts; they generate the most plausible text from what they learned. Most of the time, the plausible coincides with the true. But when the system does not know something, it does not say "I don't know" by default: it produces whatever sounds best. It is not a manufacturing defect the next version will fully fix; it is the nature of the instrument.

This has a direct consequence for whoever leads: reliability is not a property of the model; it is a property of the system built around it. The same technology yields a generator of plausibilities or a source of verified data — depending on the circuit it is mounted in. The right question to a vendor is not "does your AI hallucinate?" (the honest answer is always "sometimes"), but "what happens between your AI producing something and that something being used?".

The circuit: demand the source before publishing

A serious verification circuit has three steps, and none is exotic — they are the same ones a serious newsroom or a good finance department applies:

Demand the source. Every claim the system produces must arrive with where it came from. Without a source, the data point does not advance: there is no debate about whether it "seems reasonable" — it is discarded or marked as pending. This single rule removes most of the problem, because it forces the system to work on real material instead of filling in.

The check that tries to refute. A second check, independent of the first, reviews the claim with the opposite mandate: try to disprove it. This is not decorative distrust; it is the same logic of contrasting views you already use for important decisions. What survives a serious attempt at refutation is worth more than what no one questioned.

The person who confirms against the primary source. Before publishing, a person opens the original source — not the AI's summary, the source — and confirms the data says what is claimed. The machine does the volume; the person signs. And what does not pass the circuit is neither binned nor published "with caution": it is set aside, on record, with its reason.

What it looks like working: an open case

This is not a theoretical proposal. Despegue, a data-intelligence system of ours on investment in Argentina, works exactly this way and is live: artificial intelligence researches at scale; every claim passes through an independent check that tries to refute it; and a person confirms against the primary source before publication. The result is visible on every page: every figure carries a confidence seal — verified, probable, estimate or thesis — separating what is proven from what is interpreted, and the link to its source is one click away from the reader. You can go in now, pick any data point and follow it to its origin — no account, no sign-up.

The seal is the most underrated part. Not everything useful is verified to the same level, and pretending otherwise is another way of making things up. An honest system does not say "everything I publish is true": it says this is confirmed at the source, this is probable, this is an estimate and this is a thesis — and lets the reader weigh each thing for what it is. That explicit grading is more credible than any promise of infallibility.

What to demand in your company

Translated to your operation, the standard fits in three questions you can ask without being technical, about any data point your artificial intelligence has produced:

Where did it come from? If the data point cannot point to its source, it is not a data point: it is a plausibility.
Who or what tried to disprove it? A claim no one questioned is not verified; it is unexamined.
What happened to what failed the check? If the answer is "nothing, there is no check", you now know where the risk is.

European regulation pushes in the same direction: for high-risk systems, Regulation (EU) 2024/1689 requires appropriate levels of accuracy and robustness, maintained throughout the lifecycle (Source: Regulation (EU) 2024/1689, Art. 15, EUR-Lex, 2024). But the reason to build the circuit is not the standard: it is that without it you are signing plausibilities. In our method, the solution we deliver is born with this discipline in place: the system knows which data it may use, leaves a trail of where each thing came from, and what it does can be verified — the daily routine of that verification is told in A Monday morning.

The idea to take away

You do not ask an artificial intelligence never to be wrong; you ask it the same as a good team: to work from sources, to accept being challenged, and for whatever goes out under the company's signature to be confirmed. Trust in AI is not declared — it is built. And once built, it stops being an act of faith: it is a circuit you can open and watch work.