An AI hallucination is not a glitch you click away. For municipalities, safety regions and other public organisations it is a governance risk: an AI system that produces a convincing but incorrect answer under time pressure can damage public trust directly. This article explains what a hallucination is, why crisis communication is especially vulnerable to it, who is accountable when it goes wrong, and which safeguards keep the risk manageable.
What an AI hallucination is, and why it differs from an ordinary error
A hallucination is output from an AI system that is factually wrong or invented, but presented in the same confident tone as a correct answer. The system is not deliberately lying: it generates language on the basis of probability, not on the basis of verified knowledge. The result can be a non-existent source, a wrong figure, an invented quotation or a detail that sounds plausible but is simply untrue.
That makes a hallucination fundamentally different from an ordinary error. A crashing application or an error message is visible: everyone sees immediately that something is wrong. A hallucination is invisible until somebody checks the fact. For senior leaders and communication teams that is precisely the problem: this is not a curious technical defect for the IT department to fix, but a governance question. Teams must be able to account for and defend every AI-generated answer, and a wrong statistic or an invented fact can undermine public trust, certainly under the time pressure of a crisis.
Why this is especially dangerous during a crisis
Outside crisis time there is usually room to check an AI answer calmly before it goes out. During a crisis that room is often missing. There is time pressure: the public, media and senior leaders want clarity immediately, and a communication team using AI to draft faster feels the pull to publish an answer quickly without checking it thoroughly.
Visibility during a crisis is also far greater than normal. A message that would otherwise be read by few people is, during a crisis, shared, commented on and scrutinised by media, regulators and the people involved. An invented casualty figure, an incorrect location or a wrongly named authority stands out faster, and spreads faster too.
Finally, there is little room to recover. In ordinary circumstances you can correct a mistake calmly. During a crisis every correction touches the organisation's credibility directly, at exactly the moment trust is most needed. That combination of time pressure, visibility and limited room to recover makes an AI hallucination in crisis communication a risk of a different order than in everyday use.
The underlying governance question: who is accountable for an AI error that goes public
Here the hallucination question touches a larger issue: accountability. An AI system publishes nothing by itself, but who answers for it when an invented fact has gone out anyway? Not the supplier of the language model, and not "the technology": accountability sits with the organisation publishing the message and with the person who approved it. That sounds obvious, but in the reality of a busy crisis team that accountability blurs easily, precisely because an AI draft looks so finished and self-assured.
Practitioner accounts of using AI assistants in Dutch crisis teams stress above all that it works for drafts and overviews, provided a person checks the result — the garbage in, garbage out principle. Our editorial addition, following directly from the hallucination risk: never enter organisational or sensitive data into an AI tool, because that data can end up elsewhere or be used in ways the organisation does not control. Also give structural attention to a data protection impact assessment and the GDPR, and to what the EU AI Act means for using AI in an environment with high societal impact such as crisis management. Those two points are not a separate legal footnote; they belong to the same question: who is accountable, and can that accountability be demonstrated if something goes wrong?
Practical safeguards
The good news is that the risk of AI hallucinations in crisis communication is manageable, provided you build in a number of safeguards structurally.
Human in the loop. This is the most important practical countermeasure. AI generates a draft, but a person checks the facts and approves the message before it goes out. No AI output leaves the organisation without that human step, however great the time pressure.
Check and cite sources rather than letting them be invented. Have an AI system refer to or quote from verified sources wherever possible, such as your own crisis information or official figures, rather than leaving it free to formulate. The less room there is to guess, the smaller the chance of a hallucination.
A fixed review moment before publication. Make the fact check not a loose habit but a fixed, unskippable part of the process, even when the clock is ticking. That matches how good crisis communication already works: structured, with a fixed decision point, not ad hoc.
Be transparent if it does go wrong. If an incorrect AI-generated fact has been published after all, acknowledge it quickly and clearly, correct the message and explain what happened. That openness is exactly the founding principle of good crisis communication, and it does not change because AI was involved in the drafting.
Never publish without review
CrisisMaatje is deliberately built never to publish by itself: it delivers a draft, you decide and check.
Sources
- Mysidewalk: Why AI Hallucinations Are a Governance Problem, Not a Tech Problem.
- Frankwatching: Crisis communication — how ChatGPT can strengthen your team (2024), in Dutch.
- Articles 33 and 35 of the General Data Protection Regulation — breach notification and data protection impact assessment.
About this article
Published by CrisisRadar B.V.
This article has been checked against primary sources: Dutch crisis management practice, European regulation and academic literature. The sources used are listed above, with dates. Spotted something that is wrong? Tell us; we will correct it and record the change. Read how we work on Editorial policy.
Try it yourself with CrisisMaatje
CrisisMaatje connects the AI you already use — Claude, ChatGPT or Copilot — to the CrisisRadar platform. You ask for help in plain language, your AI drafts a message and records the facts, the timeline and the statutory reporting deadlines in a crisis file. You review it and decide what goes out. Free, no credit card.