Using ChatGPT and Claude during a crisis: what is fine, what is not

What a general chat assistant is good at during a crisis

A crisis always brings a peak in writing and thinking: overviews have to be produced quickly, situations summarised and questions answered. That is exactly where a general AI chat assistant can strengthen a crisis team. Practitioner accounts of using such tools in Dutch crisis teams show they work well for quickly mapping the parties involved and their roles, for condensing a long stream of updates into a readable situation summary, and for handling routine questions that would otherwise absorb core team time. That saves time at precisely the moment time is scarce.

It matters to see what a general chat assistant is and is not. It is a generic tool that is good with language: summarising, structuring, rephrasing. It is not a crisis expert and it does not know your organisation, your scenarios or your current situation, unless you feed that information in yourself, every time. Every answer is therefore a starting point, not a finished product.

Why you never use it for the final message

The most important distinction a crisis team has to make is between a draft and a published message. AI is strong at producing a first version quickly: a draft message, a first outline, a rough summary. That same speed becomes a risk the moment a draft goes out unchecked. A chat assistant can confuse facts, state a detail too confidently, or strike a tone that does not match the gravity of the situation. So in all circumstances: an AI draft is checked and approved by a person before it goes out. The greater the time pressure, the more important that step becomes, not the sooner it is skipped.

Alongside general assistants there are also AI tools built specifically for crisis communication, rather than generic assistants usable for anything. Such specialised tools sit in a wider landscape next to general assistants: which form suits best depends on the organisation and the situation. The same principle applies there too: a draft is a draft, and a person judges what is actually published. Our article on crisis communication software compares the categories.

The line that is never crossed: sensitive data in a public assistant

Practitioner accounts of using chat assistants in crisis teams share a consistently level-headed tone: it works well for overviews and drafts, provided a person checks the result. Our editorial addition to that is just as important: never enter organisational or sensitive data into a public chat assistant without a data processing agreement in place. That applies to names of those involved or affected, internal files, confidential reports and any other data the organisation does not want leaving the building. A public chat assistant is not automatically a safe processor of such data, and under the time pressure of a crisis the temptation to "just paste it in" is at its strongest.

This is not only a matter of caution but also of legal obligation. A data protection impact assessment and the GDPR set requirements for processing personal data, and the EU AI Act adds requirements for the use of AI systems, certainly in a context with societal impact such as crisis management. Those obligations do not disappear because there is haste. If anything: precisely because a crisis brings haste, it is wise to have this line clear in advance, so nobody has to improvise with sensitive data in the heat of the moment.

Four practical steps for responsible use

The risks above are no reason to avoid AI chat assistants, but they are a reason to settle a few things in advance. These steps help a crisis team use ChatGPT, Claude or comparable tools responsibly.

1. 1
   Use AI for drafts and overviews, never for the final message. Let the assistant summarise situations, draw up overviews and answer routine questions, but always have every outgoing message checked and approved by a person first.
2. 2
   Never enter names, files or confidential details of those affected into a public chat assistant. Without a data processing agreement that information does not belong there, however high the time pressure. Take account of your data protection impact assessment and GDPR obligations, and of what the EU AI Act asks of the organisation.
3. 3
   Record in advance which AI tool your organisation may use and under what conditions. Settle this outside crisis time, with the right legal and privacy colleagues, so the crisis team does not have to work out during an incident what is and is not permitted.
4. 4
   Consider an AI connection that knows your own crisis data and context. A bare chat assistant with no background knowledge of your organisation produces a generic draft. An AI tool that carries your own crisis context produces drafts that are more usable and better fitted.

Sources

• Frankwatching: Crisis communication — ChatGPT strengthens the team (2024), in Dutch.
• Articles 33 and 35 of the General Data Protection Regulation — breach notification and data protection impact assessment.
• Dutch National Cyber Security Centre, reporting duties under the Cybersecurity Act. Accessed 22 September 2026.

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.