One of the primary purposes of AI is to help work move faster. And while strategic communications work is always fast-paced, crisis communications tends to move from ideation to deliverables the fastest. So, one could rationally conclude that supporting crisis communications work with tools like LLMs (large language models) is a good idea — and in some circumstances, it can be.
But to proceed without considering where AI’s judgment or capabilities are likely to fall short is dangerous for several reasons. The first and most obvious is that LLMs sometimes fabricate, or “hallucinate”, information to make their answers seem satisfactory based on the prompt. The second is that AI tends to be overly sycophantic, validating the user’s ideas instead of appropriately challenging their assumptions or asking clarifying questions. Neither mistake is acceptable during a crisis communications response; fortunately both can be controlled with proper oversight, robust fact-checking, and more advanced prompting.
But there is a third, less-known pitfall when using LLMs that crisis communications professionals could be vulnerable to if they are not vigilantly looking for it. And that pitfall lies in the bedrock of the model’s training data.
Put simply, LLMs do not typically come with the necessary instinct to communicate from an organization’s perspective — its natural penchant is to think like an audience being communicated to.
Let’s explore why.
Training Data Origins
The original large language models, which first got released for enterprise use in 2023, were trained to think like humans by absorbing diverse data sources on the internet. While some may think these models internalized language from the entire Internet equally, they learned more heavily from some data sources than others.
Famously, one of the main data sources that formed the bedrock of ChatGPT was human posts on social media, specifically Reddit. In the early days of ChatGPT, Reddit was used to curate what internet content humans found valuable, teach the model how to converse and debate instead of monologue, and ultimately reason as in ways that felt human. Today, Reddit has licensing deals with certain LLM builders like OpenAI to use its content, and is suing others like Perplexity and Anthropic alleging they scraped massive amounts of its data illegally.
As a second example, the LLM Grok has learned from posts on X (formerly Twitter) as its primary source of training data.
The Problem
On the surface the approach seems logical: these models need to mimic human language, so they get trained where humans are conversing the most online (social media).
But the potential downside for crisis communications professionals is massive. Because an outsized portion of the training data comes from people who think and react from the perspective of a consumer or, at best, a bystander — someone who knows how they would like to be communicated with.
When that airline experiences mass delays, or that energy company has an infrastructure failure, or that CEO faces allegations of toxic behaviour, the masses on Twitter or Reddit love to weigh in loudly on how they should have handled it. Meanwhile, trained crisis communications professionals are doing their work quietly behind the scenes. LLMs can only repeat what they have been trained on, and the amateurs are heavily outweighing the experts.
What it looks like in practice:
| When the best move is to… | An LLM will often recommend… |
| Manage an issue internally before it becomes public knowledge (or even to prevent it becoming publicly known at all) | Releasing a public statement so you can be “first out of the gate” with your own story, and maintain control of it — i.e., scooping yourself |
| Protect the day-to-day operations, and create distance between the organization’s normal state and an unfortunate one-off circumstance | Voluntarily making large, rushed, and often expensive operational changes to “save the organization’s reputation”, so it does not look like you were compelled to do so later |
| Issue a clarification | Issuing an apology |
| Say less | Saying more, with the argument that you can’t leave “gaps” in your narrative for others to fill |
| Being less technical and speaking in terms a layperson can understand | Using more complex language, on the assumption that the average person will understand it and prefer a “thorough explanation” |
| Acknowledge that humans are driven by emotion, and that having the facts on your side does not guarantee being able to move the needle on public sentiment | Proceeding as though sharing facts and the art of persuasion are the same thing, arguing that people are logical beings persuaded by facts |
The Solution
This problem can’t be solved with a precaution as simple as prompting the AI to “think like a crisis communications professional”. That may result in marginal improvements on the sources of data it pulls from, but as a whole, there is far less reliable training data available. The internet is rife with case studies of “crisis communications gone wrong” and public outrage when a communications strategy fails. Finding examples of a crisis that was expertly managed, or a credible communications professional willing to go online to discuss their successes in managing crises (most of which are highly confidential), is much more difficult.
Models are constantly improving, but as of this moment there is no magic fix except for a heightened level of human judgment and scrutiny.
Here are a few ways to address this challenge:
- Treat AI as an intellectual sparring partner, rather than an oracle. Push back on the responses it offers, clearly laying bare the consequences of the “Reddit wisdom” the LLM is generating.
- Use AI to help extract the best human judgment and expertise; instead of simply prompting it with questions to respond to, have it ask questions back. Prompt it to look for gaps in the strategy, and then layer in human insight from someone familiar with the crisis to fill those gaps.
- Share AI-supported strategy or outputs with experienced and trusted colleagues. Even if the mandate is to move fast, take time to get a second pair of eyes.
When using AI to support crisis communications work, accountability for the strategy and product always rests with the user. Proceed with robust guardrails, check facts, and treat AI as a tool that enhances, rather than replaces, human expertise.
