Governance & law

Crisis management in the generative-AI era

August 23, 2026 6 min read
Kamel Malek Director of SetRecord.ai

A crisis communications plan usually has a chapter for the press, a chapter for social media, sometimes a chapter for employees. Almost none have a chapter for what ChatGPT, Gemini or Perplexity will tell the next person who asks about the incident — even though, within days, that may be the version of events more people actually read.

A new crisis channel, whether you plan for it or not

During any reputational event — a product recall, a leadership departure, a lawsuit, a data incident — the story unfolds across the usual channels, but it also gets absorbed, compressed and repeated by AI assistants as soon as enough content about it exists online. Unlike a news cycle, this channel doesn't fade in a week. A model's summary of "what happened" can persist, largely unchanged, for as long as the sources it drew from remain uncorrected — long after the news cycle itself has moved on.

That persistence is the core difference crisis teams need to internalise: press coverage ages; a model's standing answer to "what happened at [company]" does not age the same way, and can keep resurfacing an early, incomplete or inaccurate version of events to every new person who asks.

The mechanism behind this is straightforward once you see it. Assistants that answer mainly from what they learned during training tend to freeze whatever version of events was most widely published by the time that training data was gathered — including early, incomplete reporting. Assistants that search the web live at the moment of the query behave differently: they can surface same-day coverage, but that also means their answer about an unfolding crisis can shift hour to hour depending on whatever happens to be ranking well at that moment, which is its own kind of instability to manage.

Picture a mid-sized company facing a product recall. The first wave of coverage — incident reports, a hurried initial statement — gets indexed within hours. Two weeks later, once an investigation concludes with a narrower, less alarming root cause, that update has to compete for space against an early narrative that has already been compressed into a confident one-paragraph summary across several assistants. Closing that gap is mostly a question of when the corrected account gets published and how visible it is made, not whether the correction happens to be true.

How model answers amplify — or dampen — a crisis

Three dynamics make AI assistants a distinct force during a crisis, not just another distribution channel:

  • Compression — a nuanced, multi-fact situation gets reduced to a short, confident summary. Nuance that would appear in a full article often doesn't survive that compression.
  • Early-source lock-in — the first wave of coverage, sometimes speculative or incomplete, is what's available when models are queried early. If your official, corrected account arrives later, it competes with an answer that already looks settled.
  • Confidence flattening — hedged, tentative language from an early report ("allegedly", "according to unverified sources") often doesn't survive into a model's summary, which tends to state things more flatly than the coverage it actually drew from.

None of this requires a model to be biased against you. It's simply what compression and early-source lock-in do by default, absent a deliberate, later correction to work against.

The blind spot in most crisis plans

Most crisis playbooks are built around monitoring: media mentions, social sentiment, search results. Very few extend that monitoring to AI assistants, and fewer still have a defined owner for "what do the models say about this right now." That gap means organisations often discover the AI-generated version of their crisis weeks after the fact — usually when a customer, journalist or investor mentions it back to them, at which point the inaccurate version has had time to settle in.

Don't know what AI says about you today?

A SetRecord audit queries supported AI models about your organisation and measures the gap with verifiable facts — within five business days.

Request an audit

Part of the reason is organisational, not technical. Communications owns press monitoring, legal owns regulatory disclosure, and neither team's mandate naturally extends to "what does an AI assistant currently say about this." Without an explicit decision to own that surface, it tends to fall through the gap between two functions that each assume it's someone else's job.

Before a crisis: map your baseline exposure

The best time to learn what AI assistants say about your organisation is before you're in a crisis, not during one. A baseline audit — the ten to fifteen questions a journalist or customer would ask about your organisation in a difficult moment — gives you a reference point. When a real event happens, you'll immediately know whether the model's account has shifted, and by how much, instead of trying to establish a baseline under pressure.

It's worth going one step further and folding this into existing crisis simulations. When you run a tabletop exercise for a hypothetical incident, add a round where the team has to draft the statement it would want models to eventually reflect, not just the press release. Teams that haven't done this before are usually surprised by how differently they write for an audience that compresses everything down to two sentences.

During a crisis: what to actually monitor

Once an incident is live, add AI assistants to the same monitoring cadence as press and social channels:

  • Query daily, not weekly — model answers about an active story can shift as fast as the underlying coverage does.
  • Track the same five signals used for routine monitoring — presence, accuracy, sentiment, source diversity, volatility — but at higher frequency for the duration of the event.
  • Flag the specific factual errors, not just the general tone; a correction has to target a claim, not a mood.
  • Publish your official account promptly and make it easy to find and cite — a clear, dated statement gives future model answers something authoritative to draw on.
  • Keep one canonical statement — point every channel, internal and external, to the same dated, factual account; a fragmented set of internal talking points that never gets published gives a model nothing authoritative to draw on later.

This is a short, intense addition to an already stretched crisis team's workload — which is exactly why the ownership question addressed above needs to be settled before a crisis starts, not during one.

After a crisis: verify persistence, don't assume resolution

The instinct once a crisis is resolved in the press is to stop watching. That's exactly when an inaccurate AI-generated summary tends to calcify, because query volume drops and no one is checking anymore. A disciplined close-out means re-querying the same baseline questions thirty, sixty and ninety days after the event to confirm the model's account actually caught up with the facts — and correcting again if it hasn't. This closing loop is what separates a crisis response that's genuinely resolved from one that's simply no longer being watched.

It helps to include the AI-narrative check in the same post-mortem meeting where you already review press coverage and internal response times. Asking "did the model's account eventually match the facts, and how long did that take" turns an ad hoc afterthought into a metric you can actually improve across successive incidents.

See what AI says about you

A SetRecord audit checks what supported AI models say against your verifiable facts — within five business days.