The Crisis Story AI Tells Doesn't End When Yours Does

2026-07-27

Ask any of the leading AI systems what the best example of crisis recovery looks like, and nearly all of them will give you the same answer first: Johnson & Johnson, 1982, the Tylenol recall. The case is 44 years old. It still holds the highest citation share of any corporate crisis in the answers these systems generate — ahead of Boeing, ahead of Enron, ahead of every scandal born in the smartphone era.

That finding comes from research published in June 2026 modeling how ChatGPT, Claude, Perplexity, Gemini and Google AI Overviews answer questions like "biggest corporate crisis of all time" or "what happened with [company]." Across 25 companies and 62 buyer-intent prompts, the researchers found something that should unsettle anyone who thinks a crisis is a problem you solve and move past: citation share, once established, barely decays. Enron is 25 years old and still ranks third. BP's Deepwater Horizon spill is 16 years old and still ranks fourth. Wells Fargo's account-fraud scandal, only a decade old, already sits in the top five. These are not stories AI systems are forgetting. They are references AI systems keep reaching for.

The more useful part of the research isn't the ranking — it's the split. The same study found that AI systems treat two kinds of crisis history differently. Recovery-framed cases — J&J, Toyota's 2009 recall response, Chipotle's food-safety turnaround, Domino's public apology — compound as positive reference points. Cover-up-framed cases — Volkswagen's Dieselgate, Wells Fargo, Theranos — compound as warnings. The systems aren't guessing at this distinction. They're reading it out of the corpus: congressional testimony, court filings, business-school case studies, sustained press coverage that documented, in each instance, what the company actually said and did while the crisis was live.

That's the part worth sitting with. The framing that calcifies for decades isn't determined by the severity of the original incident. It's determined by the quality and completeness of the record left behind during the response. Toyota's congressional testimony became a structural-recovery reference. Volkswagen's internal communications, introduced as evidence years after Dieselgate broke, became the cover-up frame that still anchors its citation share today. Two companies, two crises, two records — and AI systems now treat one as the textbook and the other as the warning, for as long as they keep answering questions about corporate reputation.

For anyone managing a crisis today, this changes what the crisis actually is. It was always a live-fire test of judgment under pressure. It is now also, whether anyone intended it or not, the raw material for how several AI systems will describe the company for the next several decades, to every stakeholder who never lived through the original event and has no other way of finding out what happened. The audience for this week's crisis statement is not just this week's journalists and customers. It's the version of Perplexity someone queries in 2041.

Which raises the uncomfortable practical question underneath the research: who, at most companies, is actually keeping the record straight while the crisis is live? Court records and congressional testimony get created by processes designed to produce them. What an AI system says about a company on the Tuesday of a crisis — which facts it stated, which sources it leaned on, whether the framing shifted after the Thursday statement — mostly doesn't get captured by anyone. It is generated once, seen by whoever happened to ask, and then it's gone, even as it may already be feeding the corpus that determines whether this becomes a recovery case or a cover-up case.

Crisis teams have spent decades building muscle around managing the message to reporters and the public. The research suggests the bigger, quieter stakes now sit somewhere most of these teams aren't looking: the sealed, timestamped record of what the machines were actually saying, and citing, while the story was still being written.