The author and Fullintel test the ability of an artificial intelligence (AI) system to identify crises and recommend responses. Results show very high accuracy and effective response recommendation.
In a 2019 presentation to the International Public Relations Research Conference (IPRRC), the guru of crisis response Professor W. Timothy Coombs and his co-author Elina R. Tachkova argued that the “preventable crisis,” (i.e., those that an organization brings on itself, caused by human error or management misconduct) requires a very different response than for other types of crises. They suggested that while there is lots of research for other types of crises, there is almost none which identifies the most effective response in a preventable crisis.
To me, that was a crisis in and of itself, given how many preventable crises dominate the headlines these days.
Six months later, at our annual Summit on the Future of Measurement, I asked the traditional closing question: “What do you see as the future of measurement?” About half the respondents mentioned artificial intelligence (AI). When it came around to my answer, I suggested that AI would only become a truly valuable tool when it could identify the type of crisis (preventable vs. accidental vs. victim, etc.) and recommend the most effective response.
One of the sponsors, Gaugarin Oliver, CEO of Fullintel, piped up with an answer like the old FedEx commercial: “We can do that!” So a couple of weeks later we chatted with Professor Coombs to map out a research project that would test the theory. We recently presented the results of our test at IPRRC 2020.
Testing AI’s ability to identify crises and recommend responses
Fullintel collected many thousands of articles of traditional media coverage on three recent human error and management misconduct crises: the Boeing 737-Max, the demise of WeWork, and an accidental toddler death. We started with definitions of different types of response from Professor Coombs’ Situational Crisis Communications Theory (SCCT). Fullintel taught their machine learning technology how to identify a crisis, how to classify the type of crisis, and how to identify different responses.
Here are the response types, also based on Coombs’ theory:
|Deny crisis response strategies|
|Attack the accuser: Crisis manager confronts the person or group claiming something is wrong with the organization.|
|Denial: Crisis manager asserts that there is no crisis.|
|Scapegoat: Crisis manager blames some person or group outside of the organization for the crisis.|
|Diminish crisis response strategies|
|Excuse: Crisis manager minimizes organizational responsibility by denying intent to do harm and/or claiming inability to control the events that triggered the crisis.|
|Justification: Crisis manager minimizes the perceived damage caused by the crisis.|
|Rebuild crisis response strategies|
|Compensation: Crisis manager offers money or other gifts to victims.|
|Apology: Crisis manager indicates the organization takes full responsibility for the crisis and asks stakeholders for forgiveness.|
|Secondary crisis response strategies|
|Bolstering crisis response strategies|
|Reminder: Tell stakeholders about the past good works of the organization.|
|Ingratiation: Crisis manager praises stakeholders and/or reminds them of past good works by the organization.|
|Victimage: Crisis managers remind stakeholders that the organization is a victim of the crisis too.|
How accurate is AI crisis identification?
The next step was to check the accuracy of the AI-programmed data. The results were so good they surprised us all. In terms of whether a media narrative was or was not a crisis, we found that humans and AI agreed 97% of the time. We found similar results (97.4%) when we asked the AI system to correctly identify the type of crisis. Agreement with crisis response was slightly lower (94%), mostly due to confusion between “No response” and “No comment.”
The next challenge was to see if AI could predict the best response type. We defined “best” as that which produced the fewest negative articles and had the shortest time between crisis peak and neutral coverage. The machine data confirmed what most of us know in our gut: denial, attacking the accuser, and scapegoating are the least effective response types in a self-inflicted crisis. Providing information was the most effective response.
Implications: How will we use AI for crisis response?
The first question most people ask is: “Will AI replace the PR person?” The answer is a definitive “No.” AI can’t replace years of experience or good gut instinct.
What AI can do is provide PR pros with an effective tool to communicate with recalcitrant leadership: numbers and charts. Here’s an example from our work. If faced with this chart, any CEO will find it obvious what the best responses would be:
Also, AI is far more effective than most of today’s crisis identification systems, that look at volume of social media or at negative coverage increases.
Our work demonstrates that the potential to use artificial intelligence to help one manage a crisis is not some distant pipe dream, but something that you might want to put into next year’s budget. ∞