"This Is Not a Test": Cornell's 15-Minute Tornado Warning Window on the Thursday Before July 4th

To check this alert, Claude (an AI) read it in full 25 separate times, independently. Each read decided whether the message answers each of the five questions and gave a short reason. A final reviewer then weighed all 25 and wrote the plain-English verdict you see when you open a row. The score (for example 22/25) is how many reads agreed; the 25 individual reads are tucked underneath if you want to check them.

• Sourceabsent0/0

  Who is sending the alert and who is responding. People act faster on a message from a clearly identifiable, credible sender, such as a named department, the police, or a branded alert system, than on an anonymous notice. A branded signature counts.

  See all 25 individual reads

  Open to load the 25 reads.

• Hazardabsent0/0

  What the threat actually is. A complete warning names the specific danger, such as a shooter, a fire, a tornado, or a gas leak, rather than a vague emergency, because people decide what to do based on what they are facing.

  See all 25 individual reads

  Open to load the 25 reads.

• Locationabsent0/0

  Where the threat is. Saying whether danger is in a specific building, a part of campus, or area-wide lets people judge their own proximity and choose a safe direction. Without a where, a warning is hard to act on precisely.

  See all 25 individual reads

  Open to load the 25 reads.

• Guidanceabsent0/0

  The protective action to take. A clear, specific instruction, such as shelter in place, evacuate, avoid the area, or run-hide-fight, drives faster and more correct protective behavior than describing the threat alone.

  See all 25 individual reads

  Open to load the 25 reads.

• Timeabsent0/0

  When the message applies. A timestamp, the word now or immediately, or a phrase like until further notice tells the reader whether the danger is current and how quickly to act.

  See all 25 individual reads

  Open to load the 25 reads.