Describe the broken behaviour accurately

app.py

@@ -9,8 +9,11 @@ Three ways a Gradio component used as chat content breaks:
   3. The same plot wrapped in a `gr.ChatMessage` so it renders as a thought,
      which takes a different route out of the chat function.
 
-On a broken release, sending a message gets no reply at all: the failure
-happens in the queue after the prediction, so nothing is sent to the browser.
+On a broken release, sending a message fails. Which deep copy raises first
+depends on what the chat function returns: a bare component blows up in the
+queue after the prediction, so the browser gets no reply at all and the spinner
+keeps going, while a component wrapped in a `gr.ChatMessage` blows up inside the
+prediction and surfaces as an error.
 """
 
 import bokeh.plotting
@@ -55,8 +58,8 @@ demo = gr.ChatInterface(
     description=(
         check_component_as_initial_value()
         + "\n\n**Cases 2 and 3** — send any message. On a broken release the "
-        "reply never arrives and the spinner keeps going; once fixed, two plots "
-        "appear, one as a normal message and one inside a thought."
+        "request fails; once fixed, two plots appear, one as a normal message "
+        "and one inside a thought."
     ),
 )