diff --git a/theories/base_logic/upred.v b/theories/base_logic/upred.v
index 44a5aaea896733e684c0abb9aff30f9822163995..a67341c7d97e6c4dff1354aeeafb8226592eaa8f 100644
--- a/theories/base_logic/upred.v
+++ b/theories/base_logic/upred.v
@@ -11,7 +11,7 @@ Local Hint Extern 10 (_ ≤ _) => lia : core.
     base_logic.base_logic; that will also give you all the primitive
     and many derived laws for the logic. *)
 
-(* A good way of understanding this definition of the uPred OFE is to
+(** A good way of understanding this definition of the uPred OFE is to
    consider the OFE uPred0 of monotonous SProp predicates. That is,
    uPred0 is the OFE of non-expansive functions from M to SProp that
    are monotonous with respect to CMRA inclusion. This notion of
@@ -46,6 +46,55 @@ Local Hint Extern 10 (_ ≤ _) => lia : core.
    connective.
  *)
 
+(** Note that, somewhat curiously, [uPred M] is *not* in general a Camera,
+   at least not if all propositions are considered "valid" Camera elements.
+   It fails to satisfy the extension axiom. Here is the counterexample:
+
+We use M := (option Ex {A,B})^2 -- so we have pairs
+whose components are ε, A or B.
+
+Let
+
+  P r n := r = (A,A) ∧ n = 0 ∨
+           r = (A,B) ∨
+           r = (B,A) ∨
+           r = (B,B)
+ Q1 r n := (A, ε) ≼ r ∨ (B, ε) ≼ r
+           ("Left component is not ε")
+ Q2 r n := (ε, A) ≼ r ∨ (ε, B) ≼ r
+           ("Right component is not ε")
+
+These are all sufficiently closed and non-expansive and whatnot.
+We have P ≡{0}≡ Q1 * Q2. So assume extension holds, then we get Q1', Q2'
+such that
+
+  P ≡ Q1' ∗ Q2'
+ Q1 ≡{0}≡ Q1'
+ Q2 ≡{0}≡ Q2'
+
+Now comes the contradiction:
+We know that P (A,A) 1 does *not* hold, but I am going to show that
+(Q1' ∗ Q2') (A,A) 1 holds, completing the proof.
+To this end, I will show (a) Q1' (A,ε) 1 and (b) Q2' (ε,A) 1.
+The result follows from (A,ε) ⋅ (ε,A) = (A,A)
+(a) We have P (A,B) 1, and thus Q1' r1 1 and Q2' r2 1 for some
+    r1 â‹… r2 = (A,B). There are four possible decompositions:
+    - (ε,ε) ⋅ (A,B): This would give us Q1' (ε,ε) 1, from which we
+      obtain (through down-close and the equality above) that
+      Q1 (ε,ε) 0. However, we know that's false.
+    - (A,B) ⋅ (ε,ε): Can be excluded for similar reasons
+      (the second resource must not be ε in the 2nd component).
+    - (ε,B) ⋅ (A,ε): Can be excluded for similar reasons
+      (the first resource must not be ε in the 1st component).
+    - (A,ε) ⋅ (ε,B): This gives us the desired Q1' (A,ε) 1
+(b) We have P (B,A) 1, and thus Q1' r1 1 and Q2' r2 1 for some
+    r1 â‹… r2 = (B,A). There are again four possible decompositions,
+    and like above we can exclude three of them. This leaves us with
+    (B,ε) ⋅ (ε,A) and thus Q2' (ε,A) 1.
+This completes the proof.
+
+*)
+
 Record uPred (M : ucmraT) : Type := UPred {
   uPred_holds :> nat → M → Prop;