diff --git a/changelog.d/px-2ms.md b/changelog.d/px-2ms.md
new file mode 100644
index 0000000..8f24e17
--- /dev/null
+++ b/changelog.d/px-2ms.md
@@ -0,0 +1,4 @@
+### Fixed
+
+- A `\U` escape in a string literal is refused with the same error as `\u`:
+  `"caf\U00e9"` no longer lexes silently to `cafU00e9`.
diff --git a/docs/reference/language.md b/docs/reference/language.md
index 6dd24e6..36d2919 100644
--- a/docs/reference/language.md
+++ b/docs/reference/language.md
@@ -45,8 +45,8 @@ recognized:
 
 All six are recognized in both quote styles - the quote character that does
 not delimit the literal needs no escape, but escaping it is still accepted.
-Any other escaped character stands for itself and the backslash is dropped,
-so `"a\qb"` is `aqb`.
+Any other escaped character, apart from the refused `\u` and `\U` below,
+stands for itself and the backslash is dropped, so `"a\qb"` is `aqb`.
 
 ```elixir
 iex> Predicator.evaluate(~S{"say \"hi\""}, %{})
@@ -60,12 +60,18 @@ iex> Predicator.evaluate(~S{"a\qb"}, %{})
 ```
 
 There is no numeric escape. A `\u` sequence is refused at parse time with an
-error naming it, rather than decoding to the bare letter:
+error naming it, rather than decoding to the bare letter. `\U` is refused the
+same way, with the same error: both spellings of the letter are the refused
+escape, so neither decodes silently.
 
 ```elixir
 iex> {:error, err} = Predicator.compile("\"caf\\u00e9\"")
 iex> err.message
 "Unsupported escape sequence \\u in string literal: predicator has no numeric escape; write the character itself (string literals are UTF-8)"
+
+iex> {:error, upper} = Predicator.compile("\"caf\\U00e9\"")
+iex> upper.message
+"Unsupported escape sequence \\u in string literal: predicator has no numeric escape; write the character itself (string literals are UTF-8)"
 ```
 
 String literals are UTF-8 source text, so the character is written directly -
diff --git a/lib/predicator/lexer.ex b/lib/predicator/lexer.ex
index 0011a55..8467a4d 100644
--- a/lib/predicator/lexer.ex
+++ b/lib/predicator/lexer.ex
@@ -694,11 +694,14 @@ defmodule Predicator.Lexer do
     {:ok, acc, rest, count, line, col + 1}
   end
 
-  # `\uXXXX` is refused rather than decoded. A string literal is already UTF-8
-  # source text, so the character can be written directly; a bug fix has no
-  # mandate to add grammar. The refusal is explicit so the escape never
-  # silently decodes to the bare letter instead.
-  defp take_string([?\\ | [?u | _rest]], _acc, _count, _quote_type, _line, _col) do
+  # `\uXXXX` is refused rather than decoded, in either spelling of the letter.
+  # A string literal is already UTF-8 source text, so the character can be
+  # written directly; a bug fix has no mandate to add grammar. The refusal is
+  # explicit so the escape never silently decodes to the bare letter instead,
+  # and `\U` is refused with `\u` because two spellings of one escape refused
+  # differently is the defect.
+  defp take_string([?\\ | [u | _rest]], _acc, _count, _quote_type, _line, _col)
+       when u in [?u, ?U] do
     {:error,
      "Unsupported escape sequence \\u in string literal: predicator has no " <>
        "numeric escape; write the character itself (string literals are UTF-8)"}
diff --git a/test/predicator/lexer_test.exs b/test/predicator/lexer_test.exs
index acb624b..f06092d 100644
--- a/test/predicator/lexer_test.exs
+++ b/test/predicator/lexer_test.exs
@@ -299,6 +299,19 @@ defmodule Predicator.LexerTest do
       assert message =~ "\\u"
       assert message =~ "string literal"
     end
+
+    # Sabotage: drop the `?U` from the `when u in [?u, ?U]` guard on the
+    # refusal clause in `take_string/6` and this goes red - `\\U0041` falls
+    # through to the generic escape clause and lexes as "U0041".
+    test "refuses an uppercase \\U escape with the same error as \\u" do
+      assert {:error, message, 1, 1, _span} = Lexer.tokenize(~s("\\U0041"))
+
+      assert message =~ "\\u"
+      assert message =~ "string literal"
+
+      assert {:error, lower_message, 1, 1, _span} = Lexer.tokenize(~s("\\u0041"))
+      assert message == lower_message
+    end
   end
 
   describe "additional edge cases for coverage" do