diff --git a/library/core/src/hint.rs b/library/core/src/hint.rs
index e8d724ab1ef..5a76e866923 100644
--- a/library/core/src/hint.rs
+++ b/library/core/src/hint.rs
@@ -219,6 +219,75 @@ pub fn spin_loop() {
 /// backend used. Programs cannot rely on `black_box` for *correctness* in any way.
 ///
 /// [`std::convert::identity`]: crate::convert::identity
+///
+/// # When is this useful?
+///
+/// First and foremost: `black_box` does _not_ guarantee any exact behavior and, in some cases, may
+/// do nothing at all. As such, it **must not be relied upon to control critical program behavior.**
+/// This _immediately_ precludes any direct use of this function for cryptographic or security
+/// purposes.
+///
+/// While not suitable in those mission-critical cases, `back_box`'s functionality can generally be
+/// relied upon for benchmarking, and should be used there. It will try to ensure that the
+/// compiler doesn't optimize away part of the intended test code based on context. For
+/// example:
+///
+/// ```
+/// fn contains(haystack: &[&str], needle: &str) -> bool {
+///     haystack.iter().any(|x| x == &needle)
+/// }
+///
+/// pub fn benchmark() {
+///     let haystack = vec!["abc", "def", "ghi", "jkl", "mno"];
+///     let needle = "ghi";
+///     for _ in 0..10 {
+///         contains(&haystack, needle);
+///     }
+/// }
+/// ```
+///
+/// The compiler could theoretically make optimizations like the following:
+///
+/// - `needle` and `haystack` are always the same, move the call to `contains` outside the loop and
+///   delete the loop
+/// - Inline `contains`
+/// - `needle` and `haystack` have values known at compile time, `contains` is always true. Remove
+///   the call and replace with `true`
+/// - Nothing is done with the result of `contains`: delete this function call entirely
+/// - `benchmark` now has no purpose: delete this function
+///
+/// It is not likely that all of the above happens, but the compiler is definitely able to make some
+/// optimizations that could result in a very inaccurate benchmark. This is where `black_box` comes
+/// in:
+///
+/// ```
+/// use std::hint::black_box;
+///
+/// // Same `contains` function
+/// fn contains(haystack: &[&str], needle: &str) -> bool {
+///     haystack.iter().any(|x| x == &needle)
+/// }
+///
+/// pub fn benchmark() {
+///     let haystack = vec!["abc", "def", "ghi", "jkl", "mno"];
+///     let needle = "ghi";
+///     for _ in 0..10 {
+///         // Adjust our benchmark loop contents
+///         black_box(contains(black_box(&haystack), black_box(needle)));
+///     }
+/// }
+/// ```
+///
+/// This essentially tells the compiler to block optimizations across any calls to `black_box`. So,
+/// it now:
+///
+/// - Treats both arguments to `contains` as unpredictable: the body of `contains` can no longer be
+///   optimized based on argument values
+/// - Treats the call to `contains` and its result as volatile: the body of `benchmark` cannot
+///   optimize this away
+///
+/// This makes our benchmark much more realistic to how the function would be used in situ, where
+/// arguments are usually not known at compile time and the result is used in some way.
 #[inline]
 #[stable(feature = "bench_black_box", since = "1.66.0")]
 #[rustc_const_unstable(feature = "const_black_box", issue = "none")]