Assertion messages: assume failure, or assume success

When testing in any language, how does everybody phrase their assertion messages?

I see three obvious ways:

# assume failure
assert (4-2) == 2, "Subtracting 2 from 4 doesn't equal 2"

# describe success
assert (4-2) == 2, "Subtracting 2 from 4 should equal 2"

# be vauge with failure
assert (4-2) == 2, "Subtracting 2 from 4 is broken"

This is obviously a simple example, but you get the idea. What is the standard practice? What do you do? Why?