diff --git a/include/boost/math/tools/precision.hpp b/include/boost/math/tools/precision.hpp
index f0a4c80a0c..1bc35d9ae1 100644
--- a/include/boost/math/tools/precision.hpp
+++ b/include/boost/math/tools/precision.hpp
@@ -145,7 +145,7 @@ inline constexpr T epsilon(const std::true_type& BOOST_MATH_APPEND_EXPLICIT_TEMP
 
 #if defined(__GNUC__) && ((LDBL_MANT_DIG == 106) || (__LDBL_MANT_DIG__ == 106))
 template <>
-inline constexpr long double epsilon<long double>(const std::true_type& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(long double)) noexcept(long double)
+inline constexpr long double epsilon<long double>(const std::true_type& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(long double)) noexcept(std::is_floating_point<long double>::value)
 {
    // numeric_limits on Darwin (and elsewhere) tells lies here:
    // the issue is that long double on a few platforms is