diff --git a/compiler/rustc_infer/src/infer/relate/generalize.rs b/compiler/rustc_infer/src/infer/relate/generalize.rs
index 74929daffe247..91e680cc6f472 100644
--- a/compiler/rustc_infer/src/infer/relate/generalize.rs
+++ b/compiler/rustc_infer/src/infer/relate/generalize.rs
@@ -492,9 +492,26 @@ impl<'tcx> TypeRelation<'tcx> for Generalizer<'_, 'tcx> {
                             let origin = inner.type_variables().var_origin(vid);
                             let new_var_id =
                                 inner.type_variables().new_var(self.for_universe, origin);
-                            let u = Ty::new_var(self.tcx(), new_var_id);
-                            debug!("replacing original vid={:?} with new={:?}", vid, u);
-                            Ok(u)
+                            // If we're in the new solver and create a new inference
+                            // variable inside of an alias we eagerly constrain that
+                            // inference variable to prevent unexpected ambiguity errors.
+                            //
+                            // This is incomplete as it pulls down the universe of the
+                            // original inference variable, even though the alias could
+                            // normalize to a type which does not refer to that type at
+                            // all. I don't expect this to cause unexpected errors in
+                            // practice.
+                            //
+                            // cc trait-system-refactor-initiative#108
+                            if self.infcx.next_trait_solver()
+                                && !self.infcx.intercrate
+                                && self.in_alias
+                            {
+                                inner.type_variables().equate(vid, new_var_id);
+                            }
+
+                            debug!("replacing original vid={:?} with new={:?}", vid, new_var_id);
+                            Ok(Ty::new_var(self.tcx(), new_var_id))
                         }
                     }
                 }
@@ -614,6 +631,15 @@ impl<'tcx> TypeRelation<'tcx> for Generalizer<'_, 'tcx> {
                                     universe: self.for_universe,
                                 })
                                 .vid;
+
+                            // See the comment for type inference variables
+                            // for more details.
+                            if self.infcx.next_trait_solver()
+                                && !self.infcx.intercrate
+                                && self.in_alias
+                            {
+                                variable_table.union(vid, new_var_id)
+                            }
                             Ok(ty::Const::new_var(self.tcx(), new_var_id, c.ty()))
                         }
                     }
diff --git a/tests/ui/traits/next-solver/generalize/hr-alias-universe-lowering-ambiguity.rs b/tests/ui/traits/next-solver/generalize/hr-alias-universe-lowering-ambiguity.rs
new file mode 100644
index 0000000000000..1673b196dac67
--- /dev/null
+++ b/tests/ui/traits/next-solver/generalize/hr-alias-universe-lowering-ambiguity.rs
@@ -0,0 +1,51 @@
+//@ compile-flags: -Znext-solver
+//@ check-pass
+
+// A regression test for a fairly subtle issue with how we
+// generalize aliases referencing higher-ranked regions
+// which previously caused unexpected ambiguity errors.
+//
+// The explanations in the test may end up being out of date
+// in the future as we may refine our approach to generalization
+// going forward.
+//
+// cc trait-system-refactor-initiative#108
+trait Trait<'a> {
+    type Assoc;
+}
+
+impl<'a> Trait<'a> for () {
+    type Assoc = ();
+}
+
+fn foo<T: for<'a> Trait<'a>>(x: T) -> for<'a> fn(<T as Trait<'a>>::Assoc) {
+    |_| ()
+}
+
+fn unconstrained<T>() -> T {
+    todo!()
+}
+
+fn main() {
+    // create `?x.0` in the root universe
+    let mut x = unconstrained();
+
+    // bump the current universe of the inference context
+    let bump: for<'a, 'b> fn(&'a (), &'b ()) = |_, _| ();
+    let _: for<'a> fn(&'a (), &'a ()) = bump;
+    
+    // create `?y.1` in a higher universe
+    let mut y = Default::default();
+
+    // relate `?x.0` with `for<'a> fn(<?y.1 as Trait<'a>>::Assoc)`
+    // -> instantiate `?x.0` with `for<'a> fn(<?y_new.0 as Trait<'a>>::Assoc)`
+    x = foo(y);
+
+    // Constrain `?y.1` to `()`
+    let _: () = y;
+    
+    // `AliasRelate(<?y_new.0 as Trait<'a>>::Assoc, <() as Trait<'a>>::Assoc)`
+    // remains ambiguous unless we somehow constrain `?y_new.0` during
+    // generalization to be equal to `?y.1`, which is exactly what we
+    // did to fix this issue.
+}