diff --git a/plans/changes_sync.go b/plans/changes_sync.go
index e9305eaf9..6b4ff98ff 100644
--- a/plans/changes_sync.go
+++ b/plans/changes_sync.go
@@ -133,8 +133,8 @@ func (cs *ChangesSync) RemoveOutputChange(addr addrs.AbsOutputValue) {
 	defer cs.lock.Unlock()
 
 	addrStr := addr.String()
-	for i, r := range cs.changes.Resources {
-		if r.Addr.String() != addrStr {
+	for i, o := range cs.changes.Outputs {
+		if o.Addr.String() != addrStr {
 			continue
 		}
 		copy(cs.changes.Outputs[i:], cs.changes.Outputs[i+1:])
diff --git a/terraform/eval_output.go b/terraform/eval_output.go
index 6829934f0..10573971f 100644
--- a/terraform/eval_output.go
+++ b/terraform/eval_output.go
@@ -9,6 +9,7 @@ import (
 
 	"github.com/hashicorp/terraform/addrs"
 	"github.com/hashicorp/terraform/plans"
+	"github.com/hashicorp/terraform/states"
 )
 
 // EvalDeleteOutput is an EvalNode implementation that deletes an output
@@ -61,22 +62,33 @@ func (n *EvalWriteOutput) Eval(ctx EvalContext) (interface{}, error) {
 			// if we're continuing, make sure the output is included, and
 			// marked as unknown. If the evaluator was able to find a type
 			// for the value in spite of the error then we'll use it.
-			state.SetOutputValue(addr, cty.UnknownVal(val.Type()), n.Sensitive)
+			n.setValue(addr, state, changes, cty.UnknownVal(val.Type()))
 			return nil, EvalEarlyExitError{}
 		}
 		return nil, diags.Err()
 	}
 
+	n.setValue(addr, state, changes, val)
+
+	return nil, nil
+}
+
+func (n *EvalWriteOutput) setValue(addr addrs.AbsOutputValue, state *states.SyncState, changes *plans.ChangesSync, val cty.Value) {
 	if val.IsKnown() && !val.IsNull() {
 		// The state itself doesn't represent unknown values, so we null them
 		// out here and then we'll save the real unknown value in the planned
 		// changeset below, if we have one on this graph walk.
+		log.Printf("[TRACE] EvalWriteOutput: Saving value for %s in state", addr)
 		stateVal := cty.UnknownAsNull(val)
 		state.SetOutputValue(addr, stateVal, n.Sensitive)
 	} else {
+		log.Printf("[TRACE] EvalWriteOutput: Removing %s from state (it is now null)", addr)
 		state.RemoveOutputValue(addr)
 	}
 
+	// If we also have an active changeset then we'll replicate the value in
+	// there. This is used in preference to the state where present, since it
+	// *is* able to represent unknowns, while the state cannot.
 	if changes != nil {
 		// For the moment we are not properly tracking changes to output
 		// values, and just marking them always as "Create" or "Destroy"
@@ -116,8 +128,8 @@ func (n *EvalWriteOutput) Eval(ctx EvalContext) (interface{}, error) {
 			// Should never happen, since we just constructed this right above
 			panic(fmt.Sprintf("planned change for %s could not be encoded: %s", addr, err))
 		}
-		changes.AppendOutputChange(cs)
+		log.Printf("[TRACE] EvalWriteOutput: Saving %s change for %s in changeset", change.Action, addr)
+		changes.RemoveOutputChange(addr) // remove any existing planned change, if present
+		changes.AppendOutputChange(cs)   // add the new planned change
 	}
-
-	return nil, nil
 }