Feedback that works: how to give learning a direction

Why feedback steers learning

Without feedback you only know that you did something, not whether it was good. Practising without feedback can even cement mistakes. Research in education keeps showing that feedback is one of the most powerful influences on learning, but also that the effect varies strongly: some feedback helps, other feedback does nothing or makes things worse. So it is not about more feedback, but better.

Five traits of feedback that helps

1. Concrete. "You asked a closed question three times" helps; "well done" does not.
2. Fast. The shorter after the action, the better it sticks to that action. For skills, within minutes beats within weeks.
3. About the task, not the person. "This explanation skipped a step" is more useful than "you are sloppy".
4. Applicable. Say what you can do differently next time, not only what went wrong.
5. With a retry. Feedback without a second attempt is a verdict. With a second attempt it is learning.

What AI feedback can do in a learning journey

In CoVince open answers are assessed by AI, the learner gets feedback in the lesson and after a conversation with a CrewMate the AI writes a final report. The big advantage is speed and availability: the feedback comes straight away, for every participant, at any moment. That is something a trainer with twenty participants can never do.

The feedback after a second attempt is different too: with a question you can try again and get fewer points, so the retry is built in.

Pitfalls of AI feedback

• Too friendly. An AI that praises everything teaches nothing. Make the criteria visible and set the assessment strict enough.
• Too general. If the feedback could be the same for every answer, it is worthless. Give the AI the criteria.
• Without a human. The AI does not see everything. Have a trainer or colleague read along and steer regularly.

The final report as conversation material

A report is only valuable when someone does something with it. After a conversation exercise, let learners choose one point they want to continue with and try it in the next round. A trainer or manager can take the report as a starting point in a ten-minute conversation. That way the AI feedback becomes the start of a conversation, not the end of it. You can read more about the practice itself in practising with an AI role-play.

Feedback from each other

Not all feedback has to come from the AI or the trainer. In a scenario game the debrief shows who knew what when, and so gives feedback from the game about the team. In the organisation colleagues can give each other an award, with a required reason. That makes recognition a form of feedback too: it says what someone did well. See rewards without the gamification pitfalls.

An example: from "well done" to something useful

Compare two forms of feedback on an explanation to a customer. The first: "Well done, clear story." Pleasant, but the learner does not know what to repeat or adjust. The second: "You started with the result for the customer and used one example. You skipped the step of how it gets paid, and that is where the customer’s question came from. Try that step after the example next time."

The second is concrete, is about the task, explains a cause and gives a next step. You can write feedback like that for open questions, so the AI gives it that way too: record the assessment criteria, and ask for an observation, a consequence and a suggestion.

A culture where feedback is normal

Technology helps, but feedback only really works when people want it. That needs an environment where making mistakes in an exercise is normal, and where managers ask for and receive feedback themselves. A learning journey can help: a chat per learning journey where questions are asked, awards with which colleagues thank each other with a reason, a debrief after a scenario game that is about facts and not blame.

Start with yourself: as a trainer or designer, ask for feedback on your own learning journey. An open question at the end ("what was unclear?") yields a lot and shows you mean it.

Setting criteria for AI assessment

The quality of AI feedback depends on the criteria you provide. A good set consists of three to five criteria, each in plain language and observable. For a customer conversation, for example: did the learner ask follow-up questions before giving a solution, did they name the customer’s feeling, did they summarise before closing, and did they say concretely what the next step is?

• Keep criteria concrete: what can be seen or heard?
• Describe what is good and what is insufficient, with an example.
• Test the criteria with a few real answers and see whether the AI assesses as you would.
• Adjust based on what you see, and share the criteria with the learners.

How to set it up

Per learning goal, start with the question: what feedback does the learner need to get better, and when? Put the quick feedback (right/wrong with explanation) in the swipeslides, the rich feedback (report) with the conversations, and the human feedback at the moments that matter. Check in the reports whether people really do a second round after feedback. If not, the feedback may be good but the step towards it is not.