Inferring Preferences from User Actions

When an AI application produces a result, the people using it often react to that result in some way. They might click a button, leave a rating, or simply keep using the product. Those reactions are a form of user feedback, and they are one of the main ways a team can learn whether an application is working for the people it serves.

Consider a product recommendation. Suppose someone buys a product that the application recommended to them. The user never filled in a rating and never answered a question about the recommendation. Even so, the purchase suggests the recommendation was a good one. The signal was inferred from what the user did. This article explains how feedback like this is gathered, how it differs from feedback a user gives on purpose, and what teams can do with it.

The source groups feedback into kinds based on how it is gathered. Before looking at each kind, it helps to ask one simple question: when a user gives feedback, did they know they were giving it, or was the information gathered from something they did anyway?

Keep this question in mind as you read the rest of the article. Feedback where the user knowingly answers a prompt is one kind. Feedback where the application observes what the user did is another kind. A third kind arrives through ordinary conversation with an AI. Each kind answers the question differently, and each has different strengths.

The defining feature of explicit feedback is that the application asks for it directly. The user is shown a prompt, and their answer is the feedback. Common forms include thumbs up or thumbs down, upvote or downvote, a star rating, or a yes or no answer to a question such as 'Did we solve your problem?'

Explicit feedback is fairly standard across applications, because there are only so many ways you can ask a person whether they like something. Since the formats repeat from product to product, explicit feedback is better understood than other kinds. Teams know how to design it, and they know how to interpret the answers.

The key idea is that implicit feedback depends on the product. A shopping application has purchases to observe, so a purchase can count as implicit feedback there. Another application may have entirely different actions, and its implicit feedback would be built from those instead. The point is that you cannot list implicit feedback in advance without knowing which actions a user can take inside the application.

The source notes that foundation models enable many new genres of implicit feedback, but it does not list them. What you can rely on is the general principle: implicit feedback can be used to infer user preferences, and the available actions determine what can be observed.

The conversational interface used by many AI applications makes it easier for users to give feedback. Users can encourage good behaviors and correct errors in the same way they would in everyday dialogue. Because the feedback sits inside ordinary speech, the user does not have to be shown a special control to give it.

The language a user uses to give directions to the AI can convey feedback about two things at once: how well the application is performing, and what the user prefers. This is why conversational feedback is a third kind, separate from both explicit and implicit feedback. It is extracted from conversations rather than from a prompt or from a recorded action.

The source uses a travel scenario to show how feedback can be read from a conversation. The example below is an original illustration in the same spirit. Imagine you are using an AI assistant to plan a trip to Australia, and you ask it to find a hotel for three nights in Sydney. It responds with three recommendations.

Once feedback has been collected, the source identifies three broad uses. Evaluation means deriving metrics that show how the application is doing. Development means training future models or guiding how they are developed. Personalization means adapting the application to each user. The same feedback can support more than one of these uses at the same time.

Follow the travel example. The message about the harbour can be counted as one data point when a team measures how often recommendations miss what users asked for, which supports evaluation. The same message can be included in data used to improve future models, which supports development. And the preference for locations near the centre can shape later suggestions for that same user, which supports personalization.

Feedback is classified by how it is gathered. Explicit feedback is requested directly, implicit feedback is inferred from actions, and conversational feedback is extracted from conversations. Each kind can be used for evaluation, development, and personalization.