Engagement Should Be the Tool, Not the Goal

Media has always competed for attention. Humour, rhythm, suspense, colour and repetition can make an experience memorable. Engagement itself is not the problem. The design question is what engagement is for.

A signal is not an outcome

Watch time, clicks, likes and return visits are observable. Learning, informed citizenship, autonomy and wellbeing are more difficult to measure. This creates a temptation to optimise the signal that is readily available and treat it as a sufficient account of value.

But a person can keep watching because a sequence is excellent, because stopping requires effort, because the next claim is alarming or because the system has learned how to sustain curiosity without satisfying it. The same behavioural measure can represent very different experiences.

Purpose before optimisation

Educational media offers a useful analogy. It can begin with a defined aim, use creative craft to earn attention and evaluate whether the intended audience understood or learned. Attention does not guarantee learning, and an educational label is not proof of quality. The important feature is that engagement serves an independently stated purpose.

Recommendation can work in the same order:

1. Define whose interests the system should serve.
2. Identify several outcomes that represent those interests.
3. Use engagement as one source of evidence rather than the final objective.
4. test for predictable failures and groups who experience them differently.
5. Give people meaningful control over the inferred model and the resulting sequence.

Public value is plural

No single developmental or civic score can settle the problem. Learning, entertainment, rest, creativity, connection and autonomy can conflict. Different people will reasonably choose different balances.

The answer is not for government to approve every joke or require each recommendation to teach a lesson. It is to require services to state important objectives, explain their proxies, test cumulative effects and make consequential trade offs governable.

Better measures change the product

A children’s service could measure whether a session contains a natural stopping point. A news recommender could evaluate source diversity and distinguish repeated claims from independent corroboration. A video platform could test whether users can broaden or reset an inferred interest without losing everything useful in their history.

These measures do not eliminate engagement. They stop it from standing in for every other value.

The question for a product team is therefore not simply, did the person continue? It is what did continuing enable, and would the person recognise that as serving their interests?

Further reading

• Adomavicius and colleagues: Recommender systems, ground truth and preference pollution. Examines how recommendation and consumption can affect later preference data.