Does better summary writing actually increase user engagement?

LLM-generated summaries for social network push notifications were objectively more informative and customized than existing templates. They did not improve user engagement. The explanation is structural, not quality-related: a well-summarized notification body contains sufficient information that users do not need to open the notification to understand the content. The optimization target (informativeness) directly undermines the business metric (engagement/clicks).

This is an instance of Goodhart's Law operating through content quality: when you optimize for how informative a message is, you can succeed at informativeness while failing at the behavior the informativeness was supposed to drive. The information was meant to entice users to engage; instead, it satisfied their information need at the notification level.

Two compounding factors emerged from the experiments:

Voice alienation: LLM summarization transformed first-person user voice ("I'm looking for a plumber") into third-person reportage ("neighbor asks about plumbers"). This tonal shift alienated recipients by creating distance from the original social context. The content was more polished but less relational — it sounded like a news brief about a neighbor rather than a neighbor reaching out.

Optimization gap: Without a reward model specifically trained for engagement, or specific model training to tailor user preferences into content generation, in-context learning alone cannot shortcut established templates that have been iteratively refined over years. The control templates were the product of multiple iteration cycles; the LLM-generated alternatives were one-shot productions. Even when LLMs produce "better" content by linguistic quality metrics, they cannot automatically improve engagement metrics that require alignment with user behavioral patterns.

The broader pattern: LLM-generated content is best suited for rapid prototyping of new products but directly using it to improve metrics on mature products that have undergone years of A/B testing often fails. The same dynamic appeared in invitation emails — more informative, more personalized, but not more effective at driving sign-ups. Generic LLM-generated content cannot capture individual personal preferences without further training.

This connects to the alignment tax discussion: since Does preference optimization harm conversational understanding?, we see a parallel where optimizing for one communication quality (informativeness) erodes the behavioral outcome it was meant to serve (engagement). The mechanism differs — RLHF erodes grounding acts while informativeness optimization eliminates click-through motivation — but the pattern is the same: optimizing a proxy metric degrades the downstream target.