Group Chat AI Summaries

01 —The Problem

The cost of catching up

Group chats on Messenger are some of the most active, most beloved spaces, but they're also overwhelming. A busy group chat can accumulate messages while you're away, and the prospect of reading through all of them creates friction. Introducing AI summaries for busy group chats that can help people get quick context in order to reply. I came up with 3 core principles:

Design principles

02 —Constraints & XFN

Designing in new territory

Unlike most features I'd worked on, there was almost no direct competitive precedent to learn from. That meant the strategic and technical constraints had to be discovered through XFN collaboration rather than inferred from the market.

Each constraint had direct design implications. Working through them with engineering, privacy, and policy partners early was what kept the project moving without major redirects later.

Key design decisions

Defining the feature

I landed on the unread messages divider as the entry point for two compounding reasons. First, it's where the unread conversation begins, which is precisely where a summary is most useful. Second, anchoring the entry point there means users who want to read the chat themselves can immediately do so. The summary is offered at the threshold, not imposed on top of the content.

The summary itself opens as a collapsed card that users tap to expand. Collapsed by default keeps it non-intrusive. The tap-to-expand interaction makes the choice to read the summary an active, intentional one — which also signals to the system that the user found the entry point relevant.

Entry point

Bottom sheet

AI-generated summaries can be genuinely dense. Getting the visual treatment right required significant iteration.

The core challenge: a summary that looks like a wall of text defeats its own purpose. If the summary feels harder to read than the chat, users won't use it. I went through many rounds of design iteration on hierarchy, line length, type size, and spacing to find a format that felt light and scannable even when the underlying content was substantive.

The best way to improve quality is asking the people who have used it. Having a feedback system could signal the team whether the summary was accurate and useful and act upon it.

Feedback mechanism — thumbs up / thumbs down, subtle but accessible within the card

What shipped

Live on Messenger

Group Chat AI Summaries shipped on Messenger as part of Meta's broader MetaAI integration across its messaging products. The feature is publicly documented in Messenger's Help Center.

07 —Reflections

What I learned

My biggest lesson for this project was working on an evolving design system. Messenger has a very robust system, which had helped have a starting point in most designs I'd worked on in the past. MetaAI was still defining its branding, so working in early stages of MetaAI on Messenger was very insightful. I also learned about collaborating with XFN partners: engineering, privacy, policy, Trust & Safety. AI features carry constraints that aren't visible in the design file. It was cool to learn so much from people with different responsibilities as me. Everyone doing what they do best and learning from each other is what made this feature be one of my favorites.

I'm really proud of the contribution this made to the MetaAI design system on Messenger. Establishing patterns for how AI-generated content should be presented helps it be more digestible for users and sets a precedent for whatever is built on top of it next. Can't wait to see where this feature evolves to.