臉書 Facebook 推出 AI 搜尋引擎：可從公開貼文、Groups 與 Reels 獲取綜合答案，可信度與隱私爭議起

Meta officially launched the AI Mode search feature on Facebook on Monday, allowing users to extract comprehensive answers from public posts, Groups, and Reels. Simultaneously, new AI photo presets, video editing tools, and a creator AI assistant were added. However, alongside the feature rollout, privacy advocates in the EU have filed a complaint with the Irish DPC.
(Background summary: All about data! Meta will use user FB and IG posts to "train AI." How to oppose Facebook collecting personal data?)
(Additional background: Attention — the photos you take and the things you say are now being stored by Google to train AI (how to turn off training).)

Table of Contents

Toggle

• What Muse Spark can do
• Credibility issues and lessons from Google’s experience
• The closed loop of user data and ad personalization

Meta officially announced earlier that AI Mode is now live on Facebook, enabling users to ask questions using natural language. The Muse Spark model is responsible for extracting answers from public posts, group discussions, and Reels, presenting them as comprehensive summaries rather than traditional link lists.

Facebook has billions of real discussions, local community knowledge, and real-time event contexts. AI Mode is Meta’s first serious attempt to transform this unique asset into a search product, but user backlash has also emerged.

What Muse Spark can do

The technical foundation of AI Mode is Muse Spark, Meta’s most powerful active model released in April 2026, representing a concrete implementation of the Llama series in application.

The newly launched features also include: AI photo presets (changing outfits, hairstyles, accessories), video collage editing (cropping, transitions), virtual try-on for sports jerseys, and an AI assistant for creators that suggests optimal posting times and automatically summarizes comments. These functions span three scenarios: search, personal image, and content creation.

Looking back over the past four months: profile picture updates in February, Marketplace AI auto-replies in March, Muse Spark release in April, Forum app launch (with built-in AI Ask tab) in May, and the official launch of AI Mode in June. Each update tests user acceptance of AI intervention while accumulating real interaction data to feedback into the model.

Credibility issues and lessons from Google’s experience

However, the core issue facing AI Mode is the same as Google AI Mode: how to ensure the credibility of answers generated from ordinary users’ posts?

Google AI Mode has accumulated numerous complaints on Reddit, with users pointing out that generative summaries are less accurate than traditional search results because AI summaries obscure the visibility of information sources, making it hard for users to determine where the answers come from. Facebook faces an even greater challenge: Facebook Groups are filled with personal opinions, forwarded rumors, and commercial promotions, making signal quality more complex.

Meta has not publicly explained how AI Mode handles outdated information, identifies misleading posts, or marks answer uncertainty. But this gap is especially dangerous on sensitive topics: health, finance, and political discussions are rampant in Facebook Groups. If Muse Spark directly summarizes these contents without showing original sources, the risk of spreading misinformation becomes systemic.

The closed loop of user data and ad personalization

In the same week that AI Mode launched, regulatory pressure also intensified. EU privacy advocates filed a complaint with the Irish Data Protection Commission (DPC), claiming Meta’s AI policies violate the GDPR’s “purpose limitation principle,” as users’ consent for data use did not include AI training and ad personalization.

The Electronic Privacy Information Center (EPIC) in the US also sent a letter to the FTC, demanding regulation of Meta’s AI chatbot advertising practices.

This concern has concrete basis: Meta’s updated AI policy in 2026 allows user-AI conversation data to be used for ad personalization without providing an opt-out option. In other words, every question users ask in Facebook’s AI Mode could become input signals for the ad system.

US federal regulators have also issued warnings, pointing out that integrating social data into AI poses risks of identity theft, tracking, and reputation damage. These are not just theoretical concerns—Facebook holds user names, locations, social connections, and purchasing preferences. If the boundaries of AI system access are unclear, the attack surface of this data map expands accordingly.