Taiwan AI Labs and Tzu Chi Culture and Communication Foundation Partner to Launch Custom Media GPT: Harnessing Technology to Spread Compassionate Values
[Taipei, April 21, 2025] — Taiwan AI Labs and the Tzu Chi Culture and Communication Foundation today announced the signing of a memorandum of understanding to co-develop the nation’s first GPT-powered media solution, built on a domain-specific knowledge base and tailored for video and audio applications.
The service will aggregate over 100,000 videos and media assets collected by Da Ai Television since its founding in 1998. Leveraging AI technologies, it will automatically generate video summaries, scene logs, and continuously learn from data to assist production teams in creating scripts, manuscripts, and other innovative content formats—marking a major milestone in the digital transformation of Taiwan’s media asset management.
This collaboration is a pivotal step in realizing Tzu Chi’s vision of “Spreading the Dharma Through Technology.” For more than two decades, the Foundation has shared messages of truth, kindness, and beauty to over 136 countries and regions through satellite broadcasting, radio, websites, print publications, and social media. In the age of AI, this partnership focuses on revitalizing nearly 60 years of Tzu Chi’s archival content, extracting new thematic insights through AI while allowing media professionals to focus on producing emotionally resonant content—an embodiment of collaborative goodness between humanity and technology.
Duan-Jen Wang, CEO of the Tzu Chi Culture and Communication Foundation, emphasized that the mission is not merely to adopt generative AI, but to develop a large language model that deeply understands and conveys the spirit of Tzu Chi. “It may not yet feel weighty today,” he said, “but in the long run, this initiative bears witness to our times, documenting human history and preserving Tzu Chi’s legacy.”
Eric Yao, Director of Strategic Alignment and Development at the Foundation, noted that this collaboration is more significant than imagined. “We’re entering an era of Four Intelligences,” he said: Awakening Intelligence, without which there is no purification of the mind; Aesthetic Intelligence, which transforms beauty into impactful energy; Artificial Intelligence, which becomes more beneficial to humanity when guided by awakening and aesthetic sensibilities; and Altruistic Intelligence, the ultimate goal that Master Cheng Yen has long sought to realize. This partnership, he added, aims to integrate all four.
Ethan Tu, Founder of Taiwan AI Labs, shared that the teams are developing an AI agent with multimodal and multi-expert capabilities. It will recognize the who, what, when, where, and how of audiovisual content and support Da Ai’s scene-logging workflows, helping media professionals rapidly retrieve relevant information from vast archives.
In a field that prioritizes truth, AI must be professional and trustworthy. Tu pointed out that mainstream open-source models are akin to university interns with general knowledge but little domain-specific expertise. Training such models with sensitive internal data on external cloud services poses risks. In contrast, Taiwan AI Labs’ FedGPT was built under principles of trustworthy and responsible AI governance. Like a freshly minted PhD joining a company full-time, FedGPT can be retrained internally without sharing data externally, learning efficiently and securely through federated learning—loyal to the organization, without leaking sensitive information.
In today’s competitive AI landscape, FedGPT stands out as an internal expert model aligned with corporate sustainability and ESG values, consuming minimal energy while offering high efficiency and deep specialization. At Tzu Chi, FedGPT will act as a diligent new team member, boosting productivity in content production.
FedGPT supports Mandarin, Taiwanese, and English speech recognition, handling over 1.7 million hours of processing per year. It understands local linguistic nuances such as the Taiwanese idiom “Tiām-tiām tsia̍h sann uánn-kong puànn” (Still waters run deep). Unlike international models dependent on the cloud, FedGPT is equipped with Taiwan-specific anti-bias datasets and a Guardian mechanism to curb hallucinations and biases, offering more culturally attuned and accurate results.
Three Core Collaboration Areas for Smart Media Transformation
The collaboration focuses on three core AI services:
- Scene-Logging GPT for Tzu Chi: Built on Taiwan AI Labs’ multimodal, multi-expert FedGPT platform, this tool will automatically interpret video content and build a structured media knowledge base with details like people, events, locations, and objects.
- Topic-Guided Assistant GPT: After selecting a theme, production teams can use FedGPT to curate related historical footage, streamlining the editorial workflow and enhancing real-time content delivery.
- AutoML Platform for Active Learning: Beyond pre-training on domain-specific data, FedGPT will leverage user feedback and internal video content to continuously refine Tzu Chi’s custom language model, improving analysis and productivity.
Ethan Tu emphasized that the year-long collaboration has helped Taiwan AI Labs deeply understand the real-world challenges of media workflows. FedGPT integrates a large language model (LLM), vision-language model (VLM), automatic speech recognition (ASR), enterprise knowledge base (AutoKB), and task assistant (AutoCopilot). Its training spans trillions of Traditional Chinese and English tokens, over 5,000 hours of Taiwanese speech data, and tens of thousands of annotated faces and clothing images—enhancing its ability to recognize local visual and audio patterns.
Even within the same person, changes in angle, hairstyle, or attire can challenge recognition systems. FedGPT tackles this with precise object recognition. Users can define targets—Master Cheng Yen’s robe, Tzu Chi’s signature blue-and-white volunteer uniforms, or specific volunteers’ faces—and retrieve all relevant images or footage instantly. It achieved 94.81% top-1 facial recognition accuracy in Google Images benchmark tests, demonstrating robust performance in complex environments.
For instance, if a team urgently needs footage of Master Cheng Yen’s travels, FedGPT can swiftly locate all relevant clips—even when masks and identical clothing obscure identities—saving countless hours of manual review.
FedGPT offers three levels of scene-log granularity:
- Shot-by-shot visual topic identification.
- Segment-based thematic analysis (e.g., opening, interview, product demo).
- Full-video summaries with thematic essence.
This dramatically reduces the manual labor of drafting scene logs or ensuring content quality. For teams handling dozens of clips daily, simply uploading media to FedGPT yields structured metadata like: “The reporter is conducting a street interview,” or “Cut to Reporter A’s live segment.”
The team also enhanced FedGPT’s contextual understanding of visuals, speech, and subtitles. For example, in disaster footage, it can be discerned that collapsed structures and overturned vehicles indicate a post-earthquake recovery scene, not a single accident.
Tu concluded, “AI grows through learning, and we’ve learned so much from Tzu Chi. Now it’s time to teach GPT to learn proactively—to speak a language aligned with Tzu Chi’s values. In this irreversible transformation of the media industry, this step we’re taking together is not just Taiwan’s first—it’s a world first. Tzu Chi is our first and best teacher.”
Today, many production tasks still rely on manual labor—especially scene logging. With FedGPT, challenges like parsing footage, logging scenes, and labeling dialogue or visuals can now be automated. It can adapt output formats to suit various teams and projects, semantically segmenting and summarizing content to ease editorial burdens.
With its ability to comprehend and analyze massive archives, FedGPT enhances search and editing workflows. It also learns from user feedback, gradually replacing traditional manual training processes. In the future, Tzu Chi’s production teams will interact directly with their custom GPT—not only improving efficiency but also generating content in Tzu Chi’s unique tone, further spreading the organization’s spirit of “united in heart, regardless of religion.”
