TotooBa?: Checks disaster and election claims against official sources, in Filipino

Problem Statement

Filipinos are among the heaviest social media users in the world, and misinformation spreads fastest when people are scared or emotional: during typhoons, earthquakes, and elections. Fake storm signals, edited PAGASA track maps, false “class suspension” announcements, fake evacuation orders, and fabricated COMELEC rules travel through Facebook, TikTok, and group chats faster than corrections. Deepfake audio and video of public officials are becoming easier to make, and COMELEC has issued guidelines on AI use in campaigns. Fact-checking groups like VERA Files and Tsek.ph do important work but can’t respond in real time, and their checks rarely reach the barangay group chats where rumors start.

Proposed Solution

TotooBa? works on Messenger and Viber, as a browser extension, and as a web page. Users paste a claim or link, or upload a screenshot, short video, or audio clip. The system:

1. extracts the factual claims, using OCR or speech-to-text for media
2. retrieves the latest official information from a continuously updated index of PAGASA bulletins, PHIVOLCS advisories, NDRRMC sitreps, DepEd/LGU suspension announcements, COMELEC resolutions and calendars, and PIA releases
3. compares the claim to these sources with a retrieval-augmented LLM.

The answer comes in Filipino, Taglish, or a regional language, with a verdict (Totoo / Mali / Kulang sa konteksto / Hindi pa ma-verify), a two-line explanation, the official link, and a shareable correction card. For media, a deepfake-risk module checks for synthetic voice and face artifacts and metadata. Aggregated claims feed a “rumor radar” for PIA, NDRRMC, and COMELEC so they can publish quick corrections.

Target Market

• Primary users: Filipino social media users, about 85M+, especially in disaster-prone provinces and during election season. Also barangay officials and teachers who moderate group chats.
• B2G buyers: PIA, NDRRMC/OCD, PAGASA, COMELEC, DICT, and LGU information offices.
• B2B/partners: Newsrooms and fact-check networks (VERA Files, Tsek.ph, Rappler), platforms, and broadcasters who want verified, citable answers fast.

How AI Is Used

• Retrieval-augmented LLM: Matches free-form Taglish claims to the right official bulletin and explains the result. Keyword search can’t connect “Signal no. 4 na sa Iloilo bukas” to the latest PAGASA tropical cyclone bulletin.
• Claim extraction: Breaks long posts into checkable statements and skips opinions.
• OCR and speech-to-text (Whisper): A lot of misinformation comes as screenshots, TikToks, and voice messages.
• Deepfake detection: Pretrained audio anti-spoofing and face-forgery classifiers provide a risk score, not a final verdict. Without them, there is no defense against cloned voices of officials.
• Clustering and trend detection: Groups variations of the same rumor so agencies can respond to one cluster instead of thousands of individual posts.

PLDT / Smart / ePLDT / Rokid Angle

• Smart: Free short code or zero-rated access during disasters, when data is expensive and connections are weak. Verified correction cards can be pushed alongside emergency SMS.
• ePLDT: Hosting for the official-source index and the rumor radar, with high availability during typhoons when traffic spikes.
• PLDT Enterprise: A B2G “trusted information” service for PIA, COMELEC, and LGUs. It fits PLDT’s role as a disaster-resilient network provider.
• Rokid: Optional. Field reporters or barangay responders could get live verification overlays on AR glasses during evacuations.

Demo Scenario (3 minutes for judges)

1. Show a viral-looking fake post: “BREAKING: PAGASA – Signal No. 5 sa Metro Manila, walang pasok buong linggo!” with an edited track map.
2. Forward it to TotooBa?. Within seconds: MALI. “Ayon sa PAGASA Tropical Cyclone Bulletin #7 (5 AM), Signal No. 2 lang ang nakataas sa Metro Manila.” It includes the official link and a correction card ready to share in the group chat.
3. Upload a 10-second audio clip of a “COMELEC official” announcing a postponed election, made with a voice cloning tool by the team for the demo. The deepfake module flags high synthetic-voice risk, and RAG finds no matching COMELEC resolution.
4. Paste a true claim (e.g., an actual class suspension from a city LGU page). It returns TOTOO with a link, showing the tool is not biased toward saying “fake.”
5. Show the rumor radar: the top 5 rumor clusters right now, by region, each with a suggested correction text for PIA.

30-Hour Build Plan

Suggested stack: Python, LangChain or LlamaIndex, Gemini or other multilingual LLM, pgvector or Chroma, OpenAI Whisper, Hugging Face deepfake/anti-spoofing models (e.g., ASVspoof-trained), Pillow for cards, Next.js dashboard. Public data / APIs: PAGASA bulletins and public data, PHIVOLCS earthquake and volcano bulletins, NDRRMC situational reports, COMELEC resolutions and election calendar, PIA news releases, official LGU Facebook pages (public posts), Tsek.ph and VERA Files published fact-checks, Hugging Face fake-news and deepfake datasets.

Risks & Privacy Notes

• Wrong verdicts: Always cite sources. Use “Hindi pa ma-verify” when no official source exists. Update the index every few minutes during disasters. Partner fact-checkers review high-traffic claims.
• Deepfake detector limits: Present results as a risk score with a disclaimer, never as proof.
• Political neutrality: Check only factual claims, not opinions. Publish the methodology, and treat all candidates the same way.
• Data Privacy Act 2012: Remove names and phone numbers from forwarded content before analysis. Store only aggregated claim clusters, and make the consent notice clear.
• Clause 24 / competition rules: Use synthetic deepfake samples created by the team with consent, and public bulletins only. No personal data goes into AI tools.

AI Use Disclosure (draft)

TotooBa? uses a multilingual LLM (e.g., Google Gemini) with retrieval augmentation to extract and verify claims and write Filipino explanations, OpenAI Whisper for speech-to-text, OCR for screenshots, and pretrained Hugging Face models for deepfake risk scoring. The demo deepfake audio was generated by the team, with consent, using a voice cloning tool. AI coding assistants helped scaffold scrapers and UI. The team wrote and reviewed all verdict templates and evaluation sets, and checked demo outputs by hand.

Rubric Scoring