CheckAI: um Bot para Checagem de Fatos via RAG no Telegram

• Gabriel Guimarães Souza UFCG
• João Lucas Gomes Brandão UFCG
• Daniel Mantovani UFCG
• Leandro Balby Marinho UFCG
• Claudio E. C. Campelo UFCG
• Daniel Kansaon UFMF
• Fabrício Benevenuto UFMG

DOI: https://doi.org/10.5753/webmedia_estendido.2026.34756

Resumo

Messaging apps have become important channels for the rapid spread of potentially misleading content, yet users often lack timely access to professional fact-checks within the same environment. In this paper, we present CheckAI, a Telegram bot for real-time fact-checking. The system relies on a hybrid Retrieval-Augmented Generation (RAG) pipeline that combines lexical and vector retrieval with validation mechanisms applied to both user queries and retrieved evidence. Its knowledge base contains 16,123 fact-checks from four Brazilian fact-checking organizations: Agência Lupa, Aos Fatos, Boatos.org, and G1 Fato ou Fake. In a preliminary evaluation, CheckAI achieved 90.0% accuracy with an average response time of 14.3 seconds, compared with 73.3% accuracy and 43.97 seconds for the reference system. The bot also correctly abstained from producing a verdict when no relevant evidence was available in the knowledge base. Overall, these findings provide initial evidence that the proposed approach can support timely access to fact-checks directly within messaging platforms.

Palavras-chave: misinformation, fact-checking, chatbots, Telegram, hybrid RAG, guardrails

Referências

Agência Lupa. [n. d.]. Agência Lupa. [link].

Aos Fatos. [n. d.]. Aos Fatos. [link].

Aos Fatos. 2024. Aos Fatos lança Fátima 3.0, expansão do chatbot com IA generativa. [link].

Boatos.org. [n. d.]. Sobre o Boatos.org. [link].

Gordon V. Cormack, Charles L A Clarke, and Stefan Buettcher. 2009. Reciprocal rank fusion outperforms condorcet and individual rank learning methods (SIGIR ’09). 758–759. DOI: 10.1145/1571941.1572114

Kai Greshake, Sahar Abdelnabi, Shailesh Mishra, Christoph Endres, Thorsten Holz, and Mario Fritz. 2023. Not What You’ve Signed Up For: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection. In Proceedings of the 16th ACM Workshop on Artificial Intelligence and Security (Copenhagen, Denmark) (AISec ’23). 79–90. DOI: 10.1145/3605764.3623985

Jornal da USP. 2026. Chatbot contra desinformação criado por alunos da USP vence desafio internacional de IA. [link].

Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Proceedings of the 34th International Conference on Neural Information Processing Systems (Vancouver, BC, Canada) (NIPS ’20). Curran Associates Inc., Red Hook, NY, USA, Article 793, 16 pages.

Gustavo Resende, Philipe Melo, Hugo Sousa, Johnnatan Messias, Marisa Vasconcelos, Jussara Almeida, and Fabrício Benevenuto. 2019. (Mis)Information Dissemination in WhatsApp: Gathering, Analyzing and Countermeasures. In The World Wide Web Conference (San Francisco, CA, USA) (WWW ’19). 818–828.

Gustavo Resende, Johnnatan Messias, Márcio Silva, Jussara Almeida, Marisa Vasconcelos, and Fabrício Benevenuto. 2018. A System for Monitoring Public Political Groups in WhatsApp. In Proceedings of the 24th Brazilian Symposium on Multimedia and the Web (Salvador, BA, Brazil) (WebMedia ’18). 387–390.

Stephen Robertson and Hugo Zaragoza. 2009. The Probabilistic Relevance Framework: BM25 and Beyond. Found. Trends Inf. Retr. 3, 4 (April 2009), 333–389.