In community expansion, the pipeline is an indication from “all-human” to the hybrid “human” model. Today (TOP): Human Note Writers drafted suggested notes in response to misleading posts, with other human contributors assessing their usefulness. The bridging algorithm selects notes that are very useful. Future (Bottom): LLMS also participates in the writing stage, creating notes and supporting human writers, but the evaluation stage remains human-only. Community feedback from human evaluators flows back to improve LLM note generation (RLCF). Credit: Arxiv (2025). doi:10.48550/arxiv.2506.24118
X (formerly Twitter) launched the Community Notes program in 2021, allowing users to add context notes to posts that could lead to deceptive or misunderstanding. An example would be to ensure that users are labeling videos generated by AI so that others are not fooled to believe in events that actually happened. Community Notes are evaluated by decentralized social media communities to determine their usefulness. Only notes deemed useful by the evaluator will be displayed in the post. X’s Community Notes later urged other platforms to launch similar programs.
Up until this point, these community-based fact-checking systems consisted entirely of human-generated notes and human evaluators. However, X is currently operating a new program, allowing AI to generate community notes along with humans in the form of a large-scale learning model (LLM).
The proposed model, recently published by X researchers, integrates both human and AI notes into the pipeline, but still allows humans only to determine which notes are useful. In an age of ramp-prolonged misinformation, researchers believe that the speed and scale of the sound generated by LLMS is necessary. They said, “Enable automatic note creation will allow the system to operate at scale and speeds that are impossible for human writers, providing more content context across the web.”
LLM note generation is further improved by learning from community feedback in a process called reinforcement learning from community feedback (RLCF). This process aims to improve future memo generation through various feedback from community members with different views, and is expected to be more accurate, unbiased and useful notes.
Research issues for LLM-driven community notes. (1) Customized LLM for memo generation. (2) ai ‘co-pilots’ for human writers. (3) AI support for human evaluators. (4) Intelligent Note “Matching” adapts existing useful notes to new similar contexts. (5) Evolution of core algorithms for AI-generated content. (6) Building a robust and open infrastructure. Credit: Arxiv (2025). doi:10.48550/arxiv.2506.24118
While the new model is expected to improve the overall false information checking process, there are some potential risks. Researchers point to the potential issues regarding the persuasive and inaccurate notes generated by AI (problems known in other models) and the risk of oversolute soluteization. There is also concern that the abundance of AI-generated notes will lead to less involvement of human memo writers, and that this abundance may overwhelm the abilities of human evaluators and fully determine what is useful and what is not useful.
This study also describes many future possibilities, including more AI integration into the community notebook pipeline, while maintaining human checking. Future directions may include integrating AI co-pilots to conduct research by human writers and assist more notes faster. Also, researchers have proposed validation and authentication methods for screening human evaluators and writers, customizing LLMs, and methods for adapting and reapplying notes already validated in similar contexts, so evaluators have not evaluated the same concepts many times.
These ways of collaborating with humans may provide nuance and diversity, and provide the speed and scale for LLM to process a wealth of information available online, but there are still quite a few tests to ensure that human touch is not lost. The research authors explain their ultimate goal by saying, “The goal is not to create AI assistants that tell users what to think, but to create an ecosystem that allows humans to think more critically and understand the world better.”
Written for you by author Krystal Kasal, edited by Andrew Zinin. This article is the result of the work of a careful human being. We will rely on readers like you to keep independent scientific journalism alive. If this report is important, consider giving (especially every month). You will get an ad-free account as a thank you.
More information: Scaling human judgments in community notes with Haiwen Li et al, LLMS, Arxiv (2025). doi:10.48550/arxiv.2506.24118
Journal Information: arxiv
©2025 Science X Network
Quote: The pilot program will integrate human community notes on the X platform and human community notes on the X platform obtained on July 4, 2025 from https://techxplore.com/news/2025-07-ai-generated-human-community-platform.html.
This document is subject to copyright. Apart from fair transactions for private research or research purposes, there is no part that is reproduced without written permission. Content is provided with information only.
