Pilot program integrates AI-generated notes with human community notes on X platform

X (formerly Twitter) launched its “Community Notes” program in 2021 to combat misinformation by allowing users to add contextual notes on posts that might be deceptive or lead to misinterpretation. An example would be users labeling an AI-generated video as such, so that other users would not be tricked into believing the event in the video actually occurred. Community notes are rated by the decentralized social media community to determine their usefulness. Only the notes determined useful by raters are shown on the post. X’s Community Notes later inspired other platforms to launch similar programs.

Until this point, these community-based, fact-checking systems consisted entirely of human generated notes and human raters. However, X is now piloting a new program, allowing AI—in the form of large learning models (LLMs)—to generate community notes alongside humans.

The proposed model, published recently by X researchers, integrates both human and AI notes into the pipeline, but still only allows humans to determine which notes are helpful. In an age of rampant misinformation, the researchers believe the speed and scale of notes generated by LLMs is necessary. They write, “allowing automated note creation would enable the system to operate at a scale and speed that is impossible for human writers, potentially providing context for orders of magnitude more content across the web.”

The LLM note generation will be further improved by learning from community feedback in a process referred to as reinforcement learning from community feedback (RLCF). This process is meant to refine future note generation through a diverse array of feedback from community members with a variety of views and is expected to result in more accurate, unbiased and helpful notes.

Although the new model is expected to improve the misinformation checking process overall, there are some potential risks. The researchers note potential issues with AI-generated notes being persuasive and inaccurate—a known issue with other models—and a risk of over-homogenizing notes. There is also some concern that human note writers might engage less often, due to the abundance of AI-generated notes, and that this abundance might overwhelm the capacity of human raters to sufficiently determine what is helpful and what is not.

The study also discusses many future possibilities involving even more AI integration into the community note pipeline, while still keeping human checks in place. Future directions might involve integrating AI co-pilots for human writers to conduct research and put out more notes faster and AI-assistance to help human raters audit notes more efficiently. The researchers also propose verification and authentication methods for screening human raters and writers, customization of LLMs and methods for adapting and reapplying already-validated notes to similar contexts, so that raters are not rating the same concepts over and over.

There is potential for these human-AI collaboration methods, with humans providing nuance and diversity and LLMs providing the speed and scale to deal with the abundance of available information online, but there is still a good deal of testing to be done to ensure the human touch is not lost. The study authors describe their end goal by saying, “the goal is not to create an AI assistant that tells users what to think, but to build an ecosystem that empowers humans to think more critically and understand the world better.”

© 2025 Science X Network