![\"Two](\"https:\/\/www.newsbeep.com\/uk\/wp-content\/uploads\/2026\/04\/d41586-026-01278-1_52310302.jpg\"\/)
<\/p>\n <\/p>\n
Agent4Science posts contain AI discussions about AI-generated papers. Credit: Vertigo3d\/Getty<\/p>\n
The latest scientific social network is here \u2014 but unusually, there\u2019s no room for human users. The Reddit-style site, called Agent4Science<\/a>, allows purpose-built AI-powered agents to share, debate and discuss research papers. Human researchers can observe the chatter of artificial intelligence, but only the agents can participate.<\/p>\n The AI discussions are contained in different subgroups, focusing particularly on AI research \u2014 including topics such as AI safety, prompts and deep learning. True to form, even the papers shared in each post are AI generated.<\/p>\n The site is an experiment to have AI agents \u201cfreely discuss science and see where that will lead us\u201d, says one of its creators, Chenhao Tan, an AI researcher who directs the Chicago Human+AI Lab (CHAI) at the University of Chicago in Illinois. Tan\u2019s team had already ventured into this research realm with the site OpenAIReview, to which users can upload a research paper to receive feedback from an AI reviewer. With the new platform, Tan says, the goal is to \u201cimagine a different possibility of what knowledge production could look like\u201d.<\/p>\n Different perspectives<\/p>\n Most of the papers posted on Agent4Science come from the CHAI group\u2019s NeuriCo program, which designs, executes and documents experiments autonomously on the basis of both human and AI research ideas. As agents interact with Agent4Science, they can suggest ideas for research papers and generate them, too.<\/p>\n Although human users can\u2019t contribute to the site, they can create agents to autonomously review and comment on papers and can decide on the agents\u2019 \u2018personalities\u2019 and what research topics they might discuss. Agents take on tags with descriptors including \u2018skeptic\u2019, \u2018academic\u2019 and \u2018storyteller\u2019, and their responses are labelled with indicators such as \u2018supports\u2019, \u2018probes\u2019 and \u2018challenges\u2019.<\/p>\n OpenClaw AI chatbots are running amok \u2014 these scientists are listening in<\/p>\n <\/a><\/p>\n So far, the site lists around 40,000 comments made by more than 150 agents. \u201cThere is rich, interesting discourse going on,\u201d Tan says. \u201cIt gives me new perspectives that I wouldn\u2019t get if I were reading a paper on my own.\u201d He points to an example in which agents debated how the prevalence of untrue and harmful medical information can be reduced in LLMs through engineering prompts<\/a>.<\/p>\n![\"\"](\"https:\/\/www.newsbeep.com\/uk\/wp-content\/uploads\/2026\/04\/d41586-026-01278-1_52046686.png\"\/)
<\/p>\n