AI Scientists Need a Social Network

The speed and scale benefits are obvious, but perhaps the deeper opportunity for AI scientists will be in how they communicate and make science, for the first time, genuinely social

Dear SoTA,

Is science social?

AI scientists are systems capable of autonomously supporting or leading research. By deploying at scale and enabling fast, controlled scientific cycles, this technology has the potential to conduct significantly more research and explore far more avenues than was previously possible, particularly in fields that can be simulated or are amenable to automation using physical hardware. Human-led scientific discovery has led to the technological capabilities we enjoy today, but the path trodden has been highly stochastic and fraught with dead ends. Discoveries can’t be planned, and often stem from others’ previous ideas, so exploratory research is essential for progress [1]. However, a discovery only changes the world when it is shared.

Capability, scale and speed of communication have historically been the three levers driving scientific progress. Each has progressed in their own way over time, but despite the internet now enabling instant communication, how we share scientific knowledge today would be recognisable to a researcher from one hundred years ago. It is widely acknowledged that the current scientific publication system is far from perfect [2,3,4]. The ‘publish or perish’ incentive structures and prestigious paywalled journals mean that negative results go unreported, silos lead to repeated reinvention of the wheel, and a lack of qualified peer reviewers slows dissemination. For many research areas, the published record is the tip of the iceberg, showing what worked and fit the incentives at the time, with the bulk of information on exploration disappearing without a trace. Even when research is published, it often enters a sea of literature too vast for human researchers to stay up to date with. Perhaps the most famous scientific casualty of this was Charles Darwin’s establishment of the theory of evolution, during which he remained entirely unaware of the foundational work led by the “father of genetics”, Gregor Mendel. Such barriers to effective dissemination are not a capability or scale problem but a social and communication one [5]. Despite many researchers’ collaborative spirit, science is currently sub-optimally social.

GitHub revolutionised software development when it launched, by enabling developers to save code and collaborate. Over time, it has evolved to become a low-friction social platform tied to real-world incentives. Many attempts have been made to make research more collaborative and replicate GitHub’s successes in scientific domains. But to date, nothing has captured a critical mass or seen compounding network effects for the scientific community as a whole. Indeed, around the time of GitHub’s launch, there was a lot of discussion on this topic [6,7], and yet over a decade later, little has changed. One of those leading the charge is arXiv, which has pushed the boundary on fast, open science, particularly in computationally focused fields. The idea of bringing more structure and openness to scientific information is gaining traction; for instance, the Open Science Archive is applying similar thinking to scientific data [8]. Publication in leading journals and conferences, however, currently still carries with it a prestige that is hard to replicate. Unlike software, many scientific experiments are hard to replicate and require significant investment in infrastructure, which closes off many experiments to all but the most well-funded labs. Computers, by comparison, are ubiquitous in modern life, and a contribution of a single line change can fix an overarching bug. All this is to say, we haven’t yet found the optimal social mechanism or structure for incentivising, sharing and accelerating open scientific progress. A new communication model is needed; could this be brought about with AI scientists?

Social networks for AI agents

As the internet emerged and many communities gained an online presence, so did social networks to facilitate their interactions. Think Facebook for college students, LinkedIn for professionals and a Reddit page for just about anything. In a world where a large portion of all digital content is produced by agents, we should expect social networks for these agents to emerge. In January 2026, Moltbook gave the first glimpse of this future [9]. Moltbook is similar to Reddit in that it is designed to be an internet forum, but instead of humans posting, it is exclusively for AI agents.

Large corporations are already beginning to delegate parts of their digital workload to agentic systems [10,11]. This includes everything from internal IT support to customer service and sales. As more and more of this work is conducted by agentic systems, it seems inevitable that social networks, both internal and external, should emerge for agents. For example: if an agent needs to understand a complex area of its organisation outside its purview, how can it ask other agents for input; if an agent is about to make a high-stakes decision, how can it seek a review or brainstorm; if an agent is collaborating with one at a different organisation, how can it share updates and align on priorities. These kinds of interactions currently exist for human employees, through social networks such as emails, Slack messages, face-to-face meetings and more. Without the ability to communicate across social networks, agents will either operate blindly, repeating the same mistakes, or remain bottlenecked whilst waiting on human input.

A social network for AI scientists

Research is no different to other domains already mobilising AI agents. Without a social approach and effective means of communication, deploying AI scientists at scale risks compounding current inefficiencies rather than solving them. Much like horizontal gene transfer allows bacteria to share capabilities laterally, social networks could enable insights, failures, and tactics discovered by one agent to spread rapidly. We believe social networks for AI scientists will lay the foundations to open up science to those who previously couldn’t participate, and accelerate progress. There are many forms a platform like this could take; here are two of our ideas. First, a Moltbook-style implementation could operate like a Reddit community for AI scientists, to which participating labs connect their agents to discuss and share all aspects of their research. Indeed, we already see signals that the community will converge on such a platform. Since the time of writing, one such social network has emerged in beach.science, where users can connect agents that generate and share hypotheses, then pay for real-world experiments [12]. A more structured implementation could look like GitHub-for-science, a tracked, branching record of cumulative scientific progress.

Either of these approaches would function as an agent-native communication infrastructure for science. Unlike current approaches to research, the inputs, outputs, logs, and reasoning traces of AI scientist systems are naturally encoded in a digital format. Recording these in a structured, traceable, versioned and reproducible way has the potential to act as a significant scientific accelerator. Rather than waiting for polished papers, AI scientists could incrementally post/push this data, sharing information about current scientific problems, hypotheses, methodological choices, intermediate results, and incremental findings, whether positive or negative. Agents could also ask for help, give feedback, and contribute ideas or knowledge to others’ works, creating a live, cross-institutional exchange. Such a platform would act as a high-frequency layer that reflects how research actually unfolds. Organisations could connect their AI scientists via APIs, with public and private access controls allowing both open science initiatives and commercial R&D to engage. Finally, both agents and humans could participate, combining the scale and speed of AI scientists with the tacit knowledge and nuanced judgment that human researchers bring.

One key reason similar platforms have failed before is incentives. Contributing to open platforms has not historically led to more grants, and producing shareable artefacts can feel time-consuming and strategically costly for researchers. With AI scientists, this changes fundamentally; rather than a burden, sharing becomes an easy path to faster feedback, measurable impact and new forms of credit. Utility metrics, capturing the community’s reaction (upvotes and responses) and how a contribution is built upon, could replace metrics like impact factor and the currently ineffective peer review process. With a GitHub-like architecture, ‘pull requests’ and citations could also be handled by historic branches and ‘dependencies’. Incremental progress would be credited rather than lost, and manuscripts need not be rewritten wholesale for minor advances. Much like GitHub took off as it solved problems around saving work and collaboration, agentic AI scientists will also need this functionality, incentivising the use of common platforms where research can be saved, inspected, and built upon.

Similar to good software engineering practices, community-accepted, generalisable, structured forms of this information may evolve, for example, methodologies being reported in structured schemas, and a distinction made between what has been proven and what is opinion. Language is, of course, only one way to communicate, and novel methods that overcome the semantic limitations of natural languages may emerge, such as sharing information through embeddings. Human learnings from working with AI scientists, including safety concerns, failure modes and unexpected behaviours, could also be shared to ensure we realise the world we want with this technology. Over time, a structured longitudinal temporal record of experiments and scientific discovery will be created. This would capture the full journey of discovery and enable compounding progress through transparent, structured exchange; it could also lay the groundwork for further downstream benefits we cannot yet anticipate, just as GitHub unintentionally laid the path for coding agents.

As a field, we should not be naive about the risks. As has been seen on social media platforms, AI slop will rise. Moderation systems need to be put in place to tackle this, though non-useful projects are abundant on GitHub and don’t detract from its utility. There are genuine questions around whether such systems in their current state can be truly novel, with recent evidence suggesting that while AI can boost individual scientific productivity, it may also narrow the research and reduce follow-up interactions between studies [13]. One could argue, though, that a network of research agents sharing structured intermediate results is valuable even if those agents aren’t autonomously generating breakthroughs, and by rewarding exploration, iteration, and work on underserved areas, you can prevent groupthink and obtain a reasonable balance between breadth and depth. Additionally, information hazards reaching bad actors represent a serious challenge, with some form of filtering or controlled access likely required. Finally, the broader ethical questions raised by autonomous scientific systems will also need ongoing consideration as the technology matures. Indeed, a shared platform could help the community converge on answers to questions such as who is responsible for harm, how credit should be assigned, and what governance model is appropriate.

We believe it is not a question of if social networks for AI scientists will emerge, but whether they will be built with the right incentives, safeguards and global interests in mind from the outset. Done well, this could be the catalyst that truly democratises science and delivers on the transformative potential of AI scientists.

If you are building in AI for science, open research or agent communication, or simply think we are wrong, we would love to hear your thoughts.

Yours,

William Bolton & Ben Williams

---

Authored by William Bolton and Ben Williams, Encode fellows on Pillar VC’s AI for Science initiative, powered by ARIA and DSIT, working at the University of Oxford and Imperial College London.

---

References

1. Lehman, J. & Stanley, K.O. (2015). Why Greatness Cannot Be Planned: The Myth of the Objective. Springer. https://link.springer.com/book/10.1007/978-3-319-15524-1

2. Ioannidis, J.P.A. (2005). Why Most Published Research Findings Are False. PLOS Medicine 2(8): e124. https://doi.org/10.1371/journal.pmed.0020124

3. DORA (2012). San Francisco Declaration on Research Assessment. https://sfdora.org/read/

4. National Academies of Sciences, Engineering, and Medicine (2017). Advancing Reproducibility in Research. https://www.nationalacademies.org/projects/DBASSE-BBCSS-17-03/publication/25303.

5. Channing, G., & Ghosh, A. (2026). AI for scientific discovery is a social problem. Patterns, 7(3). https://www.cell.com/patterns/fulltext/S2666-3899(26)00006-1

6. Bedford, T. (2012). A GitHub of Science? https://bedford.io/blog/github-of-science/

7. Von Muhlen, M. (2012). I Want a GitHub of Science. https://marciovm.com/i-want-a-github-of-science

8. Byrne, R. (2025). Open Science Archive: An open-source, domain-agnostic scientific data archive. https://opensciencearchive.org/

9. Moltbook (2026). https://www.moltbook.com/

10. PwC (2025). AI Agent Survey. https://www.pwc.com/us/en/tech-effect/ai-analytics/ai-agent-survey.html

11. McKinsey Global Institute (2025). The State of AI. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

12. Weidener, L (2026). arXiv:2602.19810. https://doi.org/10.48550/arXiv.2602.19810

13. Hao, Q., Xu, F., Li, Y., & Evans, J. (2026). Artificial intelligence tools expand scientists’ impact but contract science’s focus. Nature, 1-7. https://www.nature.com/articles/s41586-025-09922-y