For years, the promise of artificial intelligence in science was confined to the digital server rack. AI models analyzed pre-existing datasets, predicted protein folds in digital isolation, or wrote computer code about computer code. But inside a study published in Advanced Science, artificial intelligence co-science has stepped out of the digital incubator and into the messy, unpredictable real world.

Researchers at Explore Science, alongside collaborators across Australia and Canada, built an autonomous multi-agent AI system that completed what researchers call “closing the empirical loop”. With human researchers obtaining ethics approval and manually launching the experiment, the system conceived a scientific hypothesis, registered an experimental protocol, coordinated online recruitment and testing of 288 participants, executed an eight-hour analytical coding session, and produced a 31-page manuscript. The study illustrates an application of AI Co-science beyond digital simulations: autonomous knowledge generation tested directly against physical reality.

Publication in Advanced Science, a journal published by Wiley-VCH, provides external assessment through the journal’s peer-review process by expert human scientists.

Winning the AI co-science race

Explore Science has staked its claim on the AI co-science race. Standard large language models notoriously fabricate academic citations and degrade when tasked with long, complex projects. Explore Science says it designed its architecture to prioritize accuracy, verification, and open-science principles.

Using a Dynamic Retrieval-Augmented Generation system, the AI queries live scientific databases like PubMed, Semantic Scholar, and OpenAlex in real time rather than relying on frozen training data. Explore Science says the platform programmatically checks manuscript citations against live doi.org records to help identify fabricated references.

The system embeds structural safeguards against human-like research biases. It generates an Open Science Framework pre-registration report before acquiring data, locking in hypotheses to prevent selective reporting. When analyzing large human datasets, the AI prioritizes practical effect size over statistical tricks, actively resisting the pervasive academic obsession with p-values. Explore Science says the system is designed to document negative or non-significant outcomes alongside positive findings, potentially helping other laboratories identify approaches that may be less promising. Human ethics and security safeguards were maintained throughout, with protocol approval obtained from the Bellberry Human Research Ethics Committee and complete de-identification of participant data.

Shipped products vs. Silicon Valley roadmaps

Explore Science has delivered a live, customer-facing platform available to enterprise laboratories today. The company has also engineered its platform to manage other people’s research across diverse fields like cognitive neuroscience, biology, and clinical trials.

To achieve this, Explore Science did not attempt to win the multi-billion-dollar foundation model race. Instead, its proprietary engine, Explorer One, operates as an overarching orchestrator sitting atop the world’s leading AI models. During a research project, Explorer One deploys a hierarchical network of over 50 specialized agents powered by Claude, OpenAI, Gemini, Grok, and Pixtral. Explore Science says the platform routes specific micro-tasks to models selected for those tasks and cross-checks outputs across different model families to help reduce single-model bias and hallucinations.

Research timelines and computing costs

The system’s reported processing times and costs suggest potential efficiency benefits. Across its validation studies, the autonomous architecture executed complete end-to-end scientific projects in an average processing time of approximately 17 hours at an average marginal compute cost of just $114 USD per project, excluding human participant payments. Driven by a master orchestrator coordinating over 50 specialized AI agents, the pipeline successfully executed online data collection with 288 live human participants. The system demonstrated long-horizon autonomy, maintaining continuous goal-directed execution across an average of 72 action-observation cycles and over 8 hours of uninterrupted coding and debugging sessions per study. This analytical workflow generated an average of 7,696 lines of functional statistical code per run to process heterogeneous raw data, fit multi-level mixed-effects models, and auto-render publication-ready LaTeX figures and tables. Finally, to help check reference authenticity, the platform integrated a live doi.org validation engine that, according to the company, checked cited references against external records to help identify fabricated references prior to final document assembly.

A new framework for human discovery

Ultimately, Explore Science’s achievement suggests a possible shift in how knowledge is created. Reported marginal computing costs of approximately $114 suggest potential to make some research tasks more affordable for underfunded institutions, including those in developing nations, although this figure excludes participant payments and other research expenses.

Far from making human scientists obsolete, these systems liberate researchers from tedious scripting, data cleaning, and formatting, leaving human minds free to focus on grand theoretical synthesis, creative problem formulation, and ethical guidance. Explore Science’s work suggests that AI’s potential in science extends beyond speed to supporting more rigorous, transparent, and trustworthy research.


VentureBeat newsroom and editorial staff were not involved in the creation of this content.