# AI Found a Million Research Ideas Humans Miss > Source: > Published: 2026-10-30 03:21:59+00:00 ## Why AI keeps suggesting the same science Asking AI models for fresh research ideas often yields familiar results. Models trained on existing scientific literature tend to reproduce its popular topics, failing to escape established intellectual grooves. Their suggestions cluster around well-trodden paths because their knowledge base is inherently retrospective. Consider [Claude](https://www.stork.ai/en/claude-ai-workspaces): when prompted for machine learning research ideas, it frequently suggests topics related to **interpretability**. This echoes a common trend within the field, reflecting a high-density area of current human research. The model, in essence, reflects the biases present in its training data. A similar dynamic affects human researchers. Their training, collaborators, and disciplinary networks subtly steer them toward familiar questions. This social and cognitive embeddedness can lead to overlook plausible intersections or "alien spaces" of science that no one happens to be in the right position to conceive. Both AI and humans, therefore, face a shared challenge in breaking free from intellectual consensus. ## The hunt for ideas outside the map Researchers developed a system to probe the “**alien space**” of science—scientifically coherent ideas that researchers are unlikely to conceive. This isn't about impossible concepts, but rather combinations of existing knowledge that fall outside typical cognitive patterns or social networks. Think of it as mapping uncharted but navigable territories within the scientific landscape. The team began by deconstructing around **16,000 machine-learning papers** into fundamental “idea atoms.” These modular conceptual units allowed them to represent every research direction as a set of basic components, enabling systematic recombination into novel candidate ideas. They then employed a two-part scoring mechanism. A **coherence model**, a 43-million-parameter transformer, assessed whether a proposed combination of idea atoms formed a logically sound research project. Concurrently, a **cognitive availability model**, trained on the publishing histories of almost 40,000 researchers, estimated the likelihood that any real-world research group would spontaneously connect those specific concepts. The system prioritizes ideas that score high on coherence but low on human availability, pushing discovery beyond established intellectual boundaries. ## A striking test—and a reality check Researchers rigorously tested their system by freezing its training data at the end of 2024. Then, they evaluated its ability to surface concept combinations that materialized in actual research papers published in 2025. This **retrospective validation** aimed to determine if the system could genuinely identify future directions. One striking result involved an ICLR 2025 oral presentation. This paper's core combination, initially ranked around 1,489 based solely on coherence, surged into the **top 25 out of more than 1.14 million candidates** when the system's "unavailability" metric was applied. It is crucial to note this indicates the system's predictive ranking power, not that it directly caused the research. Further, 50 AI professionals with graduate degrees blindly evaluated research proposals. The system-generated ideas accounted for three of the five top-rated reports. This contrasts sharply with Claude's performance, where eight out of ten self-generated ideas drifted toward interpretability; the new system produced **zero interpretability-focused proposals**, demonstrating its capacity to explore truly novel conceptual space. For deeper technical insights, consult the full paper: [The Alien Space of Science: Sampling Coherent but Cognitively Unavailable Research Directions](https://arxiv.org/abs/2603.01092). Enjoying this? Get one like it in your inbox each morning. one email a day · unsubscribe in two clicks · no third-party tracking ## The bottleneck is still the evidence Approach depends critically on data access. Researchers built their system with **16,068 full-text machine learning papers**, a field where open access is prevalent. This allowed the models to learn robust "idea atoms" and their plausible combinations. Conversely, fields like biology and medicine present a challenge. Roughly half of medical research remains behind **paywalls**, severely limiting the system's ability to ingest the necessary full-text papers. Consequently, ideas generated in these areas are notably weaker due to insufficient training data. Generating plausible combinations is distinct from scientific validation. The system identifies potential avenues, but human experts must still: - Judge proposals for true novelty and feasibility - Secure funding and resources - Design and execute experiments A compelling idea is not a proven discovery; it is merely a hypothesis awaiting rigorous testing. Readers can explore the **[Lacuna](https://www.stork.ai/en/lacuna)** tool, which hosts 25,000 research proposals, or delve into the open-source project itself. Wider adoption of this methodology in critical fields like medicine hinges on greater access to full-text research and the dedicated involvement of domain experts for thorough review and experimental follow-through. ## Frequently Asked Questions ### What is the Alien Space of Science? It is a framework for finding research directions that appear scientifically coherent but are unlikely to occur to researchers given their existing habits and connections. ### How does the system generate research ideas? It combines research concepts called idea atoms, then scores candidate combinations for coherence and for how likely researchers are to think of them. ### Did the system identify real research before publication? In a retrospective test using data frozen before 2025, it ranked a combination appearing in an ICLR 2025 oral paper among its top 25 candidates. ### Can the system invent and verify discoveries on its own? No. It proposes hypotheses, but human experts must evaluate them and conduct experiments to establish whether they are useful or true. ### What is Lacuna? Lacuna is a tool associated with the project that lets users browse research proposals; the project’s code is also available on GitHub.