Recent publications in Nature detail the deployment of multi-agent AI systems designed to automate hypothesis generation and data analysis in biological research.

近期《自然》雜誌的刊文詳細闡述了多代理 AI 系統的部署，這些系統旨在將生物研究中的假設生成與數據分析自動化。

The emergence of agentic systems, specifically Google's Co-Scientist and FutureHouse's Robin, represents a shift toward the automation of the iterative scientific process. Co-Scientist utilizes a 'tournament evolution' architecture based on the Gemini model to refine hypotheses through continuous critique and reflection, whereas Robin incorporates specialized agents—Crow and Falcon—for literature synthesis and Finch for the automated evaluation of biological screening assays. Both systems demonstrate a capacity for 'combinatorial synthesis,' identifying non-obvious correlations across disparate scientific domains to propose therapeutic candidates, such as the repurposing of ripasudil for macular degeneration or the identification of agents for acute myeloid leukemia. These systems function as 'scientist-in-the-loop' tools, where human experts maintain oversight of experimental validation and hypothesis prioritization.

代理系統的出現，特別是 Google 的 Co-Scientist 與 FutureHouse 的 Robin，代表了科學迭代過程向自動化的轉型。Co-Scientist 採用基於 Gemini 模型的「錦標賽進化」架構，透過持續的評論與反思來完善假設；而 Robin 則整合了專門的代理——Crow 與 Falcon 用於文獻綜合，Finch 則用於生物篩選分析的自動化評估。兩個系統均展現了「組合合成」的能力，能識別跨不同科學領域的非顯著相關性，以提出治療候選方案，例如將 ripasudil 用於治療黃斑病變，或識別急性髓系白血病的藥劑。這些系統作為「科學家在環」的工具，由人類專家維持對實驗驗證與假設優先順序的監督。

Notwithstanding these technical advancements, the institutionalization of AI in science has precipitated significant systemic concerns. There is documented evidence of a convergence in research focus, wherein AI-assisted studies may prioritize established questions over novel inquiries, potentially narrowing the scope of scientific exploration. Furthermore, the proliferation of 'AI slop'—characterized by hallucinated citations and superficially proficient but scientifically inferior prose—has necessitated policy revisions by preprint repositories and academic journals. A critical concern pertains to the erosion of tacit knowledge acquisition; if entry-level analytical tasks are fully automated, the apprenticeship model essential for training early-career researchers to supervise AI workflows may be compromised, thereby jeopardizing the long-term capacity for responsible human oversight.

儘管有這些技術進步，AI 在科學界的制度化已引起顯著的系統性擔憂。有記錄證據顯示研究重點出現趨同，AI 輔助的研究可能會優先考慮既有問題而非新穎詢問，潛在縮小了科學探索的範圍。此外，「AI 廢話 (AI slop)」的泛濫——其特徵為虛構引用以及表面熟練但科學質量低劣的文字——使得預印本庫與學術期刊必須修訂政策。一個關鍵擔憂在於隱性知識獲取的流失；若入門級分析任務完全自動化，對於培訓初級研究員監督 AI 工作流至關重要的學徒模式可能會受損，進而危及長期負責的人類監督能力。

While AI agents significantly accelerate the identification of plausible hypotheses, they remain dependent on human intervention for experimental execution and qualitative validation.

雖然 AI 代理能顯著加速合理假設的識別，但在實驗執行與定性驗證方面，仍依賴人類干預。