Agentic AI Poses Existential Threat to Research Funding Systems, Experts Warn 专家警告：代理式AI对研究资助体系构成生存威胁

TL;DR

• Autonomous AI agents can now write and submit grant applications without human intervention, fundamentally altering the landscape of research funding.
• Grant application volumes surged by 57% between 2022 and 2025, driven by the near-zero marginal cost of producing AI-generated proposals.
• Key risks include quality compression, where AI raises the baseline but not the ceiling, and long-term convergence of ideas due to shared training data.
• Experts argue that banning AI is unenforceable and ineffective, necessitating a shift in assessment strategies.
• Future evaluation frameworks must prioritize human-specific traits like career-stage track records and demonstrable originality that agents cannot replicate.

Why It Matters

This development poses a critical threat to the integrity of academic research funding by decoupling proposal generation from genuine human effort. For researchers and institutions, it signals an urgent need to adapt to a high-volume, low-cost environment where traditional metrics of merit may become obsolete. The situation demands immediate strategic adjustments to ensure that funding continues to reward true innovation rather than superior AI simulation.

Technical Details

• Agentic vs. Generative AI: Distinction made between tools that polish text and autonomous agents that gather data, draft proposals based on specific criteria, and submit them independently.
• Volume Metrics: Data from 12 funders across seven research systems shows a 57% increase in application volume from 2022 to 2025, with acceleration projected into 2026.
• Quality Compression: Analysis indicates that while AI raises the minimum quality floor, it fails to elevate the maximum quality ceiling, leading to homogenization.
• Convergence Risk: Agents trained on identical historical data risk creating a feedback loop where funding decisions reflect how well AI mimics past successes rather than fostering new ideas.

Industry Insight

Funding bodies and universities must redesign assessment frameworks to focus on attributes that AI cannot easily replicate, such as long-term career trajectories and unique intellectual contributions. Relying solely on the written proposal is no longer viable; evaluators should prioritize verifiable evidence of originality and past performance. Institutions should also invest in transparency tools to detect AI-generated content and adjust their internal support structures to help researchers navigate this new competitive reality.

TL;DR

• 自主AI代理（Agentic AI）能独立撰写并提交基金申请，其边际成本趋近于零，对科研资助体系的完整性构成根本性威胁。
• 数据显示，自2022年ChatGPT发布至2025年底，12个资助机构在7个研究系统中的基金申请量激增57%，且2026年增速进一步加快。
• 面临三大挑战：申请量激增导致审核压力、AI润色导致质量“地板”抬高但“天花板”未升、以及训练数据同质化导致的评审思维收敛。
• 专家反对全面禁止AI，主张资助方和高校应重构评估框架，重点考察AI无法复制的职业发展轨迹和原创性思想。

为什么值得看

这篇文章揭示了从辅助性生成式AI向自主性代理AI转变带来的系统性风险，特别是针对科研资助这一高价值场景。对于政策制定者和科研管理者而言，它提供了关于如何重新定义“人类贡献”和构建抗AI干扰评估机制的关键战略方向。

技术解析

• 自主代理与生成工具的区别：Geraint Rees强调，传统生成式AI仅用于润色，而自主AI代理具备独立收集信息、根据研究者过往成果及资助标准起草提案并自动提交的能力，实现了全流程自动化。
• 量化增长趋势：UCL的研究团队分析了2022年至2025年的数据，证实了AI普及后申请量的显著上升（+57%），并指出2026年的数据表明这一加速趋势仍在持续，反映了技术对行为模式的深刻影响。
• 同质化风险机制：当AI代理基于相同的训练数据（即过去的获奖提案）进行写作和评审时，可能导致资助决策退化为衡量“AI模拟以往受奖励思维”的能力，而非真正的科学创新，造成长期思维收敛。

行业启示

• 重塑评估标准：资助机构和高校必须摒弃单纯依赖文本质量的评审方式，转而建立以“职业阶段研究轨迹”和“可验证的原创性”为核心的新评估体系，以区分人类智慧与AI生成内容。
• 应对质量通胀：随着AI拉高了所有申请的基线水平（floor），评审者需警惕“质量压缩”现象，即虽然错误减少但突破性创新（ceiling）并未增加，需调整筛选策略以识别真正的差异化贡献。
• 适应性治理：鉴于禁止AI执行不可行，行业应采取适应性治理策略，通过技术手段（如检测AI参与度）和制度设计（如要求披露AI使用范围）来管理风险，而非简单封堵。