当前人工智能模型存在根本性缺陷,其训练数据主要来源于思维的“痕迹”,即最终产物,而非包含试错与修正机制的完整思维过程 1。即便是具备所谓“推理”功能的AI系统,其输出的步骤报告本质上仍是一种结果展示,往往无法真实反映底层的实际计算逻辑 1。这种基于结果的训练模式限制了AI的真正创造力,使其难以像人类一样从挑战和失败中汲取经验 1。
这一技术局限正导致人类用户对AI产生过度依赖,进而引发专业技能的显著退化 1。历史教训早已显现,早在1997年,美国航空机长就曾警告飞行员因过度依赖自动化而沦为“洋红线之子” 1。如今类似现象在医疗领域重现,使用AI辅助结肠镜检查后,医生独立检测息肉的能力出现下降 1。在教育场景中,依赖生成式AI的学生在移除工具后数学成绩下滑,且研究表明仅需约十分钟的AI辅助,就足以让人在没有AI支持时表现变差并更早放弃解决问题 1。
为缓解这一困境,专家建议重构AI与教育系统的交互逻辑,将“过程”纳入核心设计 1。理想的解决方案是让AI学会从错误中学习,同时开发能引导而非替代思考的教育工具 1。哈佛大学的试验证实了这一方向的有效性:由设计良好的AI物理导师辅导的学生,因为被强制完成生成性工作,比在传统主动学习课堂中的学生学得更快更好 1。
Current artificial intelligence models are fundamentally limited because they learn from the "traces" of human thought—the final products—rather than the dynamic process of thinking itself. 1 This training methodology lacks a genuine mechanism for trial and error, which restricts the creative potential of machines and fails to capture the logical steps that actually lead to an answer. 1 Even AI systems equipped with so-called "reasoning" capabilities generate step-by-step reports that serve merely as another type of trace, often diverging from their actual internal computational logic. 1
The reliance on outcome-based learning has tangible consequences for human professionals, leading to significant skill degradation across various fields. 1 Historical warnings date back to 1997, when an American Airlines captain cautioned that pilots were becoming "children of the amber line," referring to their dependence on automated flight paths. 1 More recent evidence shows that doctors who utilized AI assistance during colonoscopies experienced a decline in their ability to independently detect polyps. 1 Similarly, students who relied on generative AI tools saw their mathematical performance drop once those tools were removed. 1 Research indicates that approximately ten minutes of AI assistance is sufficient to impair subsequent problem-solving performance and cause users to give up earlier when working without help. 1
To mitigate these issues, experts suggest integrating process-oriented learning into both AI development and educational design. 1 AI systems should be trained to learn from challenges and failures rather than just successful outputs, while educational tools must guide users through questioning rather than providing direct answers. 1 A Harvard University trial demonstrated the effectiveness of this approach: students tutored by a well-designed AI physics mentor learned faster and more effectively than those in traditional active-learning classrooms. 1 The success of the AI tutor stemmed from its ability to compel students to engage in generative work, thereby preserving the cognitive effort essential for deep learning. 1
评论
还没有评论,欢迎留下第一条。