作为卫斯理大学计算机科学系教员，我正与同事们一起应对人工智能对学生学习的影响。在我们系里，我们目前正在重组计算机科学专业，并起草课程中关于人工智能的指导方针。麻省理工学院的一份最新报告指出，虽然人工智能聊天机器人能轻松生成答案，但它们往往会引发“认知投降”。确实，在与学生的互动中，我意识到问题不再仅仅是外包作业，而是外包思考。
As a computer science faculty member at Wesleyan University, I am grappling alongside my colleagues with the impact of AI on student learning. In my department, we are currently restructuring the computer science major and drafting guidelines for AI in our curriculum. A recent report from MIT noted that while AI chatbots can easily produce answers, they often trigger “cognitive surrender.”Indeed, in my own interactions with students, I realize that the issue is no longer just offloading homework; it is offloading thinking.
这一发展令人担忧。深思熟虑的决策和个人自主是民主公民身份的基石。我们的民主依赖于能够评估论证、独立推理并让掌权者承担责任的公民。当我们让人工智能为我们做这些工作时，我们就冒着失去自治能力的风险。当前的人工智能模型被设计为阿谀奉承。它们迎合我们的偏见，同时悄然引导我们的观点。它们突显特定事实而压制其他事实，为国家控制制造了危险的切入口。
This development is concerning. Deliberate decision-making and personal autonomy are at the very foundations of democratic citizenship. Our democracy depends on citizens who can evaluate arguments, reason independently, and hold those in power accountable. When we let AI do this work for us, we risk our self-governance. Current AI models are engineered for sycophancy. They confirm our biases while quietly steering our perspectives. They highlight select facts and suppress others, creating a dangerous opening for state control.
在今年的隐私增强技术研讨会上，安全专家约翰·斯科特-雷尔顿描述了专制政府如何越来越多地胁迫科技公司采用“安全”框架，从而隐形地引导模型回复。这种操纵是破坞公民参与的有力工具。与传统审查直接禁止言论、公开屏蔽内容不同，算法引导在异见和公民辩论甚至还没有机会显现之前，就悄无声息地将其化解和平息。
At this year’s Privacy Enhancing Technologies Symposium, security expert John Scott-Railton described how autocratic governments are increasingly coercing tech companies into adopting “safety” frameworks that invisibly steer model responses. Such manipulation is a potent tool to undermine civic engagement. Unlike traditional censorship, which flatly prohibits speech and overtly blocks content, algorithmic steering quietly neutralizes dissent and pacifies civic debate before it even has a chance to manifest.
这种新型审查不禁止我们说话，而是阻止我们思考。诚然，广告商和政客操纵人类欲望由来已久。然而，生成式人工智能对民主审议构成了新威胁。因为我们与这些模型的互动仿佛它们是富有同理心、无所不知的知己，它们的说服力在我们意识不到的层面上发挥作用。
This new form of censorship does not forbid us to speak; it prevents us from thinking. To be sure, advertisers and politicians have manipulated human desires for ages. However, generative AI poses a new threat to democratic deliberation. Because we interact with these models as if they were empathetic and all-knowing confidants, their persuasion operates beneath our conscious recognition.
正如近期研究所示，AI 可以通过提供大量支持某一方的事实主张，改变人们对总统候选人和政策的态度。依我之见，保护我们的个人能动性需要内外兼修的应对。在内部层面，我们可以通过从“替代思维”转向“协作思维”来捍卫公民能动性：AI 并非定义我们人类判断力的那些活动——道德推理、批判性反思和创造性综合——的替代品。
As recent studies have shown, AI can move people’s attitudes towards presidential candidates and policies by providing many factual claims that support their side. In my view, protecting our personal agency requires an internal and external response. Internally, we can defend civic agency by moving from a replacement mindset to a collaborative mindset: AI is not a substitute for the activities that define our human judgment — moral reasoning, critical reflection, and creative synthesis.
相反，AI 应在这些事业中发挥互补作用。聊天机器人的输出不是最终定论，而应成为讨论和批判性探究的起点。若如此使用，AI 可以成为更多、更好的公民参与的有力工具。在外部层面，民主社会需要一套健全的监管框架，旨在保护公共话语免受隐蔽操纵。专制政权会试图强迫 AI 公司使输出结果与其意识形态保持一致，而民主政体必须致力于让 AI 提供真正的选择，而非家长式的引导。
Instead, AI should complement us in these endeavors. The output of a chatbot is not the final word but a starting point for discussion and critical inquiry. When used this way, AI can be a helpful tool for more and better civic engagement. Externally, democratic societies need a robust regulatory framework designed to protect public discourse from covert manipulation. While autocratic regimes will try to compel AI companies to align outputs with their ideology, democracies must aim for AI to provide genuine choice rather than paternalistic steering.
我们要求对汽车、药品和金融产品进行严格的安全测试和独立基准评估。AI 也需要同等力度的监管。我们的目标应是透明度和保障 AI 用户的认知独立性。AI 可以产生积极影响。当 AI 被设计为赋能而非说服时，它可以增强个人自主性和民主参与。AI 系统不应向特定方向推动我们的信念，而可作为学习和获取信息的工具。
We require rigorous safety testing and independent benchmarks for cars, pharmaceuticals and financial products. AI demands equal oversight. Our goals should be transparency and the safeguarding of AI users’ cognitive independence. AI can have a positive impact. When engineered to empower rather than to persuade, AI can strengthen personal autonomy and democratic participation. Instead of nudging our beliefs in a particular direction, AI systems can serve as tools to learn and inform.
它们能让复杂主题更易理解，真正实现个性化学习，并大幅拓展我们的知识、生产力和创造力。但要实现这一潜力，需要全社会进行深思熟虑的集体努力。工程师和用户界面设计师可以通过构建鼓励对话而非表演性附和的模型和界面来提供帮助。
They can make complex topics more easily understandable, truly personalize learning, and dramatically extend our knowledge, productivity, and creativity. But realizing this potential requires a deliberate, collective effort across society. Engineers and user interface designers can help by building models and interfaces that encourage dialogue rather than performative affirmation.
立法者和监管机构可以制定并执行相关规则，以确保信息的透明度，并禁止任何由政府或企业主导的、旨在影响公众观点或决策的行为。对于我们高等教育工作者来说，我们的责任是优先培养学生的AI素养（即让他们掌握与人工智能相关的知识和技能），并激发他们的独立思考能力。人工智能时代的真正挑战或许不在于制造出能够“思考”的机器，而在于防止我们自己停止思考。
Lawmakers and regulators can create and enforce transparency and outlaw state- or corporate-sponsored steering. For us in higher education, it is our responsibility to prioritize AI literacy and stimulate independent thought. The ultimate challenge of the AI era may not be to create machines that can “think,” but rather to prevent ourselves from stopping to think.
塞巴斯蒂安·齐梅克（Sebastian Zimmeck）是卫斯理大学（Wesleyan University）的计算机科学副教授，同时他还负责该校的隐私技术研究实验室（privacy-tech-lab）。
Sebastian Zimmeck is an associate professor of computer science at Wesleyan University, where he directs the privacy-tech-lab.