我越来越常听到人们问我:“AI 会损伤你的大脑吗?” 这是一个引人深思的问题。不是因为 AI 真的会造成器质性脑损伤(它不会),而是因为这个问题本身揭示了我们对 AI 可能如何影响我们思考能力的深切恐惧。因此,在这篇文章中,我想探讨如何利用 AI 来帮助而非损害你的心智。但为什么人们如此执着于 AI 会损伤我们大脑这件事呢?
部分原因在于对一篇备受关注的论文的误读。这篇论文出自麻省理工学院媒体实验室(作者也来自其他机构),标题是《ChatGPT 下的你》。实际的研究远没有媒体报道的那么戏剧化。该研究让一小群大学生分别独立写作、使用谷歌写作或使用 ChatGPT(且无其他工具)来完成论文。与未使用 AI 的小组相比,使用 ChatGPT 的学生投入度更低,对论文内容的记忆也更少。四个月后,其中九名 ChatGPT 用户被要求在不使用 ChatGPT 的情况下再次撰写同一篇论文,他们的表现比最初未使用 AI 的小组(尽管在新实验中也被要求使用 AI)更差,并且在写作时脑电图活动也更少。当然,这并非脑损伤。然而,更具戏剧性的解读之所以抓住了我们的想象力,是因为我们一直担心新技术会毁掉我们的思考能力:柏拉图认为文字会削弱我们的智慧;手机问世时,也有人担心不必记住电话号码会让我们变得更笨。
但这并不意味着我们不应该担心 AI 如何影响我们的思考。毕竟,技术的一个关键目的是让我们将工作外包给机器。这包括智力工作,比如让计算器做数学题,或者让手机记录电话号码。而当我们外包思考时,我们确实会失去一些东西——例如,我们实际上无法像以前那样记住电话号码。鉴于 AI 是一种如此通用的智力技术,我们可以将大量思考外包给它。那么,我们该如何利用 AI 来帮助而非损害我们呢?
学习中的大脑
人工智能使用可能明显损害你心智成长的最不令人意外的场景,就是当你试图学习或整合新知识的时候。如果你把思考外包给AI,而不是自己完成这项工作,那么你就会错失学习的机会。我们有证据支持这一直觉——我在宾夕法尼亚大学的同事在土耳其一所高中进行了一项实验,部分学生获准使用GPT-4辅助完成作业。当这些学生被要求在没有指导或特殊提示的情况下使用ChatGPT时,他们最终走了捷径,直接获取答案。因此,尽管学生们自认为从ChatGPT的帮助中学到了很多,实际上他们学到的更少——在期末考试中得分比未使用ChatGPT的学生低了17%。
这种做法之所以尤其隐蔽,是因为即使学生怀着良好的意图,伤害依然会发生。AI被训练成乐于助人、为你解答问题。就像那些学生一样,你可能只是想获得AI关于如何着手做作业的指导,但它往往直接给你答案。正如MIT媒体实验室的研究所示,这绕过了(有时令人不快的)产生学习效果的心智努力。问题不仅仅在于作弊——尽管AI确实让作弊变得更容易——问题在于,即使是诚心诚意地尝试用AI寻求帮助,也可能适得其反,因为AI的默认模式是替你完成工作,而不是与你一起完成。

这是否意味着人工智能总是会损害学习?绝非如此!虽然目前仍处于早期阶段,但我们有越来越多的证据表明,在教师的指导下,并基于可靠的教学原则使用良好的提示词,人工智能可以极大地改善学习成果。例如,世界银行的一项随机对照研究发现,在尼日利亚的一个为期六周的课后项目中,使用基于 GPT-4 的辅导工具并在教师指导下进行,其效果“是教育领域一些最有效干预措施的两倍以上”,且成本非常低。尽管没有哪项研究是完美的(在本例中,对照组完全没有进行任何干预,因此无法完全分离出人工智能的独立影响,尽管他们确实尝试了这样做),但这项研究加入了越来越多得出类似结论的行列。哈佛大学在一门大型物理课上进行的实验发现,经过良好提示词引导的人工智能辅导工具在学习成果上优于活跃的课堂;斯坦福大学在一门大型编程课上开展的研究表明,使用 ChatGPT 提高了考试成绩;马来西亚的一项研究发现,将人工智能与教师指导和扎实的教学法结合使用,能带来更多的学习效果;甚至我之前提到的土耳其实验也发现,更好的辅导提示词消除了单纯使用 ChatGPT 所导致的考试成绩下降。

归根结底,决定使用人工智能在学习时是有益还是有害大脑的,并非是否使用AI,而是如何使用AI。从让AI帮你做作业,转向让它像导师一样帮助你学习,这是有用的一步。不幸的是,大多数AI模型的默认模式是直接给你答案,而不是就某个主题对你进行辅导,因此你可能需要使用专门的提示词。虽然还没有人开发出完美的导师提示词,但我们有一个已在一些教育研究中使用的提示词,可能对你有用,你可以在沃顿生成式AI实验室的提示词库中找到更多。请随意修改它(它采用知识共享许可协议)。如果你是家长,你也可以自己充当导师,向AI提示:“用我能教给我X年级孩子的方式解释这道题的答案。”这些方法都不是完美的,AI在教育领域带来的挑战非常真实,但我们有理由相信,教育将能够以有助于而非损害我们思考能力的方式适应AI。这将需要教师的指导、精心构建的提示词,以及在使用AI和避免使用AI之间做出谨慎的选择。
创意大脑
就像在教育领域一样,AI 对你的创造力可能有益,也可能有害,具体取决于你如何使用它。在许多创造力衡量指标上,AI 超越了大多数人类。需要明确的是,创造力并没有统一的定义,但研究人员开发了多种存在缺陷的测试方法,这些方法被广泛用于衡量人类提出多样且有意义的想法的能力。这些测试存在缺陷,原本不是什么大问题,直到突然之间,AI 能够通过所有这些测试。在创造力替代用途测试的一个变体中,旧版 GPT-4 击败了 91% 的人类,并在托伦斯创造性思维测试中超过了 99% 的人。而且我们知道,这些想法并非只有理论上的趣味性。我在沃顿商学院的同事们组织了一场创意生成竞赛:让 ChatGPT-4 与一个热门创新课程的学生们对决,该课程历来催生了许多初创公司。由人类评委对创意进行评分,结果显示 ChatGPT-4 比学生们生成了更多、更便宜且更好的创意。这些外部评委对 AI 生成创意的购买意向也更高。
然而,任何使用过 AI 进行创意生成的人都会注意到这些数字未能捕捉到的一点。AI 往往表现得像一个具有可预测模式的单一创意个体。你会反复看到相同的主题,比如涉及 VR、区块链、环境以及(当然)AI 本身的创意。这是一个问题,因为在创意生成中,你实际上需要一组多样化的创意以供选择,而不是某个主题的变体。因此,存在一个悖论:虽然 AI 比大多数个体更具创造力,但它缺乏来自多元视角的多样性。然而,研究也表明,人们在使用 AI 时通常比独自工作时能产生更好的创意,有时甚至 AI 单独工作也能胜过与 AI 协作的人类。但如果不加谨慎,当你看到足够多的这些创意时,它们彼此之间会显得非常相似。
这个问题部分可以通过更好的提示词来解决。在我与 Lennart Meincke 和 Christian Terwiesch 合作的一篇论文中,我们发现,更好的提示词可以生成更加多样化的创意,尽管其效果仍略逊于一个学生小组。
以下是针对 GPT-4 设计的提示词。它同样适用于其他 AI 模型(不过我怀疑,推理型模型在创新性上可能比传统模型略逊一筹):
Generate new product ideas with the following requirements: The product will target [market or customer]. It should be a [pick: physical good/service/software], not a [pick: physical good/service/software]. I'd like a product that could be sold at a retail price of less than about [insert amount].
The ideas are just ideas. The product need not yet exist, nor may it necessarily be clearly feasible. Follow these steps. Do each step, even if you think you do not need to. First generate a list of 100 ideas (short title only). Second, go through the list and determine whether the ideas are different and bold, modify the ideas as needed to make them bolder and more different. No two ideas should be the same. This is important! Next, give the ideas a name and combine it with a product description. The name and idea are separated by a colon and followed by a description. The idea should be expressed as a paragraph of 40-80 words. Do this step by step. 但更好的提示词只能解决部分问题。更深层的风险在于,AI 实际上会通过将你锚定在其建议上,从而损害你的创造性思维能力。这种情况通过两种方式发生。
首先是锚定效应。一旦你看到 AI 的想法,就很难跳出这些边界进行思考。就像有人告诉你“别想粉红色大象”一样。AI 的建议,即使平庸,也会挤占你自己独特的视角。其次,正如麻省理工学院那项研究所显示的,人们对 AI 生成的想法不会有太强的归属感,这意味着你会从构思过程本身抽离出来。
那么,如何在不导致思维枯竭的情况下获得 AI 的好处呢?关键在于顺序。在求助于 AI 之前,一定要先自己产生想法。把它们写下来,无论多么粗糙。就像集体头脑风暴在每个人先独立思考时效果最好一样,你需要在 AI 的建议锚定你之前,先捕捉自己独特的视角。然后再利用 AI 将想法推进一步:“以极端方式结合想法 #3 和 #7”,“再极端一点”,“再给我 10 个类似 #42 的想法”,“以超级英雄为灵感,让这个想法更有趣。”
这一原则在写作中变得更为关键。许多作家坚持认为“写作即思考”,虽然这并非普遍真理(如果你想要细节,我倒是就这个话题生成了一份相当不错的深度研究报告),但通常确实如此。写作、重写、再重写的行为,能帮助你梳理和打磨自己的想法。如果你让 AI 代笔写作,你就完全跳过了思考的部分。
对于我这样以写作为思考方式的人来说,必须变得自律。我写的每一篇文章,包括这一篇,都会完全不用任何 AI(除了研究辅助)先完成一版完整草稿。这通常是一个漫长的过程,因为我会反复撰写和修改——这就是思考!只有当草稿完成后,我才会转向多个 AI 模型,把写好的文章交给它们,让它们扮演读者:有没有什么地方表述不清?具体如何修改才能让非技术读者更容易理解?有时还会让它们扮演编辑:我不喜欢这个段落的结尾,你能给我 20 个可能更合适的结尾版本吗?所以,放心使用 AI 来打磨你的文字、拓展你的可能性吧。只是要记住,先完成思考,因为这部分是无法外包的。

集体大脑
AI 可能损害我们思维的另一个领域,是通过影响社会互动过程。理想情况下,团队协作的全部意义在于提升绩效——团队应该能产生更多想法,更善于发现潜在机遇和陷阱,并提供专业技能和能力来帮助执行。会议应该是团队协调和解决问题的场所。当然,这是理想状态。实际上,最具启发性的管理学读物之一,其实是这本二战时期美国中央情报局前身为平民编写的破坏活动指南。看看其中关于破坏办公室任务以造成士气低落和延误的那些点子,再想想有多少已经成了你日常会议中的常态。
因此,AI 早期的一个重要用途就是总结会议,并且越来越多地用于总结你完全缺席的会议,这毫不奇怪。当然,这引发了一些存在主义问题,比如“如果我们只需要读一份摘要,那为什么还要开会?”或者“我是不是应该派一个 AI 替身去开会?”显然,在一个所有人都只是来读会议记录、别无其他的会议里,没有互动,没有团队协作,没有思想碰撞。这只是在消耗时间和精力,是一种组织层面的脑损伤。
但与其说人工智能损害了我们的集体思维,不如说我们有机会让它帮助我们变得更好。一个有趣的例子是让AI充当引导者。我们设计了一个提示词,让AI扮演引导者的角色,在会议进行到一半时创建定制的塔罗牌,以引导而非取代你们的讨论。你向它提供会议记录,它就能帮助你激发出最好的想法(再次说明,这是知识共享许可,你可以根据需要修改;目前它在Claude上效果最佳,在Gemini和o3上表现尚可)。
这只是一个有趣的例子,展示了AI可用于提升我们的集体智慧,但我们还需要进行更多实验来找出有效的方法:让AI扮演唱反调的角色,揭示未被言明的担忧;让它识别讨论中哪些人的声音未被听到;或者用它来发现人类在团队动态中忽略的模式。关键在于,AI是增强而非取代人际互动。
反对“脑损伤”
AI不会损伤我们的大脑,但不加思考地使用会损害我们的思维能力。关键不在于我们的神经元,而在于我们的思维习惯。有很多工作值得用AI来自动化或取代(我们很少为用计算器做数学题而感到惋惜),但也有大量工作,我们的思考至关重要。对于这些问题,研究给出了明确的答案。如果你想保留工作中属于人的部分:先思考,先写作,先开会。
我们担心AI会“损伤我们的大脑”,实际上是对自身懒惰的恐惧。这项技术为艰难的思考工作提供了一条捷径,我们担心自己会走上这条路。我们确实应该担心。但我们也应该记住,我们是有选择的。
你的大脑是安全的。然而,你的思考,取决于你自己。
I increasingly find people asking me “does AI damage your brain?” It's a revealing question. Not because AI causes literal brain damage (it doesn't) but because the question itself shows how deeply we fear what AI might do to our ability to think. So, in this post, I want to discuss ways of using AI to help, rather than hurt, your mind. But why the obsession over AI damaging our brains?
Part of this is due to misinterpretation of a much-publicized paper out of the MIT Media Lab (with authors from other institutions as well), titled “Your Brain on ChatGPT.” The actual study is much less dramatic than the press coverage. It involved a small group of college students who were assigned to write essays alone, with Google, or with ChatGPT (and no other tools). The students who used ChatGPT were less engaged and remembered less about their essays than the group without AI. Four months later, nine of the ChatGPT users were asked to write the essay again without ChatGPT, and they performed worse than those who had not used AI initially (though were required to use AI in the new experiment) and showed less EEG activity when writing. There was, of course, no brain damage. Yet the more dramatic interpretation has captured our imagination because we have always feared that new technologies would ruin our ability to think: Plato thought writing would undermine our wisdom, and when cellphones came out, some people worried that not having to remember telephone numbers would make us dumber.
But that doesn’t mean we shouldn’t worry about how AI impacts our thinking. After all, a key purpose of technology is to let us outsource work to machines. That includes intellectual work, like letting calculators do math or our cellphones record our phone numbers. And, when we outsource our thinking, we really do lose something — we can’t actually remember phone numbers as well, for example. Given that AI is such a general purpose intellectual technology, we can outsource a lot of our thinking to it. So how do we use AI to help, rather than hurt us?
The Learning Brain
The least surprising place where AI use can clearly hurt your mental growth is when you are trying to learn or synthesize new knowledge. If you outsource your thinking to the AI instead of doing the work yourself, then you will miss the opportunity to learn. We have evidence to back up this intuition, as my colleagues at Penn conducted an experiment at a high school in Turkey where some students were given access to GPT-4 to help with homework. When they were told to use ChatGPT without guidance or special prompting, they ended up taking a shortcut and getting answers. So even though students thought they learned a lot from ChatGPT's help, they actually learned less - scoring 17% worse on their final exam (compared to students who didn't use ChatGPT).
What makes this particularly insidious is that the harm happens even when students have good intentions. The AI is trained to be helpful and answer questions for you. Like the students, you may just want to get AI guidance on how to approach your homework, but it will often just give you the answer instead. As the MIT Media Lab study showed, this short-circuits the (sometimes unpleasant) mental effort that creates learning. The problem is not just cheating, though AI certainly makes that easier. The problem is that even honest attempts to use AI for help can backfire because the default mode of AI is to do the work for you, not with you.

Does that mean that AI always hurts learning? Not at all! While it is still early, we have increasing evidence that, when used with teacher guidance and good prompting based on sound pedagogical principles, AI can greatly improve learning outcomes. For example, a randomized, controlled World Bank study finds using a GPT-4 tutor with teacher guidance in a six week after school program in Nigeria had "more than twice the effect of some of the most effective interventions in education" at very low costs. While no study is perfect (in this case, the control was no intervention at all, so it is impossible to fully isolate the effects of AI, though they do try to do so), it joins a growing number of similar findings. A Harvard experiment in a large physics class found a well-prompted AI tutor outperformed active classes in learning outcomes; a study done in a massive programming class at Stanford found use of ChatGPT led to increased exam grades; a Malaysian study found AI used in conjunction with teacher guidance and solid pedagogy led to more learning; and even the experiment in Turkey that I mentioned earlier found that a better tutor prompt eliminated the drop in test scores from plain ChatGPT use.

Ultimately, it is how you use AI, rather than use of AI at all, that determines whether it helps or hurts your brain when learning. Moving away from asking the AI to help you with homework to helping you learn as a tutor is a useful step. Unfortunately, the default version of most AI models wants to give you the answer, rather than tutor you on a topic, so you might want to use a specialized prompt. While no one has developed the perfect tutor prompt, we have one that has been used in some education studies, and which may be useful to you and you can find more in the Wharton Generative AI Lab prompt library. Feel free to modify it (it is licensed under Creative Commons). If you are a parent, you can also act as the tutor yourself, prompting the AI “explain the answer to this question in a way I can teach my child, who is in X grade.” None of these approaches are perfect, and the challenges in education from AI are very real, but there is reason to hope that education will be able to adjust to AI in ways that help, and not hurt, our ability to think. That will involve instructor guidance, well-built prompts, and careful choices about when to use AI and when it should be avoided.
The Creative Brain
Just like in education, AI can help, or hurt, your creativity depending on how you use it. On many measures of creativity, AI beats most humans. To be clear, there is no one definition of creativity, but researchers have developed a number of flawed tests that are widely used to measure the ability of humans to come up with diverse and meaningful ideas. The fact that these tests were flawed wasn't that big a deal until, suddenly, AIs were able to pass all of them. The old GPT-4 beat 91% of humans on the a variation of the Alternative Uses Test for creativity and exceeds 99% of people on the Torrance Tests of Creative Thinking. And we know these ideas are not just theoretically interesting. My colleagues at Wharton staged an idea generation contest: pitting ChatGPT-4 against the students in a popular innovation class that has historically led to many startups. Human judges rating the ideas showed that that ChatGPT-4 generated more, cheaper and better ideas than the students. The purchase intent from these outside judges was higher for the AI-generated ideas as well.
And yet, anyone who has used AI for idea generation will notice something these numbers don't capture. AI tends to act like a single creative person with predictable patterns. You'll see the same themes over and over like ideas involving VR, blockchain, the environment, and (of course) AI itself. This is a problem because in idea generation, you actually want a diverse set of ideas to pick from, not variations on a theme. Thus, there is a paradox: while AI is more creative than most individuals, it lacks the diversity that comes from multiple perspectives. Yet studies also show that people often generate better ideas when using AI than when working alone, and sometimes AI alone even outperforms humans working with AI. But, without caution, those ideas look very similar to each other when you see enough of them.
Part of this can be solved with better prompting. In a paper I worked on with Lennart Meincke and Christian Terwiesch, we found that better prompting can generate much more diverse ideas, if not quite as good as a group of students.
Here is the prompt, which was for GPT-4. It still works well for other AI models (though I suspect that reasoner models might actually be slightly less innovative than more traditional models):
Generate new product ideas with the following requirements: The product will target [market or customer]. It should be a [pick: physical good/service/software], not a [pick: physical good/service/software]. I'd like a product that could be sold at a retail price of less than about [insert amount].
The ideas are just ideas. The product need not yet exist, nor may it necessarily be clearly feasible. Follow these steps. Do each step, even if you think you do not need to. First generate a list of 100 ideas (short title only). Second, go through the list and determine whether the ideas are different and bold, modify the ideas as needed to make them bolder and more different. No two ideas should be the same. This is important! Next, give the ideas a name and combine it with a product description. The name and idea are separated by a colon and followed by a description. The idea should be expressed as a paragraph of 40-80 words. Do this step by step. But better prompting only solves part of the problem. The deeper risk is that AI can actually hurt your ability to think creatively by anchoring you to its suggestions. This happens in two ways.
First, there's the anchoring effect. Once you see AI's ideas, it becomes much harder to think outside those boundaries. It's like when someone tells you “don't think of a pink elephant.” AI's suggestions, even mediocre ones, can crowd out your own unique perspectives. Second, as the MIT study showed, people don’t feel as much ownership in AI generated ideas, meaning that you will disengage from the ideation process itself.
So how do you get AI's benefits without the brain drain? The key is sequencing. Always generate your own ideas before turning to AI. Write them down, no matter how rough. Just as group brainstorming works best when people think individually first, you need to capture your unique perspective before AI's suggestions can anchor you. Then use AI to push ideas further: “Combine ideas #3 and #7 in an extreme way,” “Even more extreme,” “Give me 10 more ideas like #42,” “User superheroes as inspiration to make the idea even more interesting.”
This principle becomes even more critical in writing. Many writers insist that "writing is thinking," and while this isn't universally true (I generated a pretty good Deep Research report on the topic if you want the details), it often is. The act of writing, and rewriting, and rewriting again helps you think through and hone your ideas. If you let AI handle your writing, you skip the thinking part entirely.
As someone for whom writing is thinking, I've needed to become disciplined. Every post I write, like this one, I do a full draft entirely without any AI use at all (beyond research help). This is often a long process, since I write and rewrite multiple times - thinking! Only when it is done do I turn to a number of AI models and give it the completed post and ask it to act as a reader: Was this unclear at any point, and how, specifically could I clarify the text for a non-technical reader? And sometime like an editor: I don’t like how this section ends, can you give me 20 versions of endings that might fit better. So go ahead, use AI to polish your prose and expand your possibilities. Just remember to do the thinking first, because that's the part that can't be outsourced.

The Collective Brain
Another area where AI can hurt our thinking is through its impact on social processes. Ideally, the whole purpose of working on teams is that it can improve our performance - teams should be able to generate more ideas, be better able to see potential opportunities and pitfalls, and provide specialized skills and abilities to help execution. Meetings should be places where teams coordinate and solve problems. Of course, this is the ideal. In reality, one of the most revelatory management texts is actually this WWII guide to sabotage for civilians from the CIA's precursor. Look at the ideas for sabotaging office tasks to cause demoralization and delay and consider how many of them are normal parts of your meetings.
So it is no wonder that a significant early use of AI is to summarize meetings, and increasingly to summarize meetings you skip entirely. Of course, this raises existential questions like “why are we meeting in the first place if we can just read a summary?” or “should I just send an AI avatar of myself to meetings?” Obviously, there is no interaction, no teamwork, no meeting of the minds in a meeting where everyone is just there to read the transcript and nothing more. It just takes up time and effort, a form of organizational brain damage.
But rather than AI hurting our collective thinking, there is the option to have it help make us better. One interesting example is using AI as a facilitator. We created a prompt where AI acts as facilitator, creating customized tarot cards halfway through your meeting to help guide, rather than replace, your discussion. You give it a meeting transcript and it helps you bring out your best ideas (again, this is a Creative Commons license, so modify as needed, right now it works best on Claude, and okay on Gemini and o3)
This is just a fun example of the ways in which AI could be used to help our collective intelligence, but there is a need for many more experiments to figure out what works: using AI as a devil's advocate to surface unspoken concerns, having it identify whose voices aren't being heard in a discussion, or using it to find patterns in team dynamics that humans miss. The key is that AI enhances rather than replaces human interaction.
Against “Brain Damage”
AI doesn't damage our brains, but unthinking use can damage our thinking. What's at stake isn't our neurons but our habits of mind. There is plenty of work worth automating or replacing with AI (we rarely mourn the math we do with calculators), but also a lot of work where our thinking is important. For these problems, the research gives us a clear answer. If you want to keep the human part of your work: think first, write first, meet first.
Our fear of AI “damaging our brains” is actually a fear of our own laziness. The technology offers an easy out from the hard work of thinking, and we worry we'll take it. We should worry. But we should also remember that we have a choice.
Your brain is safe. Your thinking, however, is up to you.