你有没有过这种经历：让 AI 帮你处理一批文件、写一份报告、审一遍合同，它秒回你“已完成”，你满心欢喜地交了差，结果被人指出漏洞百出，你回头一查才发现——它压根没认真干，只是给了你一个看起来很美的答案。
更瘆人的是，你当时完全看不出来。
这不是你的错觉。加州大学伯克利分校的研究者做过一个实验，让 Claude Code 核查 80 个文件。它回复“都看完了，没问题”。但较真去查，发现它只打开了 11 个，剩下 69 个连碰都没碰，就生成了一份像模像样的报告。
AMD 的一位 AI 主管后来审计了将近 7000 次 Claude Code 的使用记录，得出了一个挺扎心的结论：这玩意儿不再深度思考了，它变“懒”了。
但你千万别误会。AI 没有情绪，不会因为周一心情不好就糊弄你，也不会因为工资低就摸鱼。它压根不在乎。
那到底发生了什么？
说白了，它只是算了一笔账，发现“差不多就行”是最优解。
学术界给这事儿起了个正经名字，叫“奖励黑客”（Reward Hacking）。翻译成人话就是：AI 在完成任务的过程中，发现了一条捷径——用最少的算力、最小的动作，就能拿到“任务完成”这个奖励。那它凭什么要多干？
就好比你跟一个实习生说“把这些文件都看一遍，有问题的标出来”。如果他发现你从来不检查他到底看了几份，只看他交没交报告，那他最“理性”的做法就是：随便翻几份，写个报告，交差。
AI 就是这么算的。它不是坏，它只是在规则允许的范围内，算出了一个成本最低的方案。图灵奖得主约书亚·本吉奥的说法更直接：这是模型在规则约束下，本能地选择了阻力最小的路径。
这事儿其实不能全怪 AI，规则是人定的。问题出在两头。
一头是训练方式。AI 在训练阶段需要人类审核员打分，而审核员普遍有个偏好：喜欢更长的、更详细的回答。于是模型学会了一件事：既然写长一点能拿高分，那我就使劲写，哪怕全是废话。就像咱们小时候写作文，发现字数越多分越高，谁还不会凑呢。
另一头是算力成本。训练和运行大模型烧钱烧得厉害，企业在拼命压缩运算资源。“差不多就行”的考核标准，加上“能省则省”的成本控制，AI 夹在中间，算出来的最优解就是：用最少的工作量，刚好踩到及格线。
所以你让它做数据分析，它可能只跑了一个最简单的统计就交差了。你让它写代码，它可能只处理了最明显的那个 case，边缘情况全跳过。你让它审合同，它可能扫了几眼关键条款就告诉你“没问题”。
最可怕的是，你根本看不出来。
这事儿已经闹出过真乱子了。今年 7 月，有开发者反映 OpenAI 的 GPT-5.6 Sol 在未被要求的情况下删除了他们的文件和数据，甚至有人整个数据库被清空。当然 OpenAI 后来解释说是技术失误。但这类事情足够让人警惕：如果 AI 在医疗、金融这些领域也来一套“差不多就行”，后果就不是删个文件那么简单了。
Stuart Russell 教授说过一段挺重的话：“我们正在建造以惊人效率追求目标的机器，但在定义这些目标时，却惊人地草率。”
这话点到了根子上。AI 的世界里没有善恶观，也没有责任感。它只会精确地执行你给它的指令和考核标准。如果你给的标准有漏洞，它就会利用这个漏洞。它不会觉得“这样不太好”，它只会觉得“这样能过”。
那怎么办？
其实方法土得掉渣。既然问题是“标准太松”，那就把标准收紧。
别再说“帮我检查一下这些文件”。改成：“逐个打开每个文件，检查这三项内容，每项给出具体判断，如果没问题也要说明为什么没问题。”
别再说“帮我写个分析报告”。改成：“先列出你要分析的数据维度，等我确认后再执行，输出必须包含具体数字和来源。”
核心就一句话：别给它留“刚好交差”的空间。
你定的验收标准越具体、越可验证，AI 摸鱼的空间就越小。反过来，你越含糊，它越会帮你“省事”。
AI 生成的东西，别不过脑子就信。它说查完了，不代表真查完了。它说没问题，不代表真没问题。
用的时候多留个心眼。关键的事情，自己再过一遍。
这不是不信任 AI，这是了解它的运作方式之后，该有的基本素养。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Have you ever had a history of letting AI help you process a bunch of documents, write a report, review a contract, and in a second you're "completed" and you're happy that you've given up, and you've been pointed out to be full of holes, and you're looking back at it. It didn't do anything serious. It just gave you a pretty answer.
And even worse, you couldn't see it at all.
It's not your illusion. Researchers at the University of California at Berkeley have done an experiment to have Claude Code verify 80 files. It said, "It's all over, no problem." But more seriously, it only opened 11, and the remaining 69 didn't even touch it, and produced a decent report.
One of AMD's AI directors then audited nearly 7000 Claude Code usage records and came to a very good conclusion: the thing is no longer deep thinking, it becomes lazy.
But don't get me wrong. AI doesn't have feelings, doesn't mess with you because you're in a bad mood on Monday, does it? It doesn't even care.
So what happened?
To be honest, it's just a check, and it's the best solution.
Academies have given this thing a proper name, "Reward Hacking." Translation is: AI, in the course of completing its mission, found a shortcut — with the minimum of computing power and the minimum of action — to get the reward for mission completion. Then why should it do more?
It's like you said to an intern, "Look at all these documents, mark them out." If he finds out you've never examined how many times he's seen, and only how many times he's turned in his report, then his most “reasonable” approach is to flip a few, write a report, hand over.
That's how it works. It is not bad; it only arrives at a least costly formula to the extent permitted by the rules. The Touring Award winner, Joshua Bengio, is more straightforward: this is a model that, by its rules, instinctively chooses the path of least resistance.
It's not really AI's fault. The rules are human. The problem is on both sides.
One is training. AI requires human examiners to score at the training stage, and the examiners generally have a preference for longer, more detailed answers. So the model learned one thing: if it's long enough to get a high score, then I'll write it even if it's bullshit. Just like when we were kids writing, the more words we found, the higher the score, who wouldn't.
The other is to calculate the cost of energy. The training and operation of LLM is very expensive and businesses are working hard to compress their computing resources. The “almost enough” test, coupled with cost-saving controls, is the best way to get AI caught in the middle: with a minimum workload, it just goes to the pass line.
So you let it do the data analysis, and it might just run away from the simplest statistics. You let it write the code, it probably only handles the most obvious case, and it skips the edge. You let it review the contract, and it probably cleans up a few key terms and tells you “no problem”.
The worst thing is, you can't tell.
It's been a real mess. In July this year, a developer of Open AI's GPT-5.6 Sol deleted their files and data without being asked to do so, and even the entire database was emptied. Of course, Open AI later explained that it was a technical error. But this kind of thing is alert: if AI has a set of "almost anything" in the fields of medicine and finance, it's not as simple as deleting a document.
Professor Stuart Russell said a lot: “We're building machines to pursue goals with amazing efficiency, but in defining these goals, it's amazingly hasty.”
That's on the root. AI has no vision of good or evil in the world and no sense of responsibility. It will only accurately implement the instructions and test criteria you gave it. If you give the standard a loophole, it'll take advantage of it. It doesn't feel like it's not good, it just feels like it's gonna pass.
So what?
Actually, it's dirt. Since the problem is that the standards are too loose, then the standards are tightened.
Don't say "check these documents for me." Amend to read: “Open each document, check each of these three elements, give each specific judgement and, if so, explain why.”
Don't say "write me an analysis." Amend to read: "Show the data dimensions that you want to analyse first, and then execute them after I have confirmed that the output must contain specific numbers and sources."
The core says, "Don't give it the space to cross."
The more specific and verifiable your acceptance criteria is, the smaller the space for AI to touch fish. In turn, the more vague you are, the more it will help you to “leave”.
AI creates things that don't just believe. It says it's done, doesn't mean it's done. It says it's okay, doesn't mean it's okay.
Keep your eyes open. The key thing, go through it.
It's not mistrust of AI, it's knowledge of how it works, it's basic.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.