AI幻觉没法彻底消除。这不是技术不够好，是原理决定的。大模型就是个概率预测机器，每一步都在猜"下一个词最可能是什么"。它不知道的时候不会停下来，会猜一个听起来最合理的。所以别想着消灭幻觉，要想的是怎么把它关在笼子里，让它在你能看见、能验证的范围内活动。
下面五件事，从最好上手的开始说。
一、给它材料，别让它凭记忆答
你问AI公司内部制度、某篇论文的细节、某个产品的参数，它不知道就会编。不是它坏，是它被训练成了"必须给出答案"的样子。
办法很简单：把材料喂给它，让它照着你给的材料答。这叫开卷考试。你不让它闭卷考，它就没法瞎编。
有研究数据：把检索到的证据塞给模型，幻觉率能从51%降到20%。你不用懂技术，做重要的事时把相关文档、网页、数据一起丢给AI，再加一句："只根据我提供的材料回答，材料里没有的就说没有。"
这一句话比什么高级参数都管用。
二、温度调低，别让它太有创造力
你用的AI工具设置里要是有"Temperature"（温度），它管的是模型选词的随机程度。
温度高，模型爱冒险，挑不常见的词，看着有创意，也更容易胡编。温度低调保守，挑最稳妥、最常见的说法。
做事实性任务，温度调到0到0.3。有个反直觉的地方：温度调到0也消灭不了幻觉，只是让幻觉稳定复现——错了也每次都错得一模一样。好处是你排查方便，同一个问题多问几遍，答案稳定你又验证不了，那它多半是真不知道，在硬撑。
三、让它把推理过程写出来
你直接问"这个数据对吗"，它可能甩个答案给你就完事。你换个说法："一步一步说，你是怎么得出这个结论的"，效果完全不一样。
这叫思维链。道理不复杂：推理过程摊在纸面上，哪一步跳了、哪一步编了，你自己看得见。模型也没法从问题直接滑到一个编好的答案上，它得先走过中间那些步骤。
有实验显示，配合"自一致性"（让它换几种方式推几遍，看结果收不收敛），幻觉率能降10%到40%。你不用记这些术语，问关键问题时加一句"先别给答案，把推理过程列出来"就行。
四、先查再答，别先答再编
这招要多做一个动作：确认你用的AI开没开联网搜索，或者接没接知识库。
区别在哪：不联网的AI从训练数据里回忆事实，回忆不起来就编。训练数据是几个月甚至几年前的东西，它不知道的事比你想的多。联网或接知识库的AI先去查，查到什么说什么，查不到就承认。
所以问有明确事实答案的问题——某公司最新财报、某篇论文的结论、某个法规的条款——先看联网或检索开没开。没开，或者你不确定，就别信它给的数字和引用。
还有个细节：AI给了引用来源，也点开看一眼。研究发现AI经常"引用存在但支持不了那句话"——标了个链接，链接里的内容跟它说的不是一回事。引用不是护身符，引用里的原文才是。
五、接受它不知道
大多数人被幻觉坑，不是因为AI太笨，是因为人太懒。AI用流畅、自信、结构清晰的语气给个答案，人的大脑自动就把它当成核实过的信息了。但它没有。
真正管用的不在技术里，在习惯里：把AI的输出当初稿，不是终稿。事实性的东西自己核一遍。数字、引用、人名、日期、法条，这些是幻觉重灾区，一个都别跳过。
AI最危险的时候，不是它说"我不知道"，是它用最漂亮的排版给你一个错答案。你没法让AI不犯错，但可以让它犯的错在你这里过不去。
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。
Al's illusions can't be eliminated. It's not technical. It's theoretical. LLM is a probabilistic predictor machine, and every step of the way is wondering what the next word is. It won't stop when it doesn't know. It'll sound the most reasonable. So don't think about getting rid of the illusions, just how you want to keep it in a cage, so that it can operate within the limits that you can see and prove.
The next five things, starting with the best.
Give it the material. Don't let it answer with memory.
When you ask AI's internal system, the details of a paper, the parameters of a product, it doesn't know what to do. It's not bad, it's trained to be "must give the answer."
The solution is simple: feed it with the material and let it answer according to what you gave it. It's called an open exam. If you don't keep it closed, it won't make it up.
There are research data: the evidence retrieved can be plugged into the model and the hallucinogenic rate can be reduced from 51% to 20%. You don't have to know the technology, throw the files, the web pages, the data together to AI when you do important things, and add: "Only on the basis of the material I've provided, nothing in the material."
This sentence works better than any high-level parameter.
Two: Keep it down. Don't make it too creative.
If you're using "Temperature" in the AI tool setup, it's the randomity of a model word.
High temperatures, models are risky, pick unusual words, look creative and make it easier to make up. Temperatures are low-key and conservative, choosing the most conservative and common term.
Factual mission, temperature to 0 to 0.3. There's a counterintuitive: the temperature goes to zero and it doesn't eliminate the illusion, it just stabilizes it -- it's the same thing every time. The good thing is you're easy to check, ask the same question a few times, and the answer is stable and you can't prove it.
iii. Let it write the reasoning.
You just ask, "Is this the right number?" It's probably done with an answer. Say it again: "How did you come to that conclusion, step by step?" The effect is completely different.
It's called the thought chain. It's not complicated: the reasoning process is on paper, and you can see what's going on and what's going on. And the model can't just slide from the question to a complete answer, and it has to go through the middle.
There are experiments that show that, in combination with "self-conformity" (which can be pushed several times and the results do not fall), hallucinogenic rates can be reduced by 10 to 40 per cent. You don't have to remember those terms, and you just have to add to the questions "Don't give the answer first and list the reasoning."
Four, check and answer. Don't answer and edit.
It's an extra move: make sure you're using the AI or not to open a search or not to access the knowledge base.
The difference is: the unconnected AI remembers the facts from the training data and does not. Training data are for months or even years, and it knows more than you think. AI, which is connected or connected to the knowledge base, investigates what it says and acknowledges it if it is not found.
So questions with clear, factual answers — the company's latest financial report, the conclusions of a paper, the terms of a regulation — go first to the network or search. If you don't, or if you're not sure, don't believe the numbers and quotes it gives.
There's another detail: AI gave the source of the quote, and it's a glimpse. The study found that AI often "quote the existence of the sentence but cannot support it" -- a link that is not exactly what it says. The reference is not the amulet, but the original.
V. ACCEPTING IT DON'T KNOW
Most people are trapped in hallucinations, not because AI is stupid, but because people are lazy. AI gives an answer in a fluid, self-confident and structured tone, and the human brain automatically sees it as verified information. But it didn't.
It really doesn't work in technology, it's in habit: it's not a final draft. The truth thing itself checks out. Figures, quotations, names, dates, laws, these are hallucinating areas, not one of them.
The most dangerous time for AI is not when it says "I don't know," it gives you the wrong answer in the most beautiful layout. You can't keep AI from making mistakes, but you can keep it from making mistakes.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.