批判性思维与信息鉴别
Critical Thinking & Information Literacy
答案无限廉价,判断答案好坏的能力反而稀缺。
Answers are cheap; judging them is the rare skill.
AI 时代,为什么批判性思维比以往更重要
Why Critical Thinking Matters More Than Ever in the AI Era
批判性思维(Critical Thinking)= 不盲从、不轻信、能验证、能推理、能下结论的能力。它不是『抬杠』,而是『对信息保持健康的怀疑 + 用方法去验证』。在 AI 时代,它分为 3 层:① 判断信息来源(谁说的、可信吗);② 判断信息内容(逻辑通吗、证据够吗);③ 判断 AI 输出(AI 说的对吗、哪里错了)。
Critical Thinking = the ability to not follow blindly, not trust easily, verify, reason, and conclude. It's not 'arguing', but 'keeping healthy skepticism about info + using methods to verify'. In the AI era, it has 3 layers: ① judge the source (who said it, is it credible); ② judge the content (is the logic sound, is evidence enough); ③ judge AI output (is AI right, where is it wrong).
为什么 AI 时代更需要它?3 个原因:① AI 会『一本正经地胡说』(幻觉),你无法从语气判断真假;② AI 生成内容成本趋近于零,虚假信息可以规模化生产;③ AI 推荐系统让你陷入『信息茧房』,只看到你想看的。批判性思维是你在 AI 时代『不被带节奏』的底层能力。
Why more needed in the AI era? 3 reasons: ① AI 'speaks nonsense confidently' (hallucination); you can't judge truth by tone; ② AI content generation cost approaches zero, false info can be mass-produced; ③ AI recommendation traps you in 'info cocoons', only showing what you want to see. Critical thinking is the foundational skill to 'not be led around' in the AI era.
6 大核心概念:批判性思维的底层能力
6 Core Concepts: The Foundation of Critical Thinking
| 概念 | Concept | 是什么 | What it is | 为什么重要 | Why it matters | 怎么练 | How to train |
|---|---|---|---|---|---|---|---|
| 怀疑(Skepticism) | 怀疑(Skepticism) | 对任何『未验证』的信息保持健康怀疑,不轻信 | Keep healthy skepticism for any 'unverified' info, don't trust easily | 怀疑是批判性思维的起点:不怀疑,就不会去验证 | Skepticism is the starting point of critical thinking: without it, you won't verify | 每天问『这是真的吗?怎么验证?』 | Ask daily: 'Is this true? How to verify?' |
| 验证(Verification) | 验证(Verification) | 用方法去核实信息的真伪:交叉信源、查原始出处、试逻辑 | Use methods to verify info truth: cross-sources, check original, test logic | 验证把『怀疑』变成『确认』,是判断真伪的核心动作 | Verification turns 'skepticism' into 'confirmation', the core of truth judgment | 用 3 个独立信源交叉验证 | Cross-verify with 3 independent sources |
| 推理(Reasoning) | 推理(Reasoning) | 从前提推出结论的能力,识别逻辑漏洞 | Ability to derive conclusions from premises, identify logic gaps | AI 会给你『看起来对』的推理,你要能识别逻辑漏洞 | AI gives 'plausible-looking' reasoning; you must spot logic gaps | 练识别 10 种常见逻辑谬误 | Practice spotting 10 common logical fallacies |
| 证据(Evidence) | 证据(Evidence) | 判断『证据是否足够支撑结论』,区分事实/观点/推测 | Judge if evidence supports conclusion; distinguish fact/opinion/speculation | AI 常『结论先行、证据凑数』,你要能识别 | AI often 'conclusion first, evidence padded'; you must spot it | 问『证据是什么?谁提供的?可重复吗?』 | Ask: 'What's the evidence? Who provided? Reproducible?' |
| 反思(Reflection) | 反思(Reflection) | 对自己的判断保持怀疑:我的结论可能错吗? | Keep skepticism toward your own judgment: could my conclusion be wrong? | 批判性思维的对象包括『自己』,不只是别人 | Critical thinking's target includes 'yourself', not just others | 下结论前问『我可能错在哪?』 | Before concluding, ask: 'Where might I be wrong?' |
| 元认知(Metacognition) | 元认知(Metacognition) | 监控自己的思维过程:我是怎么得出这个结论的? | Monitor your own thinking: how did I reach this conclusion? | 元认知让你『看见』自己的思维陷阱,是最高层能力 | Metacognition lets you 'see' your own thinking traps; the highest-level ability | 写『判断日志』:记录我怎么判断的 | Write a 'judgment log': how I judged |
10 个『信息验证』提示词(直接复制用)
10 'Info Verification' Prompts (Copy & Use)
| 场景 | Scenario | 提示词(中文) | Prompt (ZH) | 提示词(英文) | Prompt (EN) | 验证什么 | What it verifies |
|---|---|---|---|---|---|---|---|
| 事实核查 | Fact-check | 『请核查以下说法是否属实,给出 3 个独立信源和它们的原始链接:[粘贴说法]。如果无法确认,请明确说『无法验证』。』 | 'Verify if the following claim is true, give 3 independent sources and their original links: [paste claim]. If unable to confirm, explicitly say "unable to verify".' | 事实是否真实 | Whether the fact is true | ||
| 查出处 | Trace source | 『这个说法最早出自哪里?找到原始出处(论文/新闻/官方数据),而不是二手转述。』 | 'Where does this claim originally come from? Find the original source (paper/news/official data), not secondary retelling.' | 原始出处 | Original source | ||
| 逻辑检查 | Logic check | 『请检查以下论证的逻辑:列出所有前提和结论,指出逻辑漏洞(如有),判断结论是否必然成立。』 | 'Check the logic of this argument: list all premises and conclusions, point out logic gaps (if any), judge if the conclusion necessarily holds.' | 逻辑是否严密 | Whether logic is rigorous | ||
| 查偏见 | Bias check | 『分析这段内容可能存在的偏见:立场偏见、选择偏见、幸存者偏见。指出哪些是事实、哪些是观点、哪些是推测。』 | 'Analyze possible biases in this content: stance bias, selection bias, survivorship bias. Point out which are facts, opinions, speculations.' | 内容偏见 | Content bias | ||
| 反方论证 | Counter-argument | 『请为以下观点提出最强的 3 个反方论证(steelman),每个都要给出证据。然后评估哪一方证据更强。』 | 'Provide the strongest 3 counter-arguments (steelman) for this view, each with evidence. Then assess which side has stronger evidence.' | 论点强弱 | Argument strength | ||
| 查数据 | Check data | 『这个数据是哪年、哪个机构发布的?样本量多大?是横截面还是纵向研究?数据是否可重复?』 | 'What year, which institution released this data? Sample size? Cross-sectional or longitudinal? Is data reproducible?' | 数据可靠性 | Data reliability | ||
| 查时效 | Check timeliness | 『这个信息是什么时候的?现在(2026)是否仍然成立?有没有被新信息推翻?』 | 'When was this info? Does it still hold now (2026)? Has it been overturned by newer info?' | 信息时效 | Info timeliness | ||
| AI 自查 | AI self-check | 『你刚才的回答,哪些部分是你『确定知道』的,哪些是『推测』的,哪些可能『编造』的?分别标注。』 | 'In your previous answer, which parts do you 'know for sure', which are 'speculation', which might be 'fabricated'? Label each.' | AI 回答可信度 | AI answer credibility | ||
| 查利益 | Check interest | 『提出这个观点的机构/人,有什么利益动机?他们从「你相信这个」里能获益什么?』 | 'What's the interest motive of the institution/person making this claim? What do they gain from 'you believing this'?' | 利益动机 | Interest motive | ||
| 综合判断 | Comprehensive judge | 『综合以上所有验证结果,给出一个 0-100 的可信度评分,并说明:哪些证据支持、哪些反对、最大的不确定性在哪。』 | 'Synthesize all verification results, give a 0-100 credibility score, explain: which evidence supports, which opposes, where's the biggest uncertainty.' | 综合可信度 | Overall credibility |
12 种最常见的逻辑谬误(AI 时代高发)
12 Most Common Logical Fallacies (High-Frequency in AI Era)
| 谬误 | Fallacy | 是什么 | What it is | 例子 | Example | 怎么识别 | How to spot |
|---|---|---|---|---|---|---|---|
| 稻草人(Straw man) | 稻草人(Straw man) | 歪曲对方观点,攻击歪曲后的版本 | Distort the other's view, attack the distorted version | A:我觉得 AI 有用。B:你是说 AI 能取代所有人?A 没这么说 | A: I think AI is useful. B: So you mean AI can replace everyone? A never said that | 问『他原话是什么?』 | Ask: 'What was his original wording?' |
| 滑坡(Slippery slope) | 滑坡(Slippery slope) | A 会导致 B,B 导致 C... 一路推到灾难,中间环节没证据 | A leads to B, B leads to C... all the way to disaster, middle steps have no evidence | 允许用 AI 写作业 → 没人会思考 → 人类灭亡。中间每一步都没证据 | Allowing AI homework → no one thinks → humanity ends. Each middle step has no evidence | 问『每一步都有证据吗?』 | Ask: 'Does each step have evidence?' |
| 诉诸权威(Appeal to authority) | 诉诸权威(Appeal to authority) | 『某某专家说了,所以是对的』,但专家可能说错/不在其领域 | 'A certain expert said so, therefore it's right', but the expert may be wrong/out of field | 某明星说『某保健品治癌』——明星不是医学专家 | A celebrity says 'a certain supplement cures cancer' — the celebrity isn't a medical expert | 问『他在这个领域是权威吗?证据呢?』 | Ask: 'Is he an authority in this field? What's the evidence?' |
| 以偏概全(Hasty generalization) | 以偏概全(Hasty generalization) | 用个例推出普遍结论 | Use one case to derive a universal conclusion | 我认识 3 个用 AI 的人都没变聪明,所以 AI 没用。3 个样本不代表全体 | I know 3 people who use AI and none got smarter, so AI is useless. 3 samples don't represent the whole | 问『样本量多大?有代表性吗?』 | Ask: 'How big is the sample? Is it representative?' |
| 非黑即白(False dilemma) | 非黑即白(False dilemma) | 把选项简化成两个,其实还有中间选项 | Reduce options to two, when there are actually intermediate options | 你要么全人工,要么全 AI。其实可以混合用 | Either fully manual or fully AI. Actually you can mix | 问『真的只有两个选项吗?』 | Ask: 'Are there really only two options?' |
| 诉诸情感(Appeal to emotion) | 诉诸情感(Appeal to emotion) | 用恐惧/同情/愤怒代替证据 | Use fear/sympathy/anger instead of evidence | 『你不用 AI 就会被时代抛弃!』(制造恐惧) | 'If you don't use AI, you'll be left behind by the era!' (manufacturing fear) | 问『抛开情绪,证据是什么?』 | Ask: 'Setting aside emotion, what's the evidence?' |
| 循环论证(Circular reasoning) | 循环论证(Circular reasoning) | 结论当证据,证据当结论 | Use the conclusion as evidence, and the evidence as the conclusion | AI 是可靠的,因为 AI 自己说自己可靠 | AI is reliable because AI itself claims to be reliable | 问『证据独立于结论吗?』 | Ask: 'Is the evidence independent of the conclusion?' |
| 偷换概念(Equivocation) | 偷换概念(Equivocation) | 同一个词在论证中悄悄换了含义 | The same word secretly changes meaning in the argument | 『自由』:前一句指『言论自由』,后一句变成『想干嘛就干嘛』 | 'Freedom': in one sentence means 'freedom of speech', in the next becomes 'do whatever you want' | 问『这个词前后含义一样吗?』 | Ask: 'Is this word's meaning the same before and after?' |
| 诉诸无知(Appeal to ignorance) | 诉诸无知(Appeal to ignorance) | 『没人证明它是错的,所以它是对的』(或反之) | 'No one has proven it wrong, so it's right' (or the reverse) | 没人证明 AI 不会造反,所以它不会造反。『没证据』不等于『证明』 | No one has proven AI won't revolt, so it won't. 'No evidence' ≠ 'proof' | 问『是证明了,还是只是没发现?』 | Ask: 'Is it proven, or just not found yet?' |
| 红鲱鱼(Red herring) | 红鲱鱼(Red herring) | 插入无关话题,转移注意力 | Insert an unrelated topic to divert attention | A:你的产品有 bug。B:但你公司裁员了 100 人,多不人道! | A: Your product has bugs. B: But your company laid off 100 people, how inhumane! | 问『这跟原问题有关吗?』 | Ask: 'Is this related to the original question?' |
| 合成谬误(Fallacy of composition) | 合成谬误(Fallacy of composition) | 部分成立,整体就成立(不一定) | What holds for parts holds for the whole (not necessarily) | 每个球员都很强,所以这支球队一定赢。团队是整体 | Each player is strong, so this team must win. A team is a whole | 问『部分成立能推出整体吗?』 | Ask: 'Does what holds for parts hold for the whole?' |
| 幸存者偏差(Survivorship bias) | 幸存者偏差(Survivorship bias) | 只看到『活下来』的,看不到『没活下来』的 | Only see the 'survivors', not the 'non-survivors' | 马斯克用 AI 成功了,所以 AI 一定让你成功。看不到失败者 | Musk succeeded with AI, so AI will surely make you succeed. You don't see the failure cases | 问『没成功的人呢?样本在哪?』 | Ask: 'What about those who didn't succeed? Where's the sample?' |
15 个『AI 输出』的红旗信号(见到就要警惕)
15 'Red Flag' Signals in AI Output (Be Alert When You See Them)
| 红旗 | Red flag | 表现 | How it shows | 为什么危险 | Why it's dangerous | 怎么应对 | How to respond |
|---|---|---|---|---|---|---|---|
| 过度自信 | Overconfidence | 语气绝对:『这绝对是』『毫无疑问』 | Absolute tone: 'This is absolutely', 'Without a doubt' | AI 的置信度和正确率无关,自信≠正确 | AI's confidence is unrelated to accuracy; confident ≠ correct | 问『你有多确定?证据呢?』 | Ask: 'How sure are you? What's the evidence?' |
| 编造引用 | Fabricated citations | 给出看似真实的论文/书籍/数据,但查无此物 | Give seemingly real papers/books/data, but they don't exist | 幻觉的核心表现:AI『想象』出来源 | The core of hallucination: AI 'imagines' sources | 逐条验证引用是否真实 | Verify each citation is real |
| 数据无出处 | Data without source | 给出具体数字,但说不出哪年、哪个机构 | Give specific numbers, but can't say which year, which institution | 数字是 AI 编的最常见形式 | Numbers are the most common form of AI fabrication | 问『这个数据哪来的?』 | Ask: 'Where does this data come from?' |
| 逻辑跳跃 | Logic leap | 前提和结论之间有大缺口,用『显然』『显然可知』跳过 | Big gap between premise and conclusion, skipped with 'obviously', 'clearly' | AI 用『显然』掩盖逻辑漏洞 | AI uses 'obviously' to cover logic gaps | 要求它列出每一步推理 | Require it to list each step of reasoning |
| 以偏概全 | Hasty generalization | 用个例支撑普遍结论 | Use one case to support a universal conclusion | AI 爱用『我见过/有案例』当证据 | AI loves using 'I've seen/a case' as evidence | 问『还有多少案例?』 | Ask: 'How many more cases are there?' |
| 诉诸权威 | Appeal to authority | 反复提『专家说』『研究表明』但不给具体出处 | Repeatedly cite 'experts say', 'research shows' but give no specific source | 权威是 AI 最常搬的『大旗』 | Authority is the biggest 'banner' AI most often waves | 问『哪个专家?哪篇研究?』 | Ask: 'Which expert? Which study?' |
| 回避反例 | Avoiding counterexamples | 只给支持结论的证据,对反例视而不见 | Only give evidence supporting the conclusion, ignore counterexamples | AI 默认『讨好』,只说你想听的 | AI defaults to 'please', only says what you want to hear | 主动问『反例是什么?』 | Actively ask: 'What are the counterexamples?' |
| 答非所问 | Answering something else | 你的问题被悄悄『改写』成它更容易回答的版本 | Your question is quietly 'rewritten' into a version it can answer more easily | AI 回避难问题,偷换问题 | AI avoids hard questions, swaps the question | 核对它答的是不是你的原问题 | Check if it's answering your original question |
| 过度具体 | Overly specific | 细节多到可疑:精确到个位的数字、不存在的日期 | Details so specific they're suspicious: numbers precise to the unit, non-existent dates | AI 用『具体感』伪装真实性 | AI uses 'specificity' to fake authenticity | 对可疑细节单独验证 | Verify suspicious details individually |
| 无差别支持 | Indiscriminate support | 无论你怎么问,它都支持你的立场 | No matter how you ask, it supports your stance | AI 的『讨好偏差』:镜像你的观点 | AI's 'people-pleasing bias': mirrors your view | 问『如果我是错的,你凭什么说对?』 | Ask: 'If I'm wrong, what's your basis for saying I'm right?' |
| 术语堆砌 | Jargon stacking | 大量专业术语,但解释不清楚 | Massive professional jargon, but no clear explanation | 术语是 AI 的『遮羞布』,掩盖不懂 | Jargon is AI's 'fig leaf', covering its ignorance | 要求它用人话解释 | Require it to explain in plain language |
| 时间模糊 | Time ambiguity | 『现在』『最近』『研究表明』但不说具体年份 | 'Now', 'recently', 'research shows' but no specific year | AI 训练有截止时间,可能把旧信息当新的 | AI's training has a cutoff; it may treat old info as new | 问『这是哪年的信息?』 | Ask: 'What year is this info from?' |
| 绝对化 | Absolutism | 『一定』『必然』『100%』等绝对词 | Absolute words like 'must', 'inevitably', '100%' | 现实中几乎没有 100% 确定的事 | There's almost nothing 100% certain in reality | 要求它给置信度和例外 | Require it to give confidence and exceptions |
| 单一信源 | Single source | 所有结论都来自『同一个』来源(可能是 AI 自己) | All conclusions come from 'the same' source (possibly the AI itself) | AI 会把自己当信源 | AI treats itself as a source | 要求它给独立信源 | Require it to give independent sources |
| 情感操控 | Emotional manipulation | 用『你知道吗』『震惊』『你绝对想不到』等钩子 | Hooks like 'Did you know', 'Shocking', 'You absolutely won't believe' | AI 学自营销文案,会带节奏 | AI learned from marketing copy, can lead the rhythm | 剥离情感,只看事实 | Strip emotion, look at facts only |
10 个『自我批判』的元问题(下结论前必问)
10 'Self-Critique' Meta-Questions (Must Ask Before Concluding)
| 问题 | Question | 问什么 | What it asks | 为什么重要 | Why it matters |
|---|---|---|---|---|---|
| 我可能错在哪? | Where might I be wrong? | 列出你结论最薄弱的 1 个点 | List the 1 weakest point of your conclusion | 暴露盲点,避免过度自信 | Expose blind spots, avoid overconfidence |
| 什么证据能推翻我? | What evidence would overturn me? | 如果看到什么,你会改主意? | If you saw what, would you change your mind? | 可证伪性是好判断的标志 | Falsifiability is a sign of good judgment |
| 我的信息源单一吗? | Is my source single? | 我是不是只看了『一边』? | Am I only looking at 'one side'? | 单一信源 = 信息茧房 | Single source = info cocoon |
| 我是否被情绪带着走? | Am I being led by emotion? | 我是被愤怒/恐惧/同情驱动的吗? | Am I driven by anger/fear/sympathy? | 情绪是判断的最大敌人 | Emotion is the biggest enemy of judgment |
| 我的结论是事实还是观点? | Is my conclusion a fact or opinion? | 我能验证它吗? | Can I verify it? | 区分事实/观点是批判性思维基础 | Distinguishing fact/opinion is the foundation of critical thinking |
| 我是否在『确认偏差』? | Am I in 'confirmation bias'? | 我只看支持我的证据吗? | Am I only looking at evidence supporting me? | 人天然只找『对的』证据 | Humans naturally only seek 'right' evidence |
| 反方最强的论证是什么? | What's the strongest counter-argument? | steelman:替对方说最强的话 | steelman: speak the strongest on behalf of the other | 理解反方 = 理解全貌 | Understanding the counter = understanding the whole |
| 我的样本有代表性吗? | Is my sample representative? | 我看到的『个例』能代表全体吗? | Can the 'case' I see represent the whole? | 以偏概全是常见谬误 | Hasty generalization is a common fallacy |
| 这个信息还时效吗? | Is this info still timely? | 它是什么时候的?现在还成立吗? | When is it from? Does it still hold now? | AI 信息可能过时 | AI info may be outdated |
| 我为什么相信这个? | Why do I believe this? | 我的判断基于证据,还是基于『我觉得』? | Is my judgment based on evidence, or on 'I feel'? | 元认知:看见自己的判断过程 | Metacognition: see your own judgment process |
50 个『批判性思维』高频问题(一问一答)
50 'Critical Thinking' High-Frequency Questions (One Q One A)
30 条『批判性思维』金句(可背可分享)
30 'Critical Thinking' Golden Quotes (Memorize & Share)
| 金句 | Quote | 出处/背景 | Origin/Context | 怎么用 | How to use |
|---|---|---|---|---|---|
| 未经审视的生活不值得过。 | The unexamined life is not worth living. | 苏格拉底(Socrates) | 用于提醒自己『别活在惯性里』,定期审视自己的判断。 | ||
| 我思故我在。 | I think, therefore I am. | 笛卡尔(Descartes) | 判断力的起点是『思考』,不是『反应』。 | ||
| 怀疑不是目的,更接近真相才是目的。 | Doubt is not the goal; getting closer to truth is. | 科学方法论精神(Scientific method spirit) | 怀疑是手段,不是立场。怀疑完要得出『最可能的结论』。 | ||
| 相关不等于因果。 | Correlation is not causation. | 统计学基本原则(Basic statistical principle) | 看到两个现象一起出现,先问:是因果、巧合、还是第三方? | ||
| 能推翻你的证据,你找了吗? | Did you look for evidence that could overturn you? | 可证伪性(Falsifiability, 波普尔) | 好判断必须『可被推翻』,只找支持自己的证据 = 自欺。 | ||
| 你相信的,和你知道的,是两回事。 | What you believe and what you know are two different things. | 认知科学(Cognitive science) | 把『信念』当『事实』是最常见的判断错误。 | ||
| 证据的等级:RCT > 观察研究 > 专家意见 > 个例。 | Evidence hierarchy: RCT > observational study > expert opinion > anecdote. | 循证医学(Evidence-based medicine) | 判断任何主张时,先看它的『证据等级』。 | ||
| 最好的论证,是你能替对手说出最强版本。 | The best argument is one you can state the strongest version of for your opponent. | 钢铁人原则(Steelmanning) | 先 steelman 对方,再反驳,你的论证才有质量。 | ||
| 情绪是信号,不是证据。 | Emotion is a signal, not evidence. | 认知科学(Cognitive science) | 愤怒/恐惧让你『想』判断,但不提供『判断的依据』。 | ||
| 简单不是正确,但简单更容易被检验。 | Simple isn't correct, but simple is easier to test. | 奥卡姆剃刀(Occam's razor) | 在多个解释里,选『假设最少』的那个,更容易被验证。 | ||
| 无知不是中立,无知是偏见。 | Ignorance isn't neutral; ignorance is bias. | 认知科学(Cognitive science) | 『不知道』不等于『各占一半』,无知会让你被带节奏。 | ||
| 样本量小于 30,结论只是『个例』。 | Sample size under 30, the conclusion is just an 'anecdote'. | 统计学(Statistics) | 小样本的『规律』大多是大脑在『找规律』。 | ||
| 越自信的人,越该被检验。 | The more confident, the more should be tested. | 决策科学(Decision science) | 自信程度和正确率几乎无关,越自信越要警惕。 | ||
| 把问题说清楚,就解决了一半。 | Stating the problem clearly is half solved. | 批判性思维核心(Core of critical thinking) | 大多数『想不通』是『没说清』。先定义问题,再找答案。 | ||
| 前提错了,结论再漂亮也没用。 | If the premise is wrong, no matter how pretty the conclusion is useless. | 逻辑学(Logic) | 评估论证先评估前提,前提错 = 全错。 | ||
| 你不需要『赢』,你需要『对』。 | You don't need to 'win'; you need to 'be right'. | 批判性思维 vs 抬杠(Critical thinking vs contrarian) | 讨论的目的是『更接近真相』,不是『压过对方』。 | ||
| 承认『我不知道』,是智慧的开始。 | Admitting 'I don't know' is the beginning of wisdom. | 苏格拉底(Socrates) | 最诚实的判断力来自『知道自己不知道』。 | ||
| 信息不是知识,理解才是知识。 | Information is not knowledge; understanding is knowledge. | 认知科学(Cognitive science) | 刷 100 条新闻 ≠ 理解一件事,理解 = 能解释 + 能应用。 | ||
| 让数据说话,先问『数据是谁选的』。 | Let the data speak, first ask 'who selected the data'. | 数据批判(Data criticism) | 数据从来不是『中立的』,它被『选择』过。 | ||
| 警惕『最后一天的优惠』。 | Beware of 'last day only' deals. | 营销心理学(Marketing psychology) | 『紧迫感』是情绪操控,不是事实。剥离时间压力,再看决策。 | ||
| 从众不是判断,是放弃判断。 | Following the crowd is not judgment; it's giving up judgment. | 社会心理学(Social psychology) | 『大家都这样』不是理由,是『放弃思考』的借口。'Everyone does this' is not a reason, it's an excuse to 'give up thinking'. | ||
| 最危险的不是假信息,是你『相信』的假信息。 | The most dangerous is not false info, but false info you 'believe'. | 信息素养(Information literacy) | 你相信什么,比信息真假更影响你的行为。 | ||
| 把『我觉得』换成『证据是』,你就升级了。 | Replace 'I feel' with 'the evidence is', and you've upgraded. | 表达习惯(Expression habit) | 这是批判性思维最简单的自我训练。 | ||
| 先问『这是事实还是观点』,再谈对错。 | First ask 'is this fact or opinion', then discuss right/wrong. | 事实/观点区分(Fact/opinion distinction) | 事实可验证,观点只能讨论,混淆两者 = 抬杠。 | ||
| 你的大脑不是为『正确』设计的,是为『快速』设计的。 | Your brain isn't designed for 'correct', it's for 'fast'. | 进化心理学(Evolutionary psychology) | 所以『直觉』常错,重要决策必须『慢下来』。 | ||
| AI 给你的答案,你验证了吗? | Did you verify the answer AI gave you? | AI 时代批判性思维(AI-era critical thinking) | AI 时代最基础的批判性动作:不轻信 + 验证。 | ||
| 最好的提问,是让你『少说结论,多说证据』。 | The best question makes you 'say less conclusion, more evidence'. | 提问技巧(Questioning technique) | 当 AI 给你结论时,追问『证据是什么』。 | ||
| 批判性思维不是『否定一切』,是『有根据地评估』。 | Critical thinking is not 'denying everything', it's 'evaluating with grounds'. | 定义(Definition) | 它的产出是『更可靠的判断』,不是『更多怀疑』。 | ||
| 在 AI 时代,『会问』比『会答』更稀缺。 | In the AI era, 'knowing how to ask' is scarcer than 'knowing how to answer'. | AI 时代能力(AI-era ability) | 答案无限廉价,『好问题』才决定答案质量。 | ||
| 你的注意力被操纵时,你的判断就被操纵了。 | When your attention is manipulated, your judgment is manipulated. | 注意力经济(Attention economy) | 批判性思维的第一层:保护『注意力』不被带节奏。 |
30 条『批判性思维』进阶技巧(高手常用)
30 Advanced 'Critical Thinking' Techniques (Used by Experts)
| 技巧 | Technique | 做法 | How to do it | 效果 | Effect |
|---|---|---|---|---|---|
| 钢铁人论证 | Steelmanning | 先替对方说『最强版本』,再反驳 | First state the 'strongest version' for the other, then rebut | 避免『打稻草人』,论证质量翻倍 | Avoids 'hitting a straw man', argument quality doubles |
| 预-mortem | Pre-mortem | 假设项目已经失败,倒推『怎么会失败』 | Assume the project already failed, work backward 'how it could fail' | 在成功前找出致命风险 | Find fatal risks before success |
| 贝叶斯更新 | Bayesian update | 先给『先验概率』,看到新证据后更新 | First give 'prior probability', update after seeing new evidence | 把『判断』变成『可计算的』 | Turns 'judgment' into 'calculable' |
| 5 个为什么 | 5 Whys | 对问题连续问 5 次『为什么』 | Ask 'why' 5 times consecutively for a problem | 从『表面现象』挖到『根本原因』 | Dig from 'surface phenomenon' to 'root cause' |
| 红队思维 | Red-teaming | 专门扮演『攻击方』找漏洞 | Play the 'attacker' role to find gaps | 在自己的方案被攻击前,自己先攻击 | Attack your own plan before it gets attacked |
| 魔鬼代言人 | Devil's advocate | 主动请人『反对你的观点』 | Actively ask someone to 'oppose your view' | 打破『群体思维』 | Breaks 'groupthink' |
| 证据等级法 | Evidence level method | 给每条证据标等级(RCT/观察/个例) | Label each piece of evidence with a level (RCT/observation/anecdote) | 避免『被个例带节奏』 | Avoids 'being led by anecdotes' |
| 反事实思考 | Counterfactual thinking | 问『如果不是这样,会怎样』 | Ask 'what if it weren't so, what would happen' | 暴露『隐藏假设』 | Exposes 'hidden assumptions' |
| 第一性原理 | First principles | 回到『最基本的事实』重新推导 | Return to 'basic facts' and re-derive | 避免『类比思维』的陷阱 | Avoids the trap of 'analogical thinking' |
| 奥卡姆剃刀 | Occam's razor | 多个解释选『假设最少』的 | Among multiple explanations, pick the 'fewest assumptions' | 简单更易验证,减少过度解读 | Simple is easier to verify, reduces over-interpretation |
| 概率思维 | Probabilistic thinking | 用『概率』而非『对错』描述判断 | Describe judgments with 'probability' not 'right/wrong' | 避免『非黑即白』 | Avoids 'black or white' |
| 基础比率 | Base rate | 先看『一般概率』,再看『个体情况』 | First look at 'general probability', then 'individual situation' | 避免『个例偏差』 | Avoids 'anecdote bias' |
| 锚定效应免疫 | Anchoring immunity | 对『第一个数字』保持警惕 | Stay wary of 'the first number' | 避免『被价格/数字带节奏』 | Avoids 'being led by prices/numbers' |
| 损失框架检查 | Loss-frame check | 问『这个决策让我损失什么』 | Ask 'what does this decision make me lose' | 避免『损失厌恶』的盲区 | Avoids the blind spot of 'loss aversion' |
| 决策日志 | Decision journal | 记录每次决策的『理由 + 当时情绪』 | Record each decision's 'reasons + emotions at the time' | 事后复盘,找到自己的『判断模式』 | Review after the fact, find your own 'judgment patterns' |
| 双盲检查 | Double-blind check | 让别人独立评估你的论证 | Let someone independently evaluate your argument | 避免『自我说服』 | Avoids 'self-persuasion' |
| 时间折扣检查 | Time-discount check | 问『10 年后看这个决策,还成立吗』 | Ask 'looking back at this decision in 10 years, does it still hold' | 避免『短期冲动』 | Avoids 'short-term impulsiveness' |
| 身份偏见检查 | Identity-bias check | 问『我是因为立场而相信,还是因为证据』 | Ask 'do I believe because of my stance, or because of evidence' | 避免『立场先于判断』 | Avoids 'stance before judgment' |
| 情绪标注 | Emotion labeling | 在判断前标注『我现在的情绪』 | Before judging, label 'my current emotion' | 情绪被看见后,影响力下降 50% | Once emotion is seen, its influence drops 50% |
| 信息断食 | Information diet | 每天固定 1 小时『不碰任何信息』 | 1 hour daily 'not touching any info' | 恢复『独立思考』能力 | Restores 'independent thinking' ability |
| 单一信源惩罚 | Single-source penalty | 只看到一个信源时,判断自动降 30% | When only seeing one source, judgment auto-drops 30% | 强制『多信源验证』 | Forces 'multi-source verification' |
| 置信度校准 | Confidence calibration | 每次判断给『置信度%』,事后核对 | Each judgment gives 'confidence %', check after the fact | 训练『自信程度 = 正确率』 | Trains 'confidence level = accuracy' |
| 可证伪清单 | Falsifiability checklist | 判断前问『什么证据能推翻它』 | Before judging, ask 'what evidence could overturn it' | 避免『不可推翻』的伪判断 | Avoids 'unfalsifiable' pseudo-judgments |
| 样本检查 | Sample check | 问『样本量多大?有代表性吗?』 | Ask 'how big is the sample? Is it representative?' | 避免『以偏概全』 | Avoids 'hasty generalization' |
| 时间戳检查 | Timestamp check | 问『这是哪年的信息?还有效吗?』 | Ask 'what year is this info from? Still valid?' | 避免『用旧信息判断新情况』 | Avoids 'judging new situations with old info' |
| 因果方向检查 | Causal-direction check | 问『是 A 导致 B,还是 B 导致 A』 | Ask 'is A causing B, or B causing A' | 避免『因果倒置』 | Avoids 'reversed causation' |
| 第三方变量 | Third-variable | 问『有没有 C 同时导致 A 和 B』 | Ask 'is there a C causing both A and B' | 避免『虚假相关』 | Avoids 'spurious correlation' |
| 幸存者检查 | Survivorship check | 问『失败/没活下来的在哪』 | Ask 'where are the failures/ones that didn't survive' | 避免『幸存者偏差』 | Avoids 'survivorship bias' |
| 动机检查 | Motivation check | 问『说这话的人,有什么动机』 | Ask 'what motive does the person making this statement have' | 避免『被利益驱动的信息』带节奏 | Avoids 'being led by interest-driven info' |
| 框架检查 | Frame check | 问『这个问题是怎么被『框』的』 | Ask 'how is this problem being 'framed'' | 避免『框架操纵』 | Avoids 'frame manipulation' |
10 个『批判性思维』训练工具(免费)
10 Free 'Critical Thinking' Training Tools
| 工具 | Tool | 类型 | Type | 怎么用 | How to use | 费用 | Cost |
|---|---|---|---|---|---|---|---|
| Pond5 逻辑谬误测试 | Logic Fallacy Test (online) | 测试 | Test | 每年测一次,看自己最容易中哪种谬误 | Test once a year, see which fallacy you most easily fall into | 免费 | Free |
| Anki(间隔重复) | Anki (spaced repetition) | 记忆 | Memory | 把 12 种谬误 + 30 条技巧做成卡片,每天复习 10 分钟 | Make cards for 12 fallacies + 30 techniques, review 10 min daily | 免费 | Free |
| 新闻溯源(WHO 查源) | News source-checking (WHO) | 验证 | Verification | 看到新闻先查『WHO 信源核查』 | See news, first check 'WHO source verification' | 免费 | Free |
| FactCheck.org | FactCheck.org | 核查 | Fact-check | 验证『政治/商业』主张的真假 | Verify the truth of 'political/business' claims | 免费 | Free |
| Our World in Data | Our World in Data | 数据 | Data | 看『全球数据』的可视化,训练『数据直觉』 | See 'global data' visualization, train 'data intuition' | 免费 | Free |
| Coursera 批判性思维课 | Coursera Critical Thinking Course | 课程 | Course | 系统学批判性思维(有免费旁听) | Systematically study critical thinking (free audit available) | 免费旁听 | Free audit |
| Khan Academy 逻辑 | Khan Academy Logic | 课程 | Course | 学『形式逻辑』基础(免费) | Learn 'formal logic' basics (free) | 免费 | Free |
| AI 对话(验证模式) | AI Conversation (verification mode) | 练习 | Practice | 每天用 AI 做『验证练习』:给它一个主张,让它找反例 | Daily 'verification practice' with AI: give it a claim, let it find counterexamples | 免费/订阅 | Free/sub |
| 决策日志(Notion/Excel) | Decision Journal (Notion/Excel) | 复盘 | Review | 每次重要决策记录理由 + 情绪,每月复盘 | Record reasons + emotions for each important decision, review monthly | 免费 | Free |
| 辩论俱乐部/读书会 | Debate Club/Book Club | 社群 | Community | 每周 1 次『辩论 + 复盘』,训练『对抗性思维』 | Weekly 'debate + review', train 'adversarial thinking' | 免费 | Free |
60 个『判断练习』:这是事实、观点还是带节奏?
60 'Judgment Drills': Fact, Opinion, or Being Led?
| 句子 | Statement | 判断 | Verdict | 为什么 | Why |
|---|---|---|---|---|---|
| 『2026 年全球 AI 投资增长了 40%。』 | 'In 2026 global AI investment grew 40%.' | 观点(无来源) | Opinion (no source) | 有具体数字但没给来源/年份/机构,无法验证 | Has specific number but no source/year/institution, unverifiable |
| 『我认识 5 个用 AI 的都没变强。』 | 'I know 5 who use AI and none got stronger.' | 以偏概全 | Hasty generalization | 5 个样本是『个例』,不能代表全体 | 5 samples are 'anecdotes', can't represent the whole |
| 『AI 一定会取代所有工作。』 | 'AI will definitely replace all jobs.' | 绝对化 + 滑坡 | Absolutism + slippery slope | 『一定』是绝对词,且中间环节无证据 | 'Definitely' is absolute, and middle steps lack evidence |
| 『研究表明 90% 的人觉得 AI 有用。』 | 'Research shows 90% think AI is useful.' | 红旗(无出处) | Red flag (no citation) | 『研究表明』没给具体研究,高风险幻觉 | 'Research shows' without specific study, high hallucination risk |
| 『水在 100 摄氏度沸腾。』 | 'Water boils at 100 degrees C.' | 事实(有条件) | Fact (conditional) | 在标准大气压下成立,是可验证的客观事实 | Holds at standard pressure, verifiable objective fact |
| 『这家 AI 公司是最强的。』 | 'This AI company is the strongest.' | 观点 | Opinion | 『最强』是主观比较,没有统一标准 | 'Strongest' is subjective comparison, no unified standard |
| 『你不用 AI 就会被淘汰。』 | 'If you don't use AI you'll be eliminated.' | 情感操控(恐惧) | Emotional manipulation (fear) | 制造恐惧逼你行动,不是提供证据 | Manufactures fear to force action, not provide evidence |
| 『这个模型参数量 1 万亿。』 | 'This model has 1 trillion parameters.' | 事实(可验证) | Fact (verifiable) | 具体数字,可查官方发布核实 | Specific number, verifiable against official release |
| 『用 AI 的人更聪明。』 | 'People using AI are smarter.' | 无证据的因果 | Causation without evidence | 『用 AI』和『聪明』可能只是相关,未必因果 | 'Using AI' and 'smart' may only be correlated, not causal |
| 『AI 能写代码。』 | 'AI can write code.' | 事实 | Fact | 可重复验证的客观能力,无需争议 | Repeatably verifiable objective capability, no dispute |
| 『你买这个就不会后悔。』 | 'You won't regret buying this.' | 承诺(无法验证) | Promise (unverifiable) | 『不会后悔』是未来承诺,无法事前验证 | 'Won't regret' is a future promise, can't be pre-verified |
| 『95% 的用户满意度。』 | '95% user satisfaction.' | 数据(需验证来源) | Data (source needed) | 具体数字,但要问:谁测的?样本多大? | Specific number, but ask: who measured? sample size? |
| 『AI 时代必须学批判性思维。』 | 'You must learn critical thinking in the AI era.' | 观点(合理) | Opinion (reasonable) | 合理建议,但『必须』带强推,要自己判断 | Reasonable suggestion, but 'must' pushes hard, judge yourself |
| 『这项研究在顶刊发表。』 | 'This study was published in a top journal.' | 事实(可验证) | Fact (verifiable) | 『顶刊发表』是客观事件,可查证 | 'Published in top journal' is objective event, checkable |
| 『大家都说这款好。』 | 'Everyone says this one's good.' | 诉诸大众 | Appeal to crowd | 『大家』是谁?没具体,是『从众』信号 | Who's 'everyone'? Not specific, a 'herd' signal |
| 『AI 让效率提升 3 倍。』 | 'AI triples efficiency.' | 数据(需条件) | Data (needs conditions) | 在什么场景/什么任务下 3 倍?无上下文不可用 | In what scenario/task is it 3x? Unusable without context |
| 『我不信 AI。』 | 'I don't trust AI.' | 观点(个人) | Opinion (personal) | 个人态度,无对错,但要问『为什么不信』 | Personal attitude, no right/wrong, but ask 'why not trust' |
| 『这个数据来自 2024 年。』 | 'This data is from 2024.' | 事实(时效信息) | Fact (timeliness info) | 明确的年份,可判断是否过时 | Clear year, can judge if outdated |
| 『AI 会消灭人类。』 | 'AI will destroy humanity.' | 极端滑坡 + 无证据 | Extreme slippery slope + no evidence | 『消灭』是极端结论,中间环节全无证据 | 'Destroy' is extreme conclusion, all middle steps lack evidence |
| 『这款 AI 比那款准 15%。』 | 'This AI is 15% more accurate than that one.' | 数据(需验证基准) | Data (needs benchmark) | 15% 比什么?在哪个测试集?基准不明 | 15% compared to what? Which test set? Unclear benchmark |
| 『专家说 AI 很安全。』 | 'Experts say AI is very safe.' | 诉诸权威(需核查) | Appeal to authority (need check) | 哪个专家?安全是相对概念,要具体 | Which expert? Safety is relative, needs specifics |
| 『我用过,很好用。』 | 'I've used it, very easy.' | 个例(主观) | Anecdote (subjective) | 个人体验,是『个例』不是『数据』 | Personal experience, 'anecdote' not 'data' |
| 『AI 训练需要 10000 块 GPU。』 | 'AI training needs 10000 GPUs.' | 事实(可验证) | Fact (verifiable) | 具体数字,可查公开资料核实 | Specific number, verifiable via public sources |
| 『买它,最后 3 件。』 | 'Buy it, only 3 left.' | 紧迫感操控 | Urgency manipulation | 『最后 3 件』是饥饿营销,先核实库存 | 'Only 3 left' is hunger marketing, verify stock first |
| 『这个算法的准确率 99%。』 | 'This algorithm's accuracy is 99%.' | 数据(需问测试集) | Data (ask test set) | 99% 在什么数据上测的?测试集很关键 | 99% on what data? Test set is key |
| 『AI 让人类进步。』 | 'AI makes humanity progress.' | 观点(需定义) | Opinion (needs definition) | 『进步』如何定义?没有统一衡量 | How to define 'progress'? No unified measure |
| 『这家公司估值 1000 亿。』 | 'This company is valued at 100 billion.' | 事实(可验证) | Fact (verifiable) | 估值是公开信息,可查证 | Valuation is public info, checkable |
| 『AI 比人聪明。』 | 'AI is smarter than humans.' | 观点(需定义聪明) | Opinion (needs defining smart) | 『聪明』是多重维度,AI 在某些任务强不是全面 | 'Smart' is multi-dimensional, AI strong in some tasks isn't across the board |
| 『研究表明 AI 会取代 40% 的工作。』 | 'Research shows AI will replace 40% of jobs.' | 数据(需核查研究) | Data (need check study) | 哪个研究?40% 的方法论是什么? | Which study? What's the methodology for 40%? |
| 『我邻居用 AI 赚了 100 万。』 | 'My neighbor made 1M with AI.' | 幸存者偏差 | Survivorship bias | 只看到『赚了』的,看不到『没赚』的多数 | Only seeing the 'earned' ones, not the majority who didn't |
| 『AI 的准确率在提升。』 | 'AI's accuracy is improving.' | 趋势(需数据支撑) | Trend (needs data support) | 是事实趋势,但要具体指标和时间线 | Factual trend, but needs specific metrics and timeline |
| 『你不学 AI 就落伍了。』 | 'If you don't learn AI you'll fall behind.' | 情感操控(焦虑) | Emotional manipulation (anxiety) | 制造焦虑,不是提供『如何不落后』的证据 | Manufactures anxiety, not evidence of 'how to not fall behind' |
| 『这个数据集有 10 亿条。』 | 'This dataset has 1 billion entries.' | 事实(可验证) | Fact (verifiable) | 具体规模,可查公开资料 | Specific scale, verifiable via public sources |
| 『AI 让学习更高效。』 | 'AI makes learning more efficient.' | 观点(需证据) | Opinion (needs evidence) | 『更高效』要量化对比,无数据只是观点 | 'More efficient' needs quantified comparison, no data = opinion |
| 『这款模型开源了。』 | 'This model is open-sourced.' | 事实(可验证) | Fact (verifiable) | 开源是客观状态,可查仓库 | Open-source is objective state, checkable in repo |
| 『99% 的人不知道 AI 的真相。』 | '99% don't know the truth about AI.' | 绝对化 + 神秘化 | Absolutism + mystification | 『真相』是什么?制造信息差,带节奏 | What's 'the truth'? Manufactures info gap, manipulates |
| 『AI 的能耗是电力的 5 倍。』 | 'AI's energy use is 5x that of electricity.' | 数据(需验证) | Data (needs verification) | 具体倍数,但要问基准和年份 | Specific multiple, but ask baseline and year |
| 『用 AI 学习 3 个月就精通。』 | 'Master in 3 months with AI learning.' | 虚假承诺 | False promise | 『3 个月精通』是营销话术,无证据 | 'Master in 3 months' is marketing, no evidence |
| 『这个 API 每秒处理 1000 请求。』 | 'This API handles 1000 requests/sec.' | 事实(可验证) | Fact (verifiable) | 具体性能指标,可压测验证 | Specific performance metric, load-testable |
| 『AI 让普通人也能创造。』 | 'AI lets ordinary people create.' | 观点(部分成立) | Opinion (partly true) | 部分成立,但『创造』定义宽,要具体 | Partly true, but 'create' is broad, needs specifics |
| 『这款产品的用户增长了 200%。』 | 'This product's users grew 200%.' | 数据(需验证) | Data (needs verification) | 具体增长,但要问基数和时间 | Specific growth, but ask base and period |
| 『AI 会改变世界。』 | 'AI will change the world.' | 观点(太宽泛) | Opinion (too broad) | 『改变世界』太宽泛,无法证伪 | 'Change the world' too broad, unfalsifiable |
| 『这个模型的幻觉率 5%。』 | 'This model's hallucination rate is 5%.' | 数据(需验证) | Data (needs verification) | 具体幻觉率,但要问测试方法 | Specific hallucination rate, but ask test method |
| 『不用 AI 的人会被时代抛弃。』 | 'People not using AI will be left behind.' | 滑坡 + 情感 | Slippery slope + emotion | 『抛弃』是极端,制造恐惧 | 'Left behind' is extreme, manufactures fear |
| 『AI 的推理能力接近人类。』 | 'AI's reasoning is close to human.' | 观点(需定义接近) | Opinion (needs defining close) | 『接近』如何衡量?多维度,无统一 | How to measure 'close'? Multi-dimensional, no unified |
| 『这个数据集是合成的。』 | 'This dataset is synthetic.' | 事实(可验证) | Fact (verifiable) | 合成是客观属性,可查证 | Synthetic is objective property, checkable |
| 『AI 让 90% 的工作消失。』 | 'AI makes 90% of jobs disappear.' | 极端 + 无证据 | Extreme + no evidence | 90% 是极端数字,无可靠来源 | 90% is extreme number, no reliable source |
| 『这款 AI 比上代快 2 倍。』 | 'This AI is 2x faster than last gen.' | 数据(需验证基准) | Data (needs benchmark) | 2 倍比什么任务?基准不明 | 2x on what task? Unclear baseline |
| 『AI 让人更懒。』 | 'AI makes people lazier.' | 观点(需证据) | Opinion (needs evidence) | 『更懒』要量化,无数据只是感受 | 'Lazier' needs quantifying, no data = just feeling |
| 『这个模型的训练成本 100 万美元。』 | 'This model's training cost is $1M.' | 数据(需验证) | Data (needs verification) | 具体成本,但要问包含什么 | Specific cost, but ask what's included |
| 『AI 时代人人要学提示词。』 | 'In the AI era everyone must learn prompting.' | 观点(合理) | Opinion (reasonable) | 合理建议,但『人人』太绝对 | Reasonable, but 'everyone' is too absolute |
| 『这款 AI 的月活 1 亿。』 | 'This AI's monthly active is 100M.' | 数据(需验证) | Data (needs verification) | 具体 MAU,但要问统计口径 | Specific MAU, but ask statistical basis |
| 『AI 会取代教师。』 | 'AI will replace teachers.' | 滑坡 + 极端 | Slippery slope + extreme | 『取代』是极端,教育是复杂系统 | 'Replace' is extreme, education is a complex system |
| 『这个数据集有 100 个国家。』 | 'This dataset covers 100 countries.' | 事实(可验证) | Fact (verifiable) | 具体覆盖范围,可查证 | Specific coverage, checkable |
| 『AI 让信息过载。』 | 'AI causes information overload.' | 观点(需证据) | Opinion (needs evidence) | 『过载』要量化,无数据只是感受 | 'Overload' needs quantifying, no data = just feeling |
| 『这款 AI 的延迟 200ms。』 | 'This AI's latency is 200ms.' | 事实(可验证) | Fact (verifiable) | 具体性能,可测验证 | Specific performance, testable |
| 『AI 让人类思考退化。』 | 'AI makes human thinking degrade.' | 观点(需证据) | Opinion (needs evidence) | 『退化』要量化对比,无数据 | 'Degrade' needs quantified comparison, no data |
| 『这个模型的参数量 700 亿。』 | 'This model has 70B parameters.' | 事实(可验证) | Fact (verifiable) | 具体参数,可查官方 | Specific params, checkable officially |
40 个『信息验证』实战案例(AI 输出)
40 'Information Verification' Practical Cases (AI Output)
| AI 输出 | AI Output | 问题 | Problem | 验证步骤 | Verification Steps | 结论 | Verdict |
|---|---|---|---|---|---|---|---|
| 『研究表明 85% 的人用过 AI。』 | 'Research shows 85% have used AI.' | 无具体研究 | No specific study | ① 问哪篇研究 ② 搜该研究 ③ 查样本量 | ① Which study ② Search it ③ Check sample size | 无法验证,存疑 | Unverifiable, doubtful |
| 『2025 年 AI 市场 5000 亿美元。』 | '2025 AI market is $500B.' | 数字无来源 | Number no source | ① 问数据来源 ② 查行业报告 ③ 对比多家 | ① Ask source ② Check industry reports ③ Compare multiple | 数字可能有,但口径不一 | Number may exist, but basis varies |
| 『这个模型比 GPT 强 30%。』 | 'This model is 30% stronger than GPT.' | 基准不明 | Benchmark unclear | ① 问哪个测试 ② 查测试集 ③ 看是否独立 | ① Which test ② Check test set ③ See if independent | 30% 无上下文不可用 | 30% unusable without context |
| 『AI 会取代 40% 的工作(McKinsey)。』 | 'AI will replace 40% of jobs (McKinsey). | 引用需核实 | Citation needs check | ① 查 McKinsey 原报告 ② 看 40% 指什么 ③ 是否夸大 | ① Check McKinsey original ② What 40% means ③ Exaggerated? | McKinsey 有类似研究,但 40% 是『可自动化』非『取代』 | McKinsey has similar research, but 40% is 'automatable' not 'replace' |
| 『水在 100 度沸腾,这是错的,其实 95 度。』 | 'Water boils at 100, that's wrong, actually 95.' | 事实错误 | Factual error | ① 查科学常识 ② 注意气压条件 ③ AI 在编 | ① Check science common sense ② Note pressure ③ AI fabricating | 100 度(标准气压)正确,AI 错了 | 100 deg (standard pressure) is correct, AI is wrong |
| 『AI 训练用了 1 亿小时 GPU。』 | 'AI training used 100M GPU-hours.' | 数字可疑 | Number suspicious | ① 问来源 ② 查已知模型训练量 ③ 数量级对比 | ① Ask source ② Check known model training ③ Order of magnitude | 数量级可能,但无来源 | Order may be right, but no source |
| 『这款 AI 通过了图灵测试。』 | 'This AI passed the Turing test.' | 定义问题 | Definition problem | ① 图灵测试如何判定 ② 是否有严格测试 ③ 多为主观 | ① How is Turing test judged ② Strict test? ③ Mostly subjective | 『通过』定义不清,多数是宣传 | 'Passed' is unclear, mostly marketing |
| 『AI 让效率提升 10 倍(内部数据)。』 | 'AI boosts efficiency 10x (internal data). | 内部数据不可验 | Internal data unverifiable | ① 问样本 ② 问基线 ③ 问时间 ④ 是否可复现 | ① Ask sample ② Ask baseline ③ Ask period ④ Reproducible? | 『内部数据』是宣传话术 | 'Internal data' is marketing talk |
| 『AI 的准确率 99.9%。』 | 'AI's accuracy is 99.9%.' | 测试集问题 | Test set problem | ① 问测试集 ② 问是否独立 ③ 看基线 | ① Which test set ② Independent? ③ See baseline | 99.9% 可能在小数据集上 | 99.9% may be on small dataset |
| 『AI 会消灭 90% 的程序员。』 | 'AI will eliminate 90% of programmers.' | 极端 + 无源 | Extreme + no source | ① 问来源 ② 查真实数据 ③ 看『消灭』定义 | ① Ask source ② Check real data ③ 'Eliminate' definition | 90% 是极端,无可靠来源 | 90% is extreme, no reliable source |
| 『AI 的能耗是核电站的 3 倍。』 | 'AI's energy use is 3x a nuclear plant.' | 对比不当 | Improper comparison | ① 问对比口径 ② 查实际能耗 ③ 时间维度 | ① Compare basis ② Check actual energy ③ Time dimension | 对比口径不清,可能夸大 | Comparison basis unclear, may exaggerate |
| 『AI 让 80% 的企业转型成功。』 | 'AI makes 80% of enterprises transform successfully.' | 成功率无源 | Success rate no source | ① 问研究 ② 问定义『成功』 ③ 查数据 | ① Ask study ② Define 'success' ③ Check data | 80% 无来源,存疑 | 80% no source, doubtful |
| 『AI 的幻觉率只有 1%。』 | 'AI's hallucination rate is only 1%.' | 测试方法问题 | Test method problem | ① 问测试方法 ② 问测试集 ③ 看是否独立 | ① Test method ② Test set ③ Independent? | 1% 取决于测试,可能偏低 | 1% depends on test, may be low |
| 『AI 比人类医生诊断更准。』 | 'AI diagnoses more accurately than human doctors.' | 需具体场景 | Needs specific scenario | ① 问哪个病 ② 问对比对象 ③ 查研究 | ① Which disease ② Compare with ③ Check study | 某些场景 AI 准,但不是全面 | AI accurate in some scenarios, not across the board |
| 『AI 让学习速度提升 5 倍。』 | 'AI speeds up learning 5x.' | 5 倍无上下文 | 5x no context | ① 问什么学习 ② 问基线 ③ 查数据 | ① What learning ② Baseline ③ Data | 5 倍无上下文不可用 | 5x unusable without context |
| 『AI 的月活突破 10 亿。』 | 'AI's monthly active passed 1B.' | 需验证口径 | Need verify basis | ① 问统计口径 ② 查官方 ③ 是否含子产品 | ① Statistical basis ② Official ③ Includes sub-products? | 10 亿可能是累计非月活 | 1B may be cumulative not MAU |
| 『AI 会取代所有写作工作。』 | 'AI will replace all writing jobs.' | 极端 + 无源 | Extreme + no source | ① 问来源 ② 看『所有』 ③ 查真实数据 | ① Source ② 'All' ③ Real data | 『所有』是极端,无源 | 'All' is extreme, no source |
| 『AI 的推理能力超过人类。』 | 'AI's reasoning exceeds humans.' | 需定义 | Needs definition | ① 问哪类推理 ② 问测试 ③ 查研究 | ① What reasoning ② Test ③ Study | 某些推理 AI 强,但非全面 | AI strong in some reasoning, not across the board |
| 『AI 让 70% 的公司倒闭。』 | 'AI causes 70% of companies to fail.' | 极端 + 无源 | Extreme + no source | ① 问来源 ② 查数据 ③ 看时间 | ① Source ② Data ③ Time | 70% 极端,无源 | 70% extreme, no source |
| 『AI 的准确率在小样本上 100%。』 | 'AI's accuracy is 100% on small samples.' | 过拟合信号 | Overfitting signal | ① 问样本量 ② 问独立集 ③ 看是否过拟合 | ① Sample size ② Independent set ③ Overfit? | 小样本 100% 是过拟合信号 | 100% on small samples = overfitting signal |
| 『AI 的能耗每年翻倍。』 | 'AI's energy doubles yearly.' | 趋势需数据 | Trend needs data | ① 问数据 ② 查行业 ③ 看时间 | ① Data ② Industry ③ Time | 趋势可能,但需具体数据 | Trend possible, but needs specific data |
| 『AI 让 60% 的员工失业。』 | 'AI makes 60% of employees unemployed.' | 极端 + 无源 | Extreme + no source | ① 问来源 ② 查数据 ③ 看定义 | ① Source ② Data ③ Definition | 60% 极端,无源 | 60% extreme, no source |
| 『AI 的训练数据 1 万亿 tokens。』 | 'AI's training data is 1T tokens.' | 需验证 | Needs verification | ① 问来源 ② 查官方 ③ 看是否含重复 | ① Source ② Official ③ Includes duplicates? | 1T 可能,但需确认 | 1T possible, but needs confirmation |
| 『AI 比人类更快更准。』 | 'AI is faster and more accurate than humans.' | 需具体任务 | Needs specific task | ① 问什么任务 ② 问对比 ③ 查数据 | ① What task ② Compare ③ Data | 某些任务 AI 更快更准,但非全面 | AI faster/more accurate in some tasks, not across the board |
| 『AI 的幻觉率比人类低。』 | 'AI's hallucination rate is lower than humans.' | 需对比口径 | Needs comparison basis | ① 问如何测人类 ② 问测试 ③ 查研究 | ① How measure humans ② Test ③ Study | 『人类幻觉率』难量化,对比存疑 | 'Human hallucination rate' hard to quantify, comparison doubtful |
| 『AI 让 50% 的内容生产成本降了。』 | 'AI cuts content production cost by 50%.' | 需具体领域 | Needs specific field | ① 问什么内容 ② 问基线 ③ 查数据 | ① What content ② Baseline ③ Data | 50% 需具体领域和基线 | 50% needs specific field and baseline |
| 『AI 的参数量是 GPT 的 10 倍。』 | 'AI's parameters are 10x GPT.' | 需具体模型 | Needs specific model | ① 问哪个模型 ② 查参数 ③ 对比 | ① Which model ② Params ③ Compare | 需具体模型名才能比 | Need specific model name to compare |
| 『AI 让 90% 的广告变成 AI 生成。』 | 'AI makes 90% of ads AI-generated.' | 需验证 | Needs verification | ① 问来源 ② 查行业 ③ 看时间 | ① Source ② Industry ③ Time | 90% 需具体行业和时间 | 90% needs specific industry and time |
| 『AI 的准确率在医疗上 99%。』 | 'AI's accuracy in medicine is 99%.' | 需具体病种 | Needs specific disease | ① 问什么病 ② 问测试 ③ 查研究 | ① What disease ② Test ③ Study | 99% 需具体病种和测试 | 99% needs specific disease and test |
| 『AI 让编程效率提升 3 倍。』 | 'AI triples programming efficiency.' | 需具体任务 | Needs specific task | ① 问什么任务 ② 问基线 ③ 查数据 | ① What task ② Baseline ③ Data | 3 倍需具体任务和基线 | 3x needs specific task and baseline |
| 『AI 的能耗超过一个城市。』 | 'AI's energy exceeds a city.' | 需具体城市 | Needs specific city | ① 问哪个城市 ② 查能耗 ③ 时间 | ① Which city ② Energy ③ Time | 需具体城市和时间 | Needs specific city and time |
| 『AI 让 80% 的客服被取代。』 | 'AI replaces 80% of customer service.' | 需验证 | Needs verification | ① 问来源 ② 查数据 ③ 看时间 | ① Source ② Data ③ Time | 80% 需具体行业和时间 | 80% needs specific industry and time |
| 『AI 的推理成本降了 10 倍。』 | 'AI's inference cost dropped 10x.' | 需时间对比 | Needs time comparison | ① 问哪两年 ② 问口径 ③ 查数据 | ① Which years ② Basis ③ Data | 10 倍需时间对比 | 10x needs time comparison |
| 『AI 让 70% 的设计被 AI 做。』 | 'AI makes 70% of design AI-done.' | 需验证 | Needs verification | ① 问来源 ② 查行业 ③ 时间 | ① Source ② Industry ③ Time | 70% 需具体领域 | 70% needs specific field |
| 『AI 的准确率在数学上 95%。』 | 'AI's accuracy in math is 95%.' | 需具体题型 | Needs specific problem type | ① 问什么题 ② 问测试 ③ 查研究 | ① What problems ② Test ③ Study | 95% 需具体题型 | 95% needs specific problem type |
| 『AI 让 60% 的翻译被 AI 做。』 | 'AI makes 60% of translation AI-done.' | 需验证 | Needs verification | ① 问来源 ② 查数据 ③ 时间 | ① Source ② Data ③ Time | 60% 需具体领域和时间 | 60% needs specific field and time |
| 『AI 的幻觉率在 2026 年降了 50%。』 | 'AI's hallucination rate dropped 50% in 2026.' | 需时间对比 | Needs time comparison | ① 问哪两年 ② 问方法 ③ 查数据 | ① Which years ② Method ③ Data | 50% 需时间对比和方法 | 50% needs time comparison and method |
| 『AI 让 50% 的教育内容 AI 生成。』 | 'AI makes 50% of education content AI-generated.' | 需验证 | Needs verification | ① 问来源 ② 查行业 ③ 时间 | ① Source ② Industry ③ Time | 50% 需具体领域和时间 | 50% needs specific field and time |
| 『AI 的准确率在法律上 90%。』 | 'AI's accuracy in law is 90%.' | 需具体场景 | Needs specific scenario | ① 问什么场景 ② 问测试 ③ 查研究 | ① What scenario ② Test ③ Study | 90% 需具体场景 | 90% needs specific scenario |
| 『AI 让 80% 的代码由 AI 辅助。』 | 'AI makes 80% of code AI-assisted.' | 需验证 | Needs verification | ① 问来源 ② 查数据 ③ 时间 | ① Source ② Data ③ Time | 80% 需具体行业和时间 | 80% needs specific industry and time |
| 『AI 的幻觉率在医疗上 0。』 | 'AI's hallucination rate in medicine is 0.' | 不可能 | Impossible | ① 0 幻觉不存在 ② 质疑 ③ 查数据 | ① 0 hallucination doesn't exist ② Doubt ③ Data | 0 是虚假承诺,不可能 | 0 is a false promise, impossible |
30 个认知偏差(批判性思维的敌人)
30 Cognitive Biases (Enemies of Critical Thinking)
| 偏差 | Bias | 表现 | How it shows | 解药 | The cure |
|---|---|---|---|---|---|
| 确认偏误 | Confirmation bias | 只找支持自己观点的证据 | Seeking evidence supporting your own view | 主动找反证(红队思维) | Actively seek counter-evidence (red team thinking) |
| 锚定效应 | Anchoring effect | 被第一个数字带节奏 | Being dragged by the first number | 先独立判断再看参考值 | Judge independently before seeing reference values |
| 从众偏差 | Bandwagon bias | 别人都这么认为所以正确 | Correct because others think so | 独立验证,不因众而信 | Verify independently; not trust by crowd |
| 幸存者偏差 | Survivorship bias | 只看到成功者,忽略失败者 | Only seeing winners, ignoring losers | 同时统计失败案例 | Count failure cases simultaneously |
| 光环效应 | Halo effect | 因某人某方面优秀,认为他全对 | Seeing one excellence, assuming all correctness | 分开评估不同维度 | Evaluate different dimensions separately |
| 沉没成本 | Sunk cost | 已投入的不愿意止损 | Unwilling to cut losses due to past investment | 只看未来收益,忽略已投入 | Only look at future returns, ignore past investment |
| 过度自信 | Overconfidence | 高估自己判断的准确性 | Overestimating accuracy of your judgment | 校准:回顾历史判断正确率 | Calibrate: review historical judgment accuracy |
| 可得性偏差 | Availability bias | 容易想到的事觉得更常见 | Easy-to-remember things seem more common | 用数据而非印象判断 | Judge by data not impression |
| 虚假相关 | Spurious correlation | 相关不等于因果 | Correlation is not causation | 找因果机制,警惕巧合 | Find causal mechanism, watch for coincidence |
| 诉诸权威 | Appeal to authority | 专家说的就是对的 | Experts are always right | 看证据,不只看身份 | Look at evidence, not just identity |
| 诉诸情感 | Appeal to emotion | 用情绪代替逻辑 | Using emotion instead of logic | 情绪平静后再判断 | Judge after emotion calms |
| 虚假二分 | False dilemma | 只有 A 或 B,忽略中间 | Only A or B, ignoring middle | 找第三、第四种可能 | Find third, fourth possibilities |
| 滑坡谬误 | Slippery slope | A 会导致 B 会导致灾难 | A leads to B leads to disaster | 检验每一步因果是否成立 | Verify each causal step |
| 稻草人 | Straw man | 歪曲对方观点再反驳 | Distorting the other's view then refuting | 先复述对方观点再回应 | Restate the other's view before responding |
| 诉诸无知 | Appeal to ignorance | 没人证明假就是真 | Unproven false means true | 区分『未证明』与『已证明假』 | Distinguish 'unproven' from 'proven false' |
| 后此谬误 | Post hoc | A 在 B 前所以 A 导致 B | A before B so A caused B | 找机制和对照 | Find mechanism and control |
| 以偏概全 | Hasty generalization | 样本太小就下结论 | Small sample, big conclusion | 扩大样本,标注置信度 | Enlarge sample, mark confidence |
| 诉诸自然 | Appeal to nature | 天然的就是好的 | Natural is good | 天然不等于安全/正确 | Natural not equal safe/correct |
| 诉诸流行 | Appeal to popularity | 流行就是对的 | Popular is correct | 看证据,不看数量 | Look at evidence, not numbers |
| 动机偏差 | Motivated reasoning | 按动机推理,不按证据 | Reasoning by motive, not evidence | 问自己:如果证据相反我会信吗 | Ask: would I believe if evidence were opposite |
30 个判断案例:这条信息可信吗?
30 Judgment Cases: Is This Information Credible?
| 案例 | Case | 判断 | Verdict | 判断理由 | Reasoning |
|---|---|---|---|---|---|
| 『某专家说:喝 X 水能治百病』 | 'Expert says: drinking X cures everything' | 不可信 | Uncredible | 『治百病』= 过度承诺;无任何病能靠水治 | 'Cures everything' = over-claim; no disease is cured by water |
| 『某品牌: 99% 用户说好用(未说明样本)』 | 'Brand: 99% users satisfied (sample not stated)' | 存疑 | Questionable | 未说明样本量/筛选方式,典型营销话术 | Sample size/selection not stated; typical marketing talk |
| 『权威媒体报道: 某国 GDP 增长 5%(引用统计局)』 | 'Authoritative media: GDP up 5% (citing stats bureau)' | 可信 | Credible | 有明确数据源(统计局),可独立核实 | Clear data source (stats bureau), independently verifiable |
| 『朋友圈: 转发这条能给孩子积福』 | 'WeChat: forward this to accumulate merit for your child' | 不可信 | Uncredible | 无信息价值,纯情绪操纵 | No information value; pure emotional manipulation |
| 『某论文: 咖啡与长寿相关(n=10 万,对照实验)』 | 'Paper: coffee correlates with longevity (n=100k, controlled)' | 可信(保留) | Credible (reserved) | 样本大 + 对照,但『相关』非『因果』 | Large sample + control, but 'correlation' not 'causation' |
| 『某博主: AI 明天会毁灭人类』 | 'Blogger: AI will destroy humanity tomorrow' | 存疑 | Questionable | 无证据、时间模糊,恐惧驱动 | No evidence, vague timeline, fear-driven |
| 『官方数据: 2025 年失业率 5.1%(附调查方法)』 | 'Official: 2025 unemployment 5.1% (with survey method)' | 可信 | Credible | 附调查方法,透明可复核 | Survey method attached; transparent and checkable |
| 『某群: 内部消息,某股票明天涨停』 | 'Group: inside info, stock will limit-up tomorrow' | 不可信 | Uncredible | 无证据的『内部消息』= 典型骗局 | 'Inside info' without evidence = classic scam |
| 『研究: 睡眠不足与肥胖相关(meta 分析 30 研究)』 | 'Study: sleep deprivation correlates with obesity (meta of 30 studies)' | 可信(保留) | Credible (reserved) | 多研究荟萃,证据强度较高 | Multi-study meta; higher evidence strength |
| 『某网红: 我用某产品一个月瘦 10 斤』 | 'Influencer: I lost 10 jin in a month with this product' | 存疑 | Questionable | 个体案例,无对照,可能是广告 | Individual case, no control, possibly an ad |
| 『官方发布: 某地地震 5.0 级(附震中坐标)』 | 'Official: 5.0 earthquake (with epicenter coords)' | 可信 | Credible | 官方数据 + 具体参数,可独立验证 | Official data + specific params; independently verifiable |
| 『某公众号: 震惊!90% 的人不知道』 | 'Account: Shocking! 90% don't know this' | 存疑 | Questionable | 标题党,『震惊』= 情绪钩子 | Clickbait; 'shocking' = emotional hook |
| 『论文: X 药物有效(随机双盲,n=500,同行评审)』 | 'Paper: Drug X effective (RCT, n=500, peer-reviewed)' | 可信 | Credible | RCT + 同行评审,证据等级最高 | RCT + peer review; highest evidence level |
| 『某群: 转发此消息积功德』 | 'Group: forward this to accumulate merit' | 不可信 | Uncredible | 无信息价值,纯操纵 | No information value; pure manipulation |
| 『媒体报道: 某公司裁员 20%(引用内部文件)』 | 'Media: company cuts 20% (citing internal docs)' | 存疑 | Questionable | 有信源但单一,需多方交叉验证 | Has source but single; needs cross-verification |
| 『官方: 流感季请勤洗手(引用 WHO 指南)』 | 'Official: wash hands in flu season (citing WHO guidelines)' | 可信 | Credible | 引用权威指南,可查证 | Cites authoritative guideline; verifiable |
| 『某博主: AI 让我 3 个月赚 100 万』 | 'Blogger: AI made me 1M in 3 months' | 存疑 | Questionable | 无收入凭证,典型引流 | No income proof; typical funnel |
| 『研究: 运动与抑郁风险降低 20%(系统综述)』 | 'Study: exercise reduces depression risk 20% (systematic review)' | 可信(保留) | Credible (reserved) | 系统综述,证据强度较高 | Systematic review; higher evidence strength |
| 『朋友圈: 转发抽奖(非官方)』 | 'WeChat: forward to win lottery (unofficial)' | 不可信 | Uncredible | 非官方抽奖 = 高概率诈骗 | Unofficial lottery = high scam probability |
| 『媒体报道: 某法案通过(引用议会记录)』 | 'Media: bill passed (citing parliament records)' | 可信 | Credible | 引用官方记录,可查证 | Cites official records; verifiable |
| 『某群: 内部数据,某产品销量第一』 | 'Group: internal data, product #1 in sales' | 存疑 | Questionable | 无来源的『内部数据』,需独立核实 | 'Internal data' without source; needs independent check |
| 『论文: 疫苗安全性(多项 RCT + 长期随访)』 | 'Paper: vaccine safety (multiple RCTs + long follow-up)' | 可信 | Credible | 多 RCT + 长期随访,证据最充分 | Multiple RCTs + long follow-up; most sufficient evidence |
| 『某博主: 我用 AI 一天写 100 篇文章』 | 'Blogger: I write 100 articles a day with AI' | 存疑 | Questionable | 数量≠质量,可能灌水 | Quantity ≠ quality; possibly filler |
| 『官方: 空气质量 AQI 50(附站点数据)』 | 'Official: air quality AQI 50 (with station data)' | 可信 | Credible | 官方数据 + 具体站点,可实时验证 | Official data + specific stations; real-time verifiable |
| 『某群: 不转不是中国人』 | 'Group: not Chinese if you don't forward' | 不可信 | Uncredible | 道德绑架,纯操纵 | Moral coercion; pure manipulation |
| 『媒体报道: 某研究结果(引用论文 DOI)』 | 'Media: research result (citing paper DOI)' | 可信 | Credible | 给出 DOI 可查原文 | DOI given; original paper checkable |
| 『某博主: 我测过 100 款产品(无评测方法说明)』 | 'Blogger: I tested 100 products (no method stated)' | 存疑 | Questionable | 无评测方法,可能收广告 | No test method; possibly paid ads |
| 『官方: 疫苗接种率 80%(附分地区数据)』 | 'Official: vaccine rate 80% (with regional data)' | 可信 | Credible | 官方 + 分地区,可交叉验证 | Official + regional; cross-verifiable |
| 『某群: 紧急!限今日转发』 | 'Group: Urgent! Forward only today' | 不可信 | Uncredible | 制造紧迫感 = 操纵手法 | Creating urgency = manipulation tactic |
| 『论文: X 技术提升 30%(附代码开源)』 | 'Paper: Tech X improves 30% (code open-sourced)' | 可信 | Credible | 开源可复现,证据最硬 | Open-sourced and reproducible; hardest evidence |
信息源 10 级可信度分级(速查表)
Information Source 10-Level Credibility (Quick Table)
| 级别 | Level | 信息源类型 | Source Type | 典型例子 | Examples | 可信度 | Credibility | 使用建议 | Usage advice |
|---|---|---|---|---|---|---|---|---|---|
| L10 | L10 | 原始数据/开源代码 | Raw data / open-source code | 统计局原始数据、论文代码仓库 | Stats raw data, paper code repos | 最高(可复现) | Highest (reproducible) | 直接信任,但仍要读方法 | Trust directly; still read methods |
| L9 | L9 | 同行评审论文 | Peer-reviewed papers | 《Nature》《Science》论文 | Nature/Science papers | 很高 | Very high | 看样本和方法,注意相关≠因果 | Check sample & method; correlation ≠ causation |
| L8 | L8 | 官方/权威机构发布 | Official / authoritative releases | 政府公报、WHO、央行公告 | Govt bulletins, WHO, central bank | 高 | High | 可信,注意发布动机 | Credible; watch publishing motive |
| L7 | L7 | 主流媒体(交叉验证) | Mainstream media (cross-checked) | 多家权威媒体报道一致 | Multiple mainstream media agree | 较高 | Relatively high | 多家一致再信 | Trust when multiple agree |
| L6 | L6 | 主流媒体(单一来源) | Mainstream media (single source) | 某报独家 | Single newspaper exclusive | 中等 | Medium | 等第二来源验证 | Wait for second source |
| L5 | L5 | 专家个人观点 | Expert personal opinion | 某教授微博/访谈 | Professor's Weibo/interview | 中等(因人而异) | Medium (varies) | 看具体论证,不神化 | Judge by specific argument, don't deify |
| L4 | L4 | 自媒体(有专业背景) | Self-media (with expertise) | 专业领域博主 | Professional field bloggers | 中低 | Medium-low | 看是否给出来源和方法 | Check if sources & methods given |
| L3 | L3 | 自媒体(无专业背景) | Self-media (no expertise) | 泛流量账号 | General traffic accounts | 低 | Low | 只当线索,不当结论 | As leads only, not conclusions |
| L2 | L2 | 社交转发/群消息 | Social forwards / group messages | 朋友圈转发、群截图 | WeChat forwards, group screenshots | 很低 | Very low | 默认存疑,逐条核实 | Default questionable; verify each |
| L1 | L1 | 匿名/情绪化信息 | Anonymous / emotional info | 『震惊』『内部消息』『不转不是XX』 | 'Shocking', 'inside info', 'not XX if not forwarded' | 最低(默认假) | Lowest (default false) | 直接标记为不可信 | Mark uncredible directly |
20 招识别 AI 幻觉(AI 一本正经地胡说)
20 Techniques to Spot AI Hallucination (AI Speaking Nonsense Confidently)
| 招数 | Technique | 具体做法 | How to do it | 为什么有效 | Why it works |
|---|---|---|---|---|---|
| 交叉验证 | Cross-verify | 同一问题问 3 个不同 AI/搜索,看是否一致 | Ask 3 different AIs/searches the same question; check consistency | 幻觉常不一致;事实性内容多源一致 | Hallucinations often inconsistent; factual content agrees across sources |
| 要求引用来源 | Demand citations | 让 AI 给出具体来源(书/论文/数据),然后去核实 | Ask AI for specific sources (book/paper/data), then verify | 幻觉常给假来源;真来源可查到 | Hallucinations often give fake sources; real ones are findable |
| 查具体数字 | Check specific numbers | 对 AI 给出的数字/日期/人名,逐一独立核实 | Independently verify every number/date/name AI gives | 幻觉最爱编造具体细节 | Hallucinations love fabricating specific details |
| 问『你确定吗』 | Ask 'are you sure' | 直接问 AI 的确定度,看它是否承认不确定 | Ask AI's confidence; see if it admits uncertainty | 可靠 AI 会标注不确定;幻觉会硬撑 | Reliable AI flags uncertainty; hallucinations bluff |
| 要 AI 自我纠错 | Ask AI to self-correct | 让 AI 重新检查自己的回答,找矛盾 | Ask AI to re-check its answer for contradictions | 幻觉常有内部矛盾;事实通常自洽 | Hallucinations often have internal contradictions; facts are usually self-consistent |
| 看是否过度具体 | Watch over-specificity | AI 给出『第 3 页』『2019 年 5 月』等过度具体细节时警惕 | Be wary when AI gives over-specific details like 'page 3', 'May 2019' | 幻觉爱堆砌伪具体细节来装可信 | Hallucinations pile pseudo-specifics to seem credible |
| 查原始出处 | Check original source | AI 引『某书某页』,去查原书是否有此内容 | When AI cites 'a book, page X', check if the book has that content | 幻觉常『张冠李戴』引不存在的页 | Hallucinations often cite non-existent pages |
| 对比已知事实 | Compare known facts | 用你已知的常识/事实检验 AI 答案 | Test AI's answer against your known common sense/facts | 与常识矛盾 = 大概率幻觉 | Contradicts common sense = likely hallucination |
| 看语气是否绝对 | Check absolute tone | AI 用『绝对』『必然』『100%』等绝对词时警惕 | Be wary when AI uses absolute words like 'absolutely', 'certainly', '100%' | 可靠表述留余地;幻觉爱下绝对结论 | Reliable wording leaves room; hallucinations make absolute claims |
| 多轮追问 | Multi-round follow-up | 对 AI 答案层层追问,看是否前后矛盾 | Follow up AI's answer in layers; check for contradictions | 幻觉在追问下容易崩 | Hallucinations collapse under follow-up |
| 让 AI 列证据链 | Ask for evidence chain | 要求 AI 给出『结论的依据是什么』 | Ask AI for 'what's the basis of this conclusion' | 幻觉给不出证据链;事实可追溯 | Hallucinations can't give evidence chains; facts are traceable |
| 查数据时效 | Check data timeliness | AI 引用的数据是否过时(用了旧年份数据) | Check if AI's cited data is outdated (old-year data) | 幻觉/旧模型用过时数据 | Hallucinations/old models use outdated data |
| 要 AI 承认局限 | Ask AI to admit limits | 问 AI『这个问题你确定答案吗?有没有可能错』 | Ask AI 'are you sure of this answer? Could it be wrong' | 可靠 AI 会承认局限;幻觉会硬撑 | Reliable AI admits limits; hallucinations bluff |
| 交叉 AI 比对 | Cross-AI compare | 同一问题问不同模型(ChatGPT/Claude/DeepSeek),对比差异 | Ask different models the same question; compare differences | 幻觉在不同模型间不一致 | Hallucinations inconsistent across models |
| 查反例 | Find counterexamples | 问 AI『有没有反例/例外』 | Ask AI 'are there counterexamples/exceptions' | 幻觉给不出反例;事实有边界 | Hallucinations can't give counterexamples; facts have boundaries |
| 看是否编造人名/机构 | Check fabricated names/institutions | AI 给出的专家名/机构名,去搜是否存在 | Search if AI-given expert names/institutions exist | 幻觉常编造不存在的专家/机构 | Hallucinations often fabricate non-existent experts/institutions |
| 要求可复现步骤 | Ask for reproducible steps | 对方法类答案,要求给出可复现的具体步骤 | For method answers, ask for reproducible concrete steps | 幻觉给不出可操作细节;真方法可执行 | Hallucinations can't give actionable details; real methods are executable |
| 查逻辑链 | Check logic chain | 检验 AI 的推理链条,找逻辑断裂处 | Check AI's reasoning chain; find logic breaks | 幻觉常跳过关键推理步骤 | Hallucinations often skip key reasoning steps |
| 看是否过度自信 | Watch over-confidence | AI 语气越自信,越要警惕(『这很简单』『答案就是』) | The more confident AI's tone, the more wary ('this is simple', 'the answer is') | 自信度与正确度无关;幻觉特别自信 | Confidence ≠ correctness; hallucinations are especially confident |
| 最终核实关键结论 | Final verify key conclusions | 对任何关键结论,最后都用独立来源核实一次 | For any key conclusion, finally verify once with an independent source | 关键决策不能依赖单一 AI 来源 | Key decisions can't rely on a single AI source |
30 个实战场景:批判性思维怎么用
30 Real-World Scenarios: How to Apply Critical Thinking
| 场景 | Scenario | 常见陷阱 | Common trap | 批判性思维做法 | Critical thinking action | 效果 | Effect |
|---|---|---|---|---|---|---|---|
| 刷到健康养生文章 | Reading health/wellness articles | 『某食物治百病』绝对化 | 'Some food cures everything' absolutization | 查来源/查研究/看是否夸大 | Check source/study/overclaiming | 避免被养生谣言骗 | Avoid being fooled by wellness myths |
| 看到投资消息 | Seeing investment messages | 『稳赚不赔』『内部消息』 | 'Guaranteed profit', 'inside info' | 查资质/查风险/查谁受益 | Check qualifications/risk/who benefits | 避免被投资骗局骗 | Avoid being fooled by investment scams |
| 听领导/上级说法 | Hearing boss/superior claims | 权威=正确 | Authority = correct | 看数据,不当场否定但私下验证 | Look at data; don't refute on the spot but verify privately | 既尊重又不盲从 | Respect without blindly following |
| 读新闻报道 | Reading news reports | 单一信源下结论 | Concluding from single source | 找第二信源交叉验证 | Find second source for cross-verification | 避免被单侧信息误导 | Avoid being misled by one-sided info |
| 看产品评测 | Reading product reviews | 『亲测好用』无方法 | 'Personally tested, great' with no method | 看评测方法/样本/是否收广告 | Check test method/sample/whether paid ad | 做出真实购买决策 | Make real purchase decisions |
| 听朋友转述 | Hearing friends' retellings | 传话失真 | Message distortion | 问原始来源/直接核实 | Ask original source/verify directly | 避免二手信息偏差 | Avoid second-hand info bias |
| 看 AI 回答 | Reading AI answers | AI 自信=正确 | AI confidence = correct | 交叉验证/查来源/独立核实 | Cross-verify/check source/independently verify | 避免被 AI 幻觉带节奏 | Avoid being led by AI hallucinations |
| 做重大决策 | Making major decisions | 拍脑袋/情绪化 | Gut decision/emotional | 列证据/找反例/延迟决策 | List evidence/find counterexamples/delay decision | 降低决策失误 | Reduce decision errors |
| 面试听候选人说 | Hearing candidates in interviews | 『我做过很多大项目』空泛 | 'I've done many big projects' vague | 问具体细节/要案例/追问 | Ask concrete details/ask for cases/follow up | 识别真实能力 | Identify real ability |
| 看社交媒体热点 | Seeing social media hot topics | 情绪化跟风 | Emotional bandwagon | 冷静看事实/多方观点 | Stay calm, see facts/multiple views | 避免被情绪带节奏 | Avoid being led by emotion |
| 读论文 | Reading papers | 『相关』当『因果』 | Treating 'correlation' as 'causation' | 看方法/样本/因果论证 | Check method/sample/causal argument | 正确解读研究 | Correctly interpret research |
| 听专家访谈 | Hearing expert interviews | 专家全领域正确 | Experts correct in all fields | 看具体论证/区分领域 | Check specific argument/distinguish fields | 避免神化专家 | Avoid deifying experts |
| 看数据图表 | Reading data charts | 只看到想看的 | Only seeing what you want to see | 看样本/时间/口径/基线 | Check sample/time/caliber/baseline | 正确读数据 | Correctly read data |
| 做竞品分析 | Doing competitor analysis | 『他们不行』偏见 | 'They're not good' bias | 看客观数据/承认对手强项 | Look at objective data/acknowledge opponents' strengths | 客观评估竞争 | Objectively assess competition |
| 选专业/职业 | Choosing major/career | 『热门』=好 | 'Popular' = good | 看个人匹配/长期趋势/真实数据 | Check personal fit/long-term trends/real data | 避免追热点选错 | Avoid chasing trends and choosing wrong |
| 投资理财 | Investing | 『过去涨=未来涨』 | 'Past rise = future rise' | 看基本面/分散/风险 | Check fundamentals/diversify/risk | 降低投资风险 | Reduce investment risk |
| 看医疗建议 | Reading medical advice | 『偏方治大病』 | 'Folk cures for big diseases' | 查正规医学指南/咨询医生 | Check formal medical guidelines/consult doctor | 避免误诊误治 | Avoid misdiagnosis/mistreatment |
| 听销售话术 | Hearing sales talk | 『只剩最后 3 件』制造紧迫 | 'Only 3 left' creating urgency | 冷静判断真实需求/延迟 | Calmly judge real need/delay | 避免冲动消费 | Avoid impulsive buying |
| 做内容创作 | Doing content creation | 『大家都这么说』 | 'Everyone says so' | 独立核实/给出来源 | Independently verify/give sources | 做出可信内容 | Make credible content |
| 看政治言论 | Reading political statements | 情绪化站队 | Emotional taking-sides | 看事实/多方立场/证据 | See facts/multiple positions/evidence | 理性看待政治 | Politically rational |
| 评估自己工作 | Evaluating own work | 『我觉得做得好』自我高估 | 'I think I did well' self-overestimate | 找反馈/客观指标/对照 | Find feedback/objective metrics/benchmark | 客观自我评估 | Objective self-evaluation |
| 看广告 | Reading ads | 『全网销量第一』无证明 | 'No.1 sales online' without proof | 查数据来源/是否可验证 | Check data source/whether verifiable | 避免被广告带节奏 | Avoid being led by ads |
| 听孩子/学生说 | Hearing children/students say | 『其他同学都有』 | 'Other classmates all have' | 独立核实/不盲从 | Independently verify/don't blindly follow | 理性教育决策 | Rational education decisions |
| 做技术选型 | Doing tech selection | 『某框架最火』 | 'Some framework is hottest' | 看场景/长期维护/团队匹配 | Check scenario/long-term maintenance/team fit | 做出合适技术决策 | Make suitable tech decisions |
| 看新闻标题 | Reading news headlines | 标题党下结论 | Concluding from clickbait titles | 读全文/看原报道 | Read full text/see original report | 避免被标题误导 | Avoid being misled by headlines |
| 评估 AI 工具 | Evaluating AI tools | 『最强』营销话术 | 'Strongest' marketing talk | 看具体测试/自己场景实测 | Check specific tests/self-scenario actual test | 选出合适工具 | Choose suitable tool |
| 做商业判断 | Making business judgments | 『直觉』代替分析 | 'Intuition' replacing analysis | 列数据/找案例/算概率 | List data/find cases/calculate probabilities | 降低商业判断失误 | Reduce business judgment errors |
| 看养生/饮食建议 | Reading wellness/diet advice | 『某食物排毒』伪科学 | 'Some food detoxes' pseudoscience | 查正规医学/区分食物功效 | Check formal medicine/distinguish food effects | 避免被伪科学骗 | Avoid being fooled by pseudoscience |
| 听会议发言 | Hearing meeting remarks | 『一直如此』路径依赖 | 'It's always been this way' path dependence | 问『为什么』/找改进点 | Ask 'why'/find improvement points | 打破思维定式 | Break thinking patterns |
| 做学习决策 | Making learning decisions | 『某课程包过』承诺 | 'Some course guarantees passing' | 看真实学员反馈/课程质量 | Check real student feedback/course quality | 避免报假课程 | Avoid signing up for fake courses |
50 个『批判性思维』扩充问答
50 Extended 'Critical Thinking' Q&A
自测:你的 AI 学外语能力
Quiz: Your AI Language-Learning Ability
AI 学外语 20 项行动清单
AI Language-Learning 20-Item Action List
- 今天有没有『不验证就相信』过一条信息?Did you 'believe without verifying' a piece of info today?
- 今天的判断,证据等级够吗?Does today's judgment have enough evidence level?
- 有没有只看了『一边』的信息就下结论?Did you conclude after only seeing 'one side' of info?
- 决策前标注过情绪了吗?Did you label emotions before deciding?
- 有没有用『个例』代替『数据』?Did you use 'anecdote' instead of 'data'?
- AI 给的结论,追问过证据吗?Did you follow up on evidence for AI's conclusion?
- 信息有时效检查吗(哪年的信息)?Did you check info timeliness (which year)?
- 有没有识别到至少 1 个红旗信号?Did you spot at least 1 red-flag signal?
- 重要决策做过『预-mortem』吗?Did you do a 'pre-mortem' for important decisions?
- 有没有替反方说过最强论证(steelman)?Did you state the strongest counter-argument (steelman)?
- 判断给了置信度%吗?Did you assign a confidence % to the judgment?
- 有没有『相关≠因果』的陷阱?Any 'correlation ≠ causation' traps?
- 信源单一吗?(≥2 信源了吗)Single source? (Did you get ≥2 sources?)
- 有没有被『紧迫感』操纵过?Were you manipulated by 'urgency'?
- 决策日志记了吗?Did you record in the decision journal?
- 有没有『身份偏见』(因立场而信)?Any 'identity bias' (believing due to stance)?
- 样本量够吗?(<30 只算个例)Sample size enough? (<30 only counts as anecdote)
- 有没有『幸存者偏差』(只看成功的)?Any 'survivorship bias' (only seeing success)?
- 情绪是信号还是证据?分清了?Is emotion a signal or evidence? Did you separate them?
- 今天『会问』了还是只『会答』?Did you 'ask well' today or just 'answer'?
15 条 AI 学外语 FAQ
15 AI Language-Learning FAQ
参考来源
References
- 《思考,快与慢》(Thinking, Fast and Slow) — 丹尼尔·卡尼曼Daniel Kahneman, 'Thinking, Fast and Slow'
- 《批判性思维工具》(Critical Thinking Tools) — 理查德·保罗Richard Paul, 'Critical Thinking Tools'
- 逻辑谬误大全 (List of logical fallacies) — 维基百科List of logical fallacies — Wikipedia
- 可证伪性 (Falsifiability) — 维基百科Falsifiability — Wikipedia
- 证据等级 (Levels of evidence) — 维基百科Levels of evidence — Wikipedia
- 认知偏差 (Cognitive bias) — 维基百科Cognitive bias — Wikipedia
- 确认偏差 (Confirmation bias) — 维基百科Confirmation bias — Wikipedia
- 幸存者偏差 (Survivorship bias) — 维基百科Survivorship bias — Wikipedia
- 贝叶斯推理 (Bayesian inference) — 维基百科Bayesian inference — Wikipedia
- 第一性原理 (First principles) — 维基百科First principles — Wikipedia
- Our World in Data(数据可视化)Our World in Data (data visualization)
- FactCheck.org(事实核查)FactCheck.org (fact-checking)
- Khan Academy(逻辑课程)Khan Academy (logic courses)
- Coursera(批判性思维课程)Coursera (critical thinking courses)
- 《AI 幻觉与可靠性研究》综述 (AI hallucination & reliability surveys)Surveys on AI hallucination & reliability