🌍 世界数字教育网
模块 08 · 数字能力Module 08 · Digital Skills

AI 素养:数字时代的读写能力

AI Literacy: The New Literacy of the Digital Age

AI 素养 = 理解 AI 能做什么、不能做什么、如何安全高效地用它。

AI literacy = knowing what AI can and can't do, and how to use it safely and well.

🤖 30 大节 · 提示工程 · 幻觉 · 隐私 · 伦理 · Agent · 50 提示词 · 40 工具 · 30 误区 · 20 路径 · 40 术语 · 15 红线30 sections · Prompting · Hallucination · Privacy · Ethics · Agents · 50 prompts · 40 tools · 30 myths · 20 paths · 40 terms · 15 red lines
C1

什么是 AI 素养?为什么 2026 年人人必修

What is AI Literacy? Why It's a Core Skill in 2026

AI 素养 = 理解 AI 能做什么、不能做什么、如何安全高效地用它
AI literacy = knowing what AI can and can't do, and how to use it safely and well

2023 年,世界经济论坛把「AI 素养」列入未来工作技能清单;2024 年,欧盟《AI 法案》要求成员国把 AI 素养纳入义务教育;2025 年,全球 97% 的头部企业把「会用 AI」写进员工能力模型。AI 素养已经不是『加分项』,而是像读写能力一样的『基础能力』。

In 2023, the World Economic Forum listed AI literacy among future work skills; in 2024, the EU AI Act required member states to include AI literacy in compulsory education; in 2025, 97% of leading companies wrote 'can use AI' into their competency models. AI literacy is no longer a bonus — it is a basic capability, like reading and writing.

但『会用 AI』有三层含义:会用工具(打字问问题)、会用方法(提示工程+验证+组合)、会用思维(判断边界、守住伦理、保持自主思考)。大多数人停留在第一层,真正拉开差距的是第三层。

But 'using AI' has three layers: using tools (typing questions), using methods (prompting + verification + combining), and using thinking (judging boundaries, holding ethics, keeping independent thought). Most people stay at layer one; the real gap is created at layer three.

📊76%全球职场人 2024 年已在工作中使用过生成式 AI,但只有 18% 能正确识别其输出错误Worldwide Workers Report 2024 / Microsoft Work Trend Index
层级Layer定义Definition典型表现Typical behavior
第 1 层:操作L1: Operation会把 AI 当搜索引擎用Uses AI as a search engine『帮我写一段话』『翻译一下』'Write me a paragraph','translate this'
第 2 层:方法L2: Method掌握提示工程与验证Masters prompting & verification会拆解任务、给上下文、交叉验证Decomposes tasks, gives context, cross-checks
第 3 层:思维L3: Thinking理解原理、边界与伦理Understands principles, boundaries, ethics知道何时不用 AI、如何保持判断力Knows when NOT to use AI, keeps judgment
💡AI 素养的核心不是『学会一个工具』,而是『学会和 AI 协作的思维方式』。工具会过时(今年 ChatGPT、明年可能是别的),但协作思维不过时。The core of AI literacy is not 'learning one tool' but 'learning a way of thinking about working with AI'. Tools become outdated; the collaboration mindset does not.
⚠️警惕『工具崇拜』:把某个 App 当成 AI 本身。AI 是一类技术(大模型、多模态、Agent),不是某个产品。工具换了三代,素养不变。Beware 'tool worship': treating one app as AI itself. AI is a class of technology (LLMs, multimodal, agents), not one product. Tools have changed three generations; literacy stays.
C2

大语言模型是如何工作的(通俗版)

How Large Language Models Actually Work (Plain Edition)

不用公式,理解 LLM 的底层逻辑,你才能知道它何时可信、何时会骗你
Without formulas, understanding LLM fundamentals tells you when to trust and when it deceives

一句话原理

One-Sentence Principle

大语言模型本质上是一个『超级预测机』:它读过海量文本,学会了『前面这些字后面最可能接什么字』。它不是『理解』,而是『概率预测』。这个区别决定了它的一切优点和缺点。

A large language model is essentially a 'super-predictor': it has read massive text and learned 'what word most likely follows these words'. It does not 'understand' — it does 'probability prediction'. This distinction determines all its strengths and weaknesses.

概念Concept通俗解释Plain explanation对你意味着什么What it means for you
TokenToken文本切成的『字块』,中文约 1 字=1~2 个 tokenText split into chunks; Chinese ≈1 char = 1~2 tokens长文会超出模型的『上下文窗口』,需要分段喂Long text exceeds the context window — feed it in chunks
上下文窗口Context window模型一次能『记住』的文本总量,主流 12.8 万~100 万 tokenTotal text a model can 'remember' at once; mainstream 128K~1M tokens把关键信息放在开头和结尾,中间容易被忽略Put key info at the beginning and end; the middle gets ignored
温度Temperature控制输出随机性:0=最确定,1+ 更有创意Controls randomness: 0 = most deterministic, 1+ more creative要准确答案调低,要创意调高Lower for accuracy, higher for creativity
训练数据Training data模型『读过的书』,有截止日期The 'books' the model read — with a cutoff date它不知道截止日之后发生的事,可能给过时信息It doesn't know events after the cutoff date — may give outdated info
幻觉Hallucination模型编造不存在的事实,但语气很自信Model invents facts with confident tone必须对关键信息交叉验证Always cross-verify critical info
对齐Alignment训练模型『说人话、守规矩』的过程Training the model to 'speak human, follow rules'它被要求礼貌、拒绝危险请求——但也可能过度拒绝It's asked to be polite and refuse risky requests — but may over-refuse
🧠为什么 AI 会『一本正经地胡说』?因为它的目标是『生成看起来合理的文本』,不是『生成真实的事实』。当训练数据里某个说法不够多,它就用概率猜一个最像真的填进去。Why does AI 'confidently talk nonsense'? Because its goal is to 'generate plausible-sounding text', not 'generate true facts'. When training data is thin on a topic, it guesses the most plausible-sounding fill.
📊100B+主流大模型通常有超过 1000 亿参数,每个参数是一个可学习的『旋钮』
  • 『记住』不等于『理解』:模型可以背下整本词典,但不认识『苹果』是水果'Remembering' ≠ 'understanding': a model can memorize a whole dictionary but not know an apple is a fruit
  • 『上下文』是工作记忆:它不会把这次对话内容存进长期记忆(除非产品做了记忆功能)'Context' is working memory: it does not store this conversation into long-term memory (unless the product adds memory features)
  • 『概率』是双刃剑:让答案流畅自然,也让它敢编造'Probability' is a double-edged sword: it makes answers fluent and natural, but also makes it dare to fabricate
C3

AI 的能力边界:哪些事擅长、哪些事不行

AI Capability Boundaries: What It's Good At, What It Can't Do

一张表看清 AI 的能力地图,避免『什么都问』或『什么都不信』
One map to see AI's capability landscape — avoid 'ask it everything' or 'trust it nothing'
任务类型Task typeAI 表现Performance正确用法Right way
知识问答Knowledge Q&A⭐⭐⭐⭐ 强,但有幻觉风险Strong but hallucination risk问它当『起点』,再自己查证Ask it as a 'starting point', then verify yourself
写作起草Writing drafts⭐⭐⭐⭐ 快且流畅,但同质化Fast and fluent but homogenized让它出初稿,你改结构和灵魂Have it produce a first draft, you reshape the soul
代码生成Code generation⭐⭐⭐⭐ 常见库很熟,边缘场景会错Good with common libraries, wrong in edge cases逐段审查 + 跑测试,别整段照抄Review each segment + run tests, don't copy wholesale
翻译Translation⭐⭐⭐⭐ 通用文本好,专业术语要核对Good for general text, check technical terms通用直译 + 术语人工核对Direct translate + manual term check
数据分析Data analysis⭐⭐⭐ 会写代码但可能算错,要验证Writes code but may compute wrong, verify给它数据让它写代码,你验结果Give it data to write code, you verify the result
实时信息Real-time info⭐⭐ 训练有截止日期,除非联网Training has a cutoff, unless it has web access重大事件先查权威来源,再用 AI 整理For major events, check authoritative sources first, then let AI organize
数学/逻辑Math & logic⭐⭐⭐ 简单题强,多步推理易断链Strong on simple, breaks on multi-step reasoning复杂推理让它『一步步想』并自己复查For complex reasoning, have it 'think step by step' and double-check
创意Creativity⭐⭐⭐ 组合已有,不真正『原创』Combines existing, not truly 'original'用 AI 扩脑暴,人类定调Use AI to brainstorm, humans set the tone
情绪/心理Emotion & psychology⭐⭐ 会共情话术但无真正理解Empathy scripts but no real understanding不能替代专业心理咨询,仅辅助Cannot replace professional counseling, only auxiliary
个人决策Personal decisions⭐⭐⭐ 能给多角度分析,但不能替你负责Can give multi-angle analysis but can't take responsibility for you把它当『顾问团』,决定权在你Treat it as an 'advisory panel'; the decision is yours
🚫三条红线,AI 再聪明也别碰:① 医疗诊断(它可以查文献,不能下诊断);② 法律结论(它给参考,不能替代律师意见);③ 替人做重大决定(它没有你的价值观和处境)。Three red lines — no matter how smart the AI: ① medical diagnosis (it can look up literature, not diagnose); ② legal conclusions (it gives references, not lawyer opinions); ③ making major decisions for people (it has neither your values nor your circumstances).
📚15 万主流大模型训练语料约相当于 15 万本书的文本量,覆盖了几乎所有公开知识领域150K | Mainstream LLM training corpora ≈ the text of 150,000 books, covering nearly all public knowledge domains
❤️记住这句:AI 是『增强器』,不是『替代器』。你的判断力、价值观、处境,是它永远补不上的部分。用 AI 放大你已有的能力,而不是外包你的思考。Remember this: AI is an 'amplifier', not a 'replacer'. Your judgment, values, and circumstances are the parts it can never fill. Use AI to amplify what you already have, not to outsource your thinking.
C4

提示工程 10 个核心技巧(把 AI 用出 3 倍效果)

10 Core Prompting Techniques (Get 3× from AI)

提示不是『打字』,是『沟通工程』。这 10 招覆盖 90% 场景
Prompting is 'communication engineering'. These 10 cover 90% of scenarios
技巧Technique做法How to do it反例→正例Wrong → Right
1 给角色1. Role告诉 AI 它是谁:『你是资深数据分析师』Tell it who to be: 'You're a senior data analyst'『写报告』→『你是十年经验的财务顾问,写一份…』'write a report' → 'you're a 10-yr financial advisor, write...'
2 给上下文2. Context背景、目标、读者、限制一次给全Give background, goal, audience, constraints at once『翻译这句』→『这是给技术读者的论文摘要,正式体,译成…』'translate this' → 'this is a paper abstract for tech readers, formal tone, translate into...'
3 拆任务3. Decompose大任务拆成小步骤,逐个完成Break big tasks into small steps, one by one『帮我做完整营销方案』→『先列目标,再列渠道,最后写预算』'give me a full marketing plan' → 'first list goals, then channels, then budget'
4 给例子4. Examples少样本示例,让 AI 模仿格式Few-shot examples so AI mimics the format描述 10 遍不如给 2 个例子Describing 10 times < giving 2 examples
5 要步骤5. Chain-of-thought『请一步步思考并展示推理』'Please think step by step and show reasoning'直接要答案 → 要过程,错误率大降Direct answer → ask for the process, error rate drops a lot
6 定格式6. Format指定输出:表格/列表/JSON/MarkdownSpecify output: table/list/JSON/Markdown『整理这些』→『用三列表格:名称|价格|库存』'organize these' → 'in a 3-col table: name|price|stock'
7 要边界7. Boundaries告诉它『不知道就说不知道』Tell it 'say you don't know if you don't know'放任它编 → 明确要求标注不确定Letting it fabricate → explicitly require marking uncertainty
8 迭代8. Iterate别指望一次完美,追问和修改Don't expect perfection in one shot; follow up and revise一次不满意就放弃 → 基于上一轮追问Giving up after one unsatisfying answer → follow up on the last round
9 用变量9. Variables把重复的部分做成模板,只换变量Make repeated parts into templates, only swap variables每次重写长提示 → 做模板填空Rewriting long prompts each time → make a fill-in template
10 自我验证10. Self-check『请检查上面的回答有没有错误并修正』'Please check the answer above for errors and correct'直接采用 → 让它自审一遍Using it directly → have it self-review once
💡万能提示骨架:角色 + 背景 + 任务 + 格式 + 限制 + 示例。把六要素填进一句话,你的提示质量立刻超过 90% 的人。Universal prompt skeleton: Role + Context + Task + Format + Constraints + Examples. Fill the six elements into one sentence and your prompt quality instantly beats 90% of people.
⚠️最常见错误:提示太短。『帮我写个总结』不如『你是学术编辑,这是论文第 3 节,为期刊审稿人写 200 字摘要,语气中性,突出方法创新』。上下文越多,AI 越懂你。Most common error: prompts too short. 'Write me a summary' < 'You're an academic editor, this is Section 3, write a 200-word abstract for a journal reviewer, neutral tone, highlight method innovation'. The more context, the better AI understands you.
C5

幻觉检测与验证:8 种方法

Hallucination Detection & Verification: 8 Methods

AI 会编,但会验证的人不受伤害。这 8 招让它『说谎必被抓』
AI lies, but people who verify are not hurt. These 8 make it 'lie and get caught'
方法Method怎么做How to do it适用场景When to use
1 来源核对1. Source check让它给出处,你逐个打开验证Have it cite sources, you open and verify each学术、数据、历史事实Academic, data, historical facts
2 交叉验证2. Cross-verify同一个问题问 2~3 个模型,看是否一致Ask 2~3 models the same question, see if consistent关键结论、数字Critical conclusions, numbers
3 反事实测试2. Counterfactual故意问一个不存在的事实,看它是否硬编Deliberately ask about a non-existent fact, see if it fabricates测试某个模型的可靠性Testing a model's reliability
4 让它自证4. Self-prove『请给出推理过程和数据来源』'Please give the reasoning process and data sources'数学、逻辑、因果Math, logic, causality
5 缩小范围5. Narrow down模糊问题改具体,减少它发挥空间Turn vague questions specific, reduce its free play开放性大问题Open-ended big questions
6 要置信度6. Confidence『你对这个答案有多确定?0-100 分』'How confident are you? 0-100'需要判断可信度When you need to judge credibility
7 反向提问7. Reverse ask『这个说法有什么反例或争议?』'What counterexamples or controversies exist for this claim?'单一观点、争议话题Single viewpoints, controversial topics
8 事实/观点分离8. Fact/Opinion split让它标注哪些是事实、哪些是推断Have it label which are facts, which are inferences综合分析、预测Comprehensive analysis, predictions
🕵️识别幻觉的 5 个信号:① 引用了不存在的研究/论文;② 数字精确到小数点却无来源;③ 语气极度自信但内容可疑;④ 把不同领域的事实张冠李戴;⑤ 你追问细节时开始含糊或改口。5 signals of hallucination: ① cites non-existent studies/papers; ② precise-to-decimal numbers with no source; ③ extremely confident tone but suspicious content; ④ mixes facts from different fields; ⑤ starts being vague or reverses when you press on details.
📊40%在涉及冷门知识的问答中,大语言模型幻觉率可达 40% 以上,越是『它没学过』的领域越危险
❤️记住:AI 的『自信』和『正确』是两回事。它越流畅、越笃定,你越要警惕。把 AI 当『很勤奋但会看走眼的实习生』,而不是『权威专家』。Remember: AI's 'confidence' and 'correctness' are two different things. The more fluent and assured it is, the more you should be wary. Treat AI as 'a diligent intern who occasionally misreads', not 'an authority expert'.
C6

数据安全与隐私:给 AI 喂什么、不喂什么

Data Security & Privacy: What to Feed AI, What Not to

保护你的数据和隐私,是 AI 素养的底线
Protecting your data and privacy is the baseline of AI literacy
数据类型Data type能给 AI 吗?Can it be given to AI?原因Reason
公开信息Public info✅ 可以Yes本身已公开,无风险Already public, no risk
匿名化数据Anonymized data✅ 可以Yes去标识后风险低Low risk after de-identification
公司机密Company secrets❌ 不可以No会泄露到训练/日志中Leaked into training/logs
客户个人信息Customer PII❌ 不可以No违反隐私法规(GDPR/个保法)Violates privacy laws (GDPR/PIPL)
医疗/财务记录Medical/financial records❌ 不可以No高度敏感,需脱敏Highly sensitive, must be de-identified
内部未发布内容Internal unreleased content❌ 不可以No可能被竞争对手间接获取May be indirectly obtained by competitors
密码/密钥Passwords/keys❌ 绝对不Never任何情况下都不给Never, under any circumstance
未公开合同Unpublished contracts❌ 不可以No商业机密Trade secrets
🔒给 AI 喂数据前,先问三个问题:① 这是不是已公开的?② 泄露了会不会造成实际损失?③ 我能不能用假数据/匿名数据替代?三个问题都通过才喂。Before feeding data to AI, ask three questions: ① Is it already public? ② Would leaking it cause real damage? ③ Can I substitute fake/anonymized data? Only feed if all three pass.
📋企业级使用要点:① 选支持数据不用于训练的商用版/API;② 建立『可喂/不可喂』数据清单并培训员工;③ 对敏感项目用私有化部署;④ 定期审计 AI 访问日志。Enterprise usage points: ① Choose commercial versions/APIs that don't use your data for training; ② Build a 'feedable/non-feedable' data list and train staff; ③ Use private deployment for sensitive projects; ④ Regularly audit AI access logs.
  • 免费版 vs 商用版:免费版通常用你的数据改进模型,商用版承诺不用于训练(但要看条款)Free vs commercial: free versions usually use your data to improve the model; commercial versions promise not to train on it (but read the terms)
  • 本地部署最安全:数据不出内网,但成本高Local deployment is safest: data stays on the intranet, but costly
  • 提示词也算数据:你写的长提示里如果含机密,同样会泄露Prompts are also data: if your long prompt contains secrets, they leak too
C7

伦理与偏见:AI 不中立,它带着偏见出生

Ethics & Bias: AI Is Not Neutral — It's Born with Bias

理解 AI 的偏见来源,你才能不被它带偏
Understanding where AI bias comes from lets you not be led astray by it

AI 的偏见不是『坏』,而是『统计』:它从人类写的文本里学习,而人类文本本身充满偏见(性别、种族、地域、刻板印象)。所以 AI 会『继承』这些偏见,并在新场景里放大它们。这不是 AI 的错,是数据源的错,但后果由使用者承担。

AI bias is not 'bad', it's 'statistical': it learns from text written by humans, and human text is full of bias (gender, race, region, stereotypes). So AI 'inherits' these biases and amplifies them in new scenarios. This isn't AI's fault, it's the data source's fault, but the consequences are borne by the user.

偏见类型Bias type表现Manifestation如何缓解How to mitigate
性别偏见Gender bias『工程师』常联想男性,『护士』常联想女性'Engineer' often associated with male, 'nurse' with female要求多角度、检查代词使用Require multiple angles, check pronoun usage
地域偏见Regional bias对发达地区描述更正面,欠发达地区刻板More positive descriptions of developed regions, stereotypes for underdeveloped对比多地表述、查证来源Compare multi-region descriptions, verify sources
确认偏见Confirmation bias倾向给『你想听到的』答案Tends to give 'the answer you want to hear'让它先给反面论据Have it give counterarguments first
近因偏见Recency bias过度依赖训练数据里的近期事件Over-relies on recent events in training data标注时间、要求考虑长期趋势Mark time, require considering long-term trends
精英偏见Elitism bias偏好名校/知名机构,忽视草根方案Prefers elite schools/famous institutions, ignores grassroots solutions要求它列非主流方案Require it to list non-mainstream solutions
语言偏见Language bias英语内容质量明显高于其他语言English content quality clearly higher than other languages重要内容多语言交叉验证Cross-verify important content in multiple languages
⚖️用 AI 做涉及人的决策(招聘、评估、推荐)时要格外小心:① 把 AI 输出当『初筛』而非『终审』;② 定期用『反例测试』检查它是否系统性偏向某群体;③ 最终决定由人做出并承担。When using AI for people-related decisions (hiring, evaluation, recommendation) be extra careful: ① Treat AI output as 'pre-screening' not 'final review'; ② Periodically 'counterexample-test' whether it systematically favors a group; ③ Final decision is made by a human who bears responsibility.
📊3~5%研究显示,大语言模型在涉及职业、能力描述的文本中,性别刻板印象发生率比人类写作高 3~5 个百分点
❤️AI 素养的最高境界是『清醒』:知道它强在哪、弱在哪、偏见在哪,既不盲信也不盲拒。用它的长处,补它的短处,守它的底线。The highest level of AI literacy is 'clear-headedness': knowing where it's strong, weak, and biased — neither blind trust nor blind rejection. Use its strengths, patch its weaknesses, hold its bottom line.
C8

多模态 AI:文字之外的新能力

Multimodal AI: New Capabilities Beyond Text

2025 年 AI 已能看图、听音、生成视频、操作电脑。理解这些新能力,才能用好
In 2025, AI can already see, hear, generate video, and operate computers. Understanding these new capabilities lets you use them well
能力Capability2025 年成熟度2025 maturity典型应用Typical uses
图像理解Image understanding⭐⭐⭐⭐ 成熟Mature看图回答、OCR、图表解读Image Q&A, OCR, chart interpretation
图像生成Image generation⭐⭐⭐⭐ 成熟Mature设计、插画、产品图Design, illustration, product images
语音识别Speech recognition⭐⭐⭐⭐⭐ 非常成熟Very mature会议转写、字幕Meeting transcription, subtitles
语音合成Speech synthesis⭐⭐⭐⭐ 成熟Mature配音、朗读、客服Dubbing, reading, customer service
视频生成Video generation⭐⭐⭐ 快速成熟Rapidly maturing短视频、宣传片、演示Short videos, promos, demos
代码执行Code execution⭐⭐⭐⭐ 成熟Mature跑代码、数据分析、自动化Running code, data analysis, automation
Agent 操作Agent operations⭐⭐⭐ 成长中Growing自动浏览网页、填表、多步任务Auto browsing, form filling, multi-step tasks
实时对话Real-time conversation⭐⭐⭐⭐ 成熟Mature语音助手、陪练、面试模拟Voice assistant, sparring, interview simulation
🎨多模态的精髓是组合:用图像理解读图表,用文字生成写分析,用语音合成做讲解。一个任务串起三种能力,效果远超单一模态。The essence of multimodal is combination: image understanding to read charts, text generation to write analysis, speech synthesis to explain. Chaining three capabilities for one task far exceeds single-modality results.
⚠️AI 生成的图片和视频越来越逼真,识别 AI 生成内容成为新技能:看手部细节、文字渲染、物理一致性、元数据标记。很多平台已强制 AI 内容打水印。AI-generated images and videos are increasingly realistic. Identifying AI-generated content becomes a new skill: check hand details, text rendering, physical consistency, metadata marks. Many platforms now force watermarks on AI content.
C9

AI Agent:从工具到助手的跃迁

AI Agents: The Leap from Tool to Assistant

Agent 是 2025 年最大的趋势:AI 不再只回答,而是自主执行多步任务
Agents are the biggest trend of 2025: AI no longer just answers, but autonomously executes multi-step tasks

传统 AI 是你问它答。AI Agent 是你给目标,它自己拆解、规划、执行、检查、修正,像一个能动手的实习生。它能浏览网页、调用工具、写代码、运行程序、把结果交给你。

Traditional AI is you ask, it answers. An AI Agent is you give a goal, it decomposes, plans, executes, checks, and corrects by itself, like an intern who can actually work. It can browse the web, call tools, write code, run programs, and deliver results.

Agent 能力Agent capability示例Example人该做什么What the human should do
任务拆解Task decomposition帮我做行业报告,自动拆成搜资料、读、写、排'Make an industry report' auto-decomposes into search, read, write, format定目标、定验收标准Set goals, set acceptance criteria
工具调用Tool calling自动调搜索、计算器、APIAuto-calls search, calculator, APIs给它授权边界Give it authorization boundaries
自我检查Self-check写完自查格式、数据、逻辑After writing, self-checks format, data, logic抽查关键部分Spot-check key parts
记忆Memory记住之前对话的上下文和偏好Remembers prior conversation context and preferences定期校准它的记忆Periodically calibrate its memory
多 Agent 协作Multi-Agent collaboration一个搜索、一个写作、一个校对One searches, one writes, one proofreads设计分工和流程Design division of labor and workflow
🤖用 Agent 的正确方式:给它目标 + 边界 + 验收标准,而不是一步步指令。比如:调研 2025 年在线教育市场,找 5 家头部公司,输出对比表格,数据要标来源,不确定就标注。The right way to use Agents: give goal + boundaries + acceptance criteria, not step-by-step instructions. E.g., research the 2025 online education market, find 5 leading companies, output a comparison table, data must cite sources, mark uncertainty if unsure.
⚠️Agent 的风险:它会自信地走错路。多步任务中一步错,后面全错。所以关键节点必须人工检查。给 Agent 设检查点:每完成一个里程碑,停下来让你确认。Agent risk: it confidently goes down the wrong path. In multi-step tasks, one wrong step cascades. So key checkpoints must be human-checked. Set checkpoints for Agents: stop at each milestone for your confirmation.
C10

AI 素养的 7 个水平:你在哪一级?

7 Levels of AI Literacy: Which Level Are You At?

从完全不用到 AI 增强型专家,7 级阶梯
From never use to AI-augmented expert, a 7-level ladder
级别Level特征Characteristics典型行为Typical behavior
L0 排斥L0 Aversion认为 AI 没用或有害,从不使用Thinks AI is useless or harmful, never uses it只用传统方法,拒绝一切 AI 工具Uses only traditional methods, rejects all AI tools
L1 尝鲜L1 Novice偶尔用,当高级搜索引擎Occasionally uses, as an advanced search engine问简单问题,不深究Asks simple questions, doesn't dig deeper
L2 熟练L2 Proficient日常使用,掌握基本提示技巧Daily use, masters basic prompting techniques写邮件、翻译、写文案都用 AIUses AI for emails, translation, copywriting
L3 方法L3 Method掌握提示工程,会验证、会组合Masters prompting, verifies, combines拆解任务、交叉验证、多模型对比Decomposes tasks, cross-verifies, compares models
L4 理解L4 Understanding懂原理、知边界、能判断幻觉Understands principles, knows boundaries, judges hallucinations知道何时该用、何时不该用、何时要怀疑Knows when to use, when not to, when to doubt
L5 创造L5 Creation用 AI 创造新价值,设计工作流Creates new value with AI, designs workflows搭建 Agent 流程、自动化、产品原型Builds Agent workflows, automation, product prototypes
L6 专家L6 Expert能教别人、能评估模型、能定标准Can teach others, evaluate models, set standards做 AI 培训、模型评测、企业 AI 战略Does AI training, model evaluation, enterprise AI strategy
📊22%2025 年全球只有约 22% 的职场人达到 L3(方法级)以上,能到 L4(理解级)的不到 8%。你处于哪一级?22% | In 2025, only ~22% of workers worldwide reach L3 (method level) or above; fewer than 8% reach L4. Which level are you at?
📈升级路径:L1 到 L2(每天用 3 次,坚持一周);L2 到 L3(学提示工程,开始验证);L3 到 L4(读原理,理解边界);L4 到 L5(动手建工作流);L5 到 L6(输出、教学、定标准)。每级 1 到 2 个月可达。Upgrade path: L1 to L2 (use 3x daily for a week); L2 to L3 (learn prompting, start verifying); L3 to L4 (read principles, understand boundaries); L4 to L5 (build workflows hands-on); L5 to L6 (output, teach, set standards). Each level takes 1-2 months to reach.
🎯AI 素养不是考个证,是持续升级。模型每半年变一代,但你的素养升级速度决定你相对 AI 的竞争优势。今天 L3 的人,明年可能就是 L5。AI literacy isn't getting a certificate, it's continuous upgrading. Models change every six months, but your literacy upgrade speed determines your competitive edge relative to AI. Someone at L3 today could be at L5 next year.
C11

AI 素养自评:15 道题测测你自己

AI Literacy Self-Assessment: 15 Questions to Test Yourself

诚实作答,看看你处在哪一级,短板在哪
Answer honestly, see which level you're at and where your weaknesses are

快速自评(每题 1 到 3 分,满分 45)

Quick Self-Assessment (1-3 points each, max 45)

题目Question1 分1 pt3 分3 pts
我会用 AI 写邮件和文案I can use AI to write emails and copy会Can熟练且快Skilled and fast
我会给 AI 完整的背景上下文I give AI full background context偶尔Occasionally每次都给Every time
我会让 AI 一步步展示推理I have AI show step-by-step reasoning不会Never复杂题都会Always for complex
我会对 AI 的输出做交叉验证I cross-verify AI output不验证Don't关键信息都验证Verify all critical info
我能识别 AI 的幻觉(编造)I can identify AI hallucinations很少Rarely基本能Basically always
我了解大模型的基本原理I understand LLM basics不了解Don't能解释Can explain
我遵守数据安全红线(不喂机密)I follow data security red lines有时违反Sometimes从不违反Never violate
我会用多模态(图像/语音/视频)I use multimodal (image/audio/video)没用过Never经常用Frequently
我理解 AI 的偏见并会应对I understand AI bias and cope不理解Don't能识别并规避Can identify and avoid
我会设计 AI 工作流或 Agent 流程I design AI workflows or Agent flows不会Never能搭建Can build
我会评估不同模型的优劣I evaluate different models' pros and cons不评估Don't能对比Can compare
我能教别人用 AII can teach others to use AI不能Can't能系统教Can teach systematically
我知道 AI 能做和不能做的边界I know AI's do and don't boundaries模糊Vague清晰Clear
我警惕过度依赖 AI,保持自主思考I guard against over-reliance有时依赖Sometimes depend始终自主Always independent
我跟进 AI 新技术并应用到工作I follow AI new tech and apply to work不跟进Don't持续跟进Continuously follow
🎯12-20总分 12 到 20 分 = L2 熟练级;21 到 30 分 = L3 方法级;31 到 45 分 = L4+ 理解级。低于 12 分建议从基础提示技巧开始。12-20 pts = L2 proficient; 21-30 pts = L3 method; 31-45 pts = L4+. Below 12, start from basic prompting techniques.
💪找到你得分最低的 3 题,那就是你下一个月的修炼重点。AI 素养的提升不是多学一点,而是补短板。Find your 3 lowest-scoring questions, those are your focus for the next month. Improving AI literacy isn't learn a bit more, it's patch the weaknesses.
🧩

自测:你的 AI 素养水平

Quiz: Your AI Literacy Level

10 题,点选项即判对错 + 解析,诚实作答
10 questions, click to judge + explain, answer honestly
点击选项立即判分,答完看解析,可重置重做。
Click to grade instantly, see explanations, can reset.
大模型知道一个事实的方式是?How does a LLM 'know' a fact?
1. 记住一个数据库条目1. Remember a database entry1. 训练中学到的概率模式1. Probability pattern learned in training1. 实时联网查询1. Real-time web query
✅ 大模型没有知识数据库,它是从海量文本中学到的概率分布,所以会猜错而非查错。✅ An LLM has no knowledge database; it's a probability distribution learned from massive text, so it guesses wrong rather than looks up wrong.
上下文窗口指的是?What is the context window?
2. 模型训练的数据量2. Model training data size2. 模型一次能处理的文本长度2. Text length a model can process at once2. 模型的参数数量2. Number of model parameters
✅ 上下文窗口是模型一次装得下的文本总量,超出就要截断或分段。✅ The context window is the total text a model can hold at once; beyond it, truncation or chunking is needed.
以下哪种提示更能得到好答案?Which prompt yields a better answer?
3. 写个总结3. Write a summary3. 你是学术编辑,这是论文第3节,为审稿人写200字摘要,突出方法创新3. You're an academic editor, this is Section 3, write a 200-word abstract for a reviewer, highlight method innovation3. 总结一下上面3. Summarize the above
✅ 角色+背景+任务+格式+限制,上下文越全,AI 越懂你。✅ Role+context+task+format+constraints — the more complete the context, the better AI understands you.
AI 回答引用了一篇 2024 年的论文,你该怎么做?AI cites a 2024 paper, what should you do?
4. 直接采用,AI 不会错4. Use it directly, AI doesn't err4. 打开论文验证它是否真实存在、内容是否被正确引用4. Open the paper to verify it exists and is correctly cited4. 再问 AI 一次看它怎么说4. Ask AI again to see what it says
✅ 引用是幻觉高发区,必须人工验证。问 AI 自己不能验证 AI。✅ Citations are a high-hallucination zone; manual verification is required. Asking AI itself can't verify AI.
温度调高的作用是?The effect of raising temperature is?
5. 输出更确定、更保守5. More deterministic, conservative output5. 输出更有随机性、更创意5. More random, creative output5. 输出速度变快5. Faster output speed
✅ 温度控制随机性:低=确定,高=创意。要准确调低,要灵感调高。✅ Temperature controls randomness: low=deterministic, high=creative. Lower for accuracy, higher for inspiration.
下列哪种数据绝对不该喂给 AI?Which data should absolutely NOT be fed to AI?
6. 已公开发布的新闻6. Already published news6. 客户的手机号和身份证号6. Customer phone numbers and ID numbers6. 匿名的统计数据6. Anonymized statistics
✅ 个人信息违反隐私法规,喂给 AI 等于泄露。密码密钥更是最最不能碰。✅ Personal information violates privacy laws; feeding it to AI equals leaking. Passwords/keys are even more off-limits.
识别 AI 幻觉的最可靠方法是?The most reliable way to detect AI hallucination is?
7. 相信它的语气7. Trust its tone7. 交叉验证:多个来源/多个模型对比,并查原始出处7. Cross-verify: compare multiple sources/models and check original sources7. 让它重复一遍看是否一致7. Have it repeat to see if consistent
✅ 交叉验证 + 原始出处核对最可靠。AI 自己重复只检验一致性,不检验真实性。✅ Cross-verification + checking original sources is most reliable. AI repeating itself only checks consistency, not truth.
AI 的偏见主要来自?AI bias mainly comes from?
8. AI 自身的想法8. AI's own thoughts8. 训练数据中人类文本的偏见8. Bias in human text in training data8. 用户的提问方式8. User's questioning style
✅ AI 从人类写的文本学习,人类文本自带偏见,AI 会继承并放大。✅ AI learns from human-written text, which carries bias; AI inherits and amplifies it.
用 AI 做招聘初筛时,最稳妥的做法是?Using AI for hiring pre-screening, the safest practice is?
9. 完全信任 AI 的结果9. Fully trust AI's results9. AI 初筛 + 人工终审 + 定期反例测试检查偏见9. AI pre-screen + human final review + periodic counterexample tests to check bias9. 只让 AI 决定,省人力9. Let AI decide only, save manpower
✅ AI 做涉及人的决策必须有人工终审,且要定期检测它是否系统性偏向某群体。✅ AI for people-related decisions must have human final review, and periodically check whether it systematically favors a group.
AI 素养的最高境界是?The highest level of AI literacy is?
10. 会用最多的 AI 工具10. Using the most AI tools10. 清醒:知强弱、知边界、知偏见,不盲信不盲拒10. Clear-headedness: know strengths/weaknesses, boundaries, bias — neither blind trust nor blind rejection10. 完全不用 AI,避免依赖10. Never use AI, avoid dependence
✅ AI 素养的核心是清醒的判断力,不是工具数量,也不是排斥。✅ The core of AI literacy is clear-headed judgment, not tool count, and not rejection.
CHECK

20 项 AI 素养自检清单

20-Item AI Literacy Self-Check List

逐项打勾,看看你离 L4 还差多少
Check each item, see how far you are from L4
  • ✓我会给 AI 完整的角色和背景I give AI full role and context
  • ✓我会把大任务拆成小步骤喂给 AII break big tasks into small steps for AI
  • ✓复杂推理我会让 AI 一步步展示过程For complex reasoning I have AI show step-by-step
  • ✓我会对关键信息做交叉验证I cross-verify critical info
  • ✓我会检查 AI 的引用是否真实存在I check AI citations actually exist
  • ✓我知道 AI 的训练截止日期,警惕过时信息I know AI's training cutoff, wary of outdated info
  • ✓我会让 AI 标注不确定和争议的地方I have AI mark uncertainty and controversy
  • ✓我不把机密和个人信息喂给 AII don't feed secrets/PII to AI
  • ✓我了解 token、上下文、温度的基本含义I understand basic meaning of token/context/temperature
  • ✓我会根据任务调温度和格式I adjust temperature and format by task
  • ✓我会识别 AI 的性别和地域偏见I recognize AI's gender/regional bias
  • ✓我会让 AI 给反面论据,避免确认偏见I have AI give counterarguments to avoid confirmation bias
  • ✓我用多模态(图像/语音)辅助学习I use multimodal (image/audio) to aid learning
  • ✓我会用 Agent 做多步任务并设检查点I use Agents for multi-step tasks with checkpoints
  • ✓关键节点我会人工复核 Agent 的输出I manually review Agent output at key nodes
  • ✓我警惕过度依赖,保持独立判断I guard against over-reliance, keep independent judgment
  • ✓我会对比 2 个以上模型再下结论I compare 2+ models before concluding
  • ✓我能向别人解释 AI 的基本原理I can explain AI basics to others
  • ✓我跟进 AI 新技术并应用到工作I follow AI new tech and apply to work
  • ✓我建立了自己的 AI 使用红线清单I built my own AI usage red-line list
已完成 0 / 20 项
0 / 20 done
❓

15 条高频疑问(AI 素养)

15 High-Frequency Questions (AI Literacy)

把最常被问到的 15 个问题讲透
Thoroughly explain the 15 most frequently asked questions
两者都有。它理解你的字面意图(通过概率模式),但不理解你的处境和价值观。把它当很会读空气但不了解你的实习生。Both. It understands your literal intent (via probability patterns) but not your circumstances and values. Treat it as an intern who reads the room well but doesn't know you.
通常会(看条款)。商用版/API 一般承诺不用于训练。敏感内容一律用商用版或本地部署。Usually yes (check terms). Commercial versions/APIs generally promise not to train on your data. For sensitive content, use commercial or local deployment.
因为温度大于 0 时有随机性。要一致答案,调低温度或加固定提示。Because randomness exists when temperature is above 0. For consistent answers, lower temperature or add fixed prompts.
不可信,这是幻觉高发信号。要求它给具体出处,再验证。Not credible — this is a high-hallucination signal. Require specific sources, then verify.
你。AI 没有法律责任主体资格,最终决策者和责任人是使用者。所以重大决定永远要自己判断。You. AI has no legal liability subject; the decision-maker and responsible party is the user. So major decisions always require your own judgment.
会用 AI 替代思考会退化,会用 AI 放大思考会进步。关键是:AI 处理重复和资料,你处理判断、创造、决策。Using AI to replace thinking atrophies; using AI to amplify thinking progresses. Key: AI handles repetition/materials, you handle judgment/creation/decisions.
取决于怎么用。让 AI 出答案是坏;让 AI 当苏格拉底老师引导思考是好。关键是培养验证和批判的习惯。Depends on how. Letting AI give answers is bad; letting AI be a Socratic teacher guiding thinking is good. Key is cultivating verification and critical habits.
通用文本准确率高,专业术语和法律文书要人工核对。正式文件建议 AI 初译加专业审校。General text has high accuracy; technical terms and legal documents need manual checking. For formal documents, recommend AI first draft plus professional review.
常见物体和场景识别很准,细节(文字、小物体、复杂结构)会出错。关键信息要人工复核。Common objects/scenes recognized accurately; details (text, small objects, complex structures) can err. Critical info needs manual review.
会操作,但有风险。给它最小权限、设检查点,关键操作(删除/支付/发送)必须人工确认。It will operate, but with risks. Give it minimum permissions, set checkpoints; key operations (delete/pay/send) must have manual confirmation.
不需要。L1 到 L4 完全不需要编程。L5(建工作流)会写点脚本有帮助,但也不是必须。No. L1 to L4 requires no programming. L5 (building workflows) helps with some scripting, but not required.
三件事:学提示工程(给角色/背景/格式/例子);开始验证(引用/数字/关键结论);用多模态和 Agent。坚持 30 天。Three things: learn prompting (role/context/format/examples); start verifying (citations/numbers/key conclusions); use multimodal and Agent. 30 days of persistence.
不一定会取代你,但会用 AI 的人会取代不会用 AI 的人。把 AI 当放大器,放大你的不可替代部分。It may not replace you, but people who can use AI will replace people who can't. Treat AI as an amplifier, amplifying your irreplaceable parts.
看四点:数据怎么处理(隐私条款);是否标注 AI 生成内容;有没有人工审核/纠错机制;用户评价里是否有人提到错误/幻觉。Check four points: how data is handled (privacy terms); whether AI-generated content is labeled; whether there's human review/correction; whether user reviews mention errors/hallucination.
有一些(如 Google/Microsoft 的 AI 基础认证),但更看重实际能力。证书是入门证明,不是能力证明。优先提升实操。Some exist (e.g., Google/Microsoft AI foundation certs), but actual ability matters more. Certificates are entry proof, not ability proof. Prioritize hands-on improvement.
📚

参考来源(AI 素养)

References (AI Literacy)

以下来源真实可查,用于验证本文关键论断
The following sources are verifiable, used to verify key claims in this article
  1. World Economic Forum - Future of Jobs Report 2025 (AI literacy as core skill): weforum.org
  2. Microsoft - Work Trend Index 2024 (76% of workers use generative AI): microsoft.com/worklab
  3. European Union - AI Act 2024 (AI literacy in compulsory education): artificialintelligenceact.eu
  4. OpenAI - GPT-4 Technical Report (model capabilities and limitations): openai.com
  5. Anthropic - Claude Model Card (bias and safety evaluation): anthropic.com
  6. Google - Gemini Technical Report (multimodal capabilities): deepmind.google
  7. Stanford HAI - AI Index Report 2025 (AI adoption statistics): aiindex.stanford.edu
  8. Pew Research Center - AI and the Public 2024 (public perception and trust): pewresearch.org
  9. OECD - AI and Future of Work 2025 (skills gap analysis): oecd.org
  10. Deloitte - State of Generative AI in Enterprise 2025: deloitte.com
  11. MIT Sloan - AI Literacy Framework 2024: mitsloan.mit.edu
  12. Brookings Institution - AI and Democracy 2024 (bias and misinformation): brookings.edu
  13. NIST - AI Risk Management Framework 2024: nist.gov
  14. Common Sense Media - AI and Children 2025: commonsensemedia.org
  15. IEEE - Ethics of AI 2024 (bias mitigation guidelines): ieee.org
  16. Accenture - The State of AI 2025 (enterprise adoption): accenture.com
  17. Harvard Business Review - AI Augmentation 2024: hbr.org
  18. Pew Research - Generative AI and the Public 2025: pewresearch.org
  19. McKinsey - The Economic Potential of Generative AI 2024: mckinsey.com
  20. World Bank - AI and Development 2025 (digital divide): worldbank.org
C15

50 条分场景提示词库(直接复制用)

50 Scenario-Based Prompts (Copy & Use)

按场景分类的 50 条实战提示词,覆盖 10 大场景
50 real-world prompts by scenario, covering 10 scenarios

一、写作类(8 条)

I. Writing (8)

场景Scene提示词模板Prompt
邮件Email你是职场沟通专家。根据要点写邮件:语气[正式/友好],收件人[谁],要点[列表],长度[字数],结尾[行动项]。先出初稿我确认再润色。You're a workplace comm expert. Write an email from these points: tone [formal/friendly], recipient [who], points [list], length [words], ending [action]. Draft first, I confirm before polishing.
摘要Summary你是学术编辑。把这段[领域]内容压成[字数]摘要,保留[关键信息],去掉[冗余],用[目标读者]能懂的语言。You're an academic editor. Compress this [field] content into a [words] summary, keep [key info], drop [redundancy], in [readers'] language.
改写Rewrite把这段改写为[正式/口语/营销]风格,面向[读者],突出[核心信息],[字数]内,保留原意换表达。Rewrite this as [formal/casual/marketing], for [readers], highlight [core info], within [words], keep meaning, change expression.
大纲Outline你是内容策划。为[主题]写[类型]大纲,分[几部分],每部分[要点数]个要点,逻辑[递进/并列/对比]。You're a content planner. Write a [type] outline for [topic], in [parts], each [count] points, logic [progressive/parallel/contrast].
标题Title给[内容]起[数量]个标题,风格[吸睛/专业/悬念],每个[字数]内,突出[卖点],避免标题党。Give [content] [count] titles, style [catchy/pro/suspense], each within [words], highlight [sell point], avoid clickbait.
故事Story用[风格]讲[主题]的[短]故事,主角[谁],冲突[什么],结局[怎样],[字数]内,有[情感点]。Tell a [style] short story on [topic], protagonist [who], conflict [what], ending [how], within [words], with [emotional beat].
文案Copy为[产品]写[渠道]文案,卖点[列表],目标[人群],痛点[什么],号召[行动],语气[调性],分[几段]。Write [channel] copy for [product], selling points [list], target [audience], pain [what], CTA [action], tone [tone], in [parts].
问答Q&A把[主题]整理成[数量]个高频问答,问题是[用户]会问的,答案[简洁/详细],每个[字数]内,附[来源]。Organize [topic] into [count] high-freq Q&A, questions [users] ask, answers [brief/detailed], each within [words], with [sources].

二、学习类(8 条)

II. Learning (8)

场景Scene提示词模板Prompt
苏格拉底Socratic你是苏格拉底式老师。别直接告诉[主题]答案,用提问引导我想出来。从[我的水平]开始,每次只问一个问题,等我回答再推进。You're a Socratic teacher. Don't give [topic] answer directly; use questions to lead me. Start from [my level], one question at a time, wait for my answer.
费曼Feynman用[比喻]给我解释[概念],假设我是[背景],先[核心]再[细节],最后[例子]收尾。我没懂就换[另一个比喻]。Explain [concept] with [metaphor], assume I'm [background], [core] then [details], end with [example]. If I don't get it, switch [another metaphor].
出题Quiz基于[材料/主题]出[数量]道[选择/判断/简答]题,难度[分布],每题附答案+解析,最后给总分+薄弱点分析。Based on [material/topic] create [count] [mcq/tf/short] questions, difficulty [dist], each with answer + explanation, finally total + weakness analysis.
计划Plan制定学[技能]的[天数]天计划,我[每天可用时间],起点[水平],目标[达到什么],按周分解,每天给具体任务。Make a [days]-day plan for [skill], I have [daily time], start [level], goal [achieve], break by week, each day a concrete task.
拆解Breakdown把[大主题]拆成[层级]知识树,第一层[几大块],每块[几小点],每小点[一句话定义],帮我理清结构。Break [big topic] into a [level]-tier knowledge tree, first tier [blocks], each [subpoints], each [one-sentence def], help me see structure.
连接Connect解释[概念A]和[概念B]的关系:相同点、不同点、一个怎么影响另一个,用[例子]说明,最后[一句话总结]。Explain relationship of [A] and [B]: similarities, differences, how one affects the other, with [example], finally [one-line summary].
检验Check我写了[主题]的理解:[内容]。指出哪里准确、哪里偏差、哪里缺失,按重要度排序,给修正建议。I wrote my understanding of [topic]: [content]. Point out accurate/off/missing, order by importance, give correction advice.
迁移Transfer我刚学[概念/方法],用[新场景]帮我应用,给步骤,预测可能踩的坑,告诉我怎么避免。I just learned [concept/method], apply to [new scenario] for me, give steps, predict pitfalls, tell me how to avoid.

三、工作类(10 条)

III. Work (10)

场景Scene提示词模板Prompt
PPTPPT为[主题]的[时长]演示写 PPT 大纲,[页数]页,每页[标题+3要点+讲稿1句],逻辑[故事线],风格[专业/活泼],末页[行动号召]。Write PPT outline for [topic] [duration] presentation, [slides] pages, each [title+3 points+1-line script], logic [storyline], style [pro/lively], last [CTA].
会议Meeting根据[纪要要点]整理:决议[列表]、待办[负责人+截止]、风险[列表]、下次议程[预排],用[格式]输出。Organize from [minutes]: decisions [list], to-dos [owner+deadline], risks [list], next agenda [pre-set], output in [format].
报告Report写[主题]的[类型]报告,结构[背景-发现-分析-建议],每部分[字数],数据[来源],结论[明确/谨慎],给[读者]看。Write a [type] report on [topic], structure [bg-find-analysis-reco], each [words], data [source], conclusion [clear/cautious], for [readers].
分析Analysis用[SWOT/5W/PEST]分析[对象],每维度[几点],基于[事实/数据],最后给核心结论+2-3 个行动建议。Analyze [object] with [SWOT/5W/PEST], each dimension [points], based on [facts/data], finally core conclusion + 2-3 action advice.
邮件回复Email reply这是[对方邮件]。帮我回复:立场[同意/拒绝/协商],语气[专业/温和],要点[列表],长度[字数],结尾[下一步]。Here's [other's email]. Reply: stance [agree/refuse/negotiate], tone [pro/gentle], points [list], length [words], ending [next step].
谈判Negotiation我[角色],对方[角色],目标[什么],底线[什么],对方诉求[什么]。给策略+话术,预测对方反应和我的应对。I'm [role], other [role], goal [what], bottom line [what], their demands [what]. Give strategy + scripts, predict their reaction + my response.
提案Proposal为[方案]写提案:问题[现状]、方案[做法]、收益[量化]、成本[预算]、风险[预案]、时间表[里程碑],面向[决策者],突出[ROI]。Write proposal for [plan]: problem [now], plan [how], benefits [quant], cost [budget], risks [contingency], timeline [milestones], for [decision-maker], highlight [ROI].
复盘Review复盘[项目/事件]:目标[当初]、结果[实际]、差距[原因]、做对的[亮点]、做错的[教训]、下次[改进],用[5W/AAR]格式。Review [project/event]: goal [orig], result [actual], gap [cause], right [highlights], wrong [lessons], next [improve], in [5W/AAR] format.
招聘Hiring为[岗位]写招聘描述:职责[列表]、要求[硬/软]、加分[项]、薪资[范围]、吸引力[点],语气[调性],避免歧视性表述。Write JD for [role]: duties [list], requirements [hard/soft], plus [items], salary [range], attraction [points], tone [tone], avoid discriminatory language.
客户Customer客户[投诉/问题]是[内容]。回复:先共情,再解释,给解决方案+补偿,结尾挽回,语气专业,[字数]内。Customer [complaint/issue] is [content]. Reply: empathize, explain, solution + compensation, retain ending, professional, within [words].

四、生活类(8 条)

IV. Life (8)

场景Scene提示词模板Prompt
旅行Travel规划[天数]天[目的地]旅行:预算[多少],同行[谁],兴趣[类型],每天给行程+预算+贴士,含交通+住宿+餐饮,避开[不想去]。Plan [days]-day [destination] trip: budget [amount], companions [who], interests [type], each day itinerary + budget + tips, incl. transport + lodging + food, avoid [don't want].
健身Fitness制定[天数]天[减脂/增肌]计划,我[基础体能],每周[几次],每次[多久],给具体动作+组数+强度,附饮食建议。Make a [days]-day [cut/bulk] plan, my [base fitness], [times]/week, each [duration], specific moves + sets + intensity, with diet advice.
理财Finance分析我的[收支情况],给预算分配建议,紧急备用金该留[多少],负债怎么还最优,投资怎么起步,基于[风险偏好],不承诺收益。Analyze my [income/expense], budget allocation advice, emergency fund [amount], debt repayment optimal, investment start, based on [risk preference], no income promises.
亲子Parenting孩子[年龄],遇到[问题],给我沟通话术+引导方法,避免指责/包办,培养自主/习惯,给具体例子。Child [age], faces [issue], give communication scripts + guidance, avoid blame/doing-for-them, cultivate autonomy/habit, with examples.
社交Social帮我准备[场合]的开场+话题,对象[谁],我想表达/了解什么,给 3 个话题+接话技巧,避免尴尬话题。Prepare [opening/topics] for [occasion], other is [who], I want to express/learn, 3 topics + follow-up tips, avoid awkward topics.
烹饪Cooking用[现有食材]设计[几道]菜,口味[偏好],[时间]内完成,给步骤+火候+调味,适合[几人],附营养。Design [dishes] with [ingredients], taste [preference], within [time], steps + heat + seasoning, for [people], with nutrition.
搬家Moving规划[城市A到B]搬家:预算[多少],物品量,时间,给清单+步骤+时间线+成本拆解,含打包+运输+入住。Plan [city A to B] moving: budget [amount], item volume, time, checklist + steps + timeline + cost breakdown, incl. packing + transport + settling.
健康Health我[症状/困扰],基于一般知识给可能方向+自查建议,提醒我该看医生,不诊断不开药,给日常调理建议。I have [symptoms], based on general knowledge give possible directions + self-check, remind me to see a doctor, no diagnosis/prescription, daily care advice.

五、创作类(6 条)

V. Creative (6)

场景Scene提示词模板Prompt
脑暴Brainstorm为[主题]脑暴[数量]个创意/方案/名字,风格多样,从不同角度,每个给一句话说明,最后挑最可行的 3 个并说明理由。Brainstorm [count] ideas/plans/names for [topic], diverse style, different angles, each with one-line desc, finally pick top 3 feasible with reasons.
结构Structure帮我把散乱的点整理成[文章/演讲/视频],逻辑主线,起承转合安排,每部分字数/时长,突出核心。Organize scattered points into [article/speech/video], logic mainline, setup-development-turn-conclusion, each part words/duration, highlight core.
润色Polish润色这段文字:提升流畅度/感染力/专业度,保留核心意思,去掉口语/冗余,调整节奏,控制在[字数],语气[调性]。Polish this text: improve flow/appeal/professionalism, keep core meaning, remove colloquial/redundant, adjust rhythm, within [words], tone [tone].
视频脚本Video script为[时长]视频写脚本:钩子开头 5 秒、主体 3 段、结尾号召,每段给画面+台词+时长,节奏快/慢,风格[调性]。Script for [duration] video: hook first 5s, body 3 segments, ending CTA, each gives visual + line + duration, pace fast/slow, style [tone].
歌曲Song写一首[风格]歌,主题[什么],[几段],押韵,情绪变化,每段[字数],副歌要抓耳,给意象。Write a [style] song, theme [what], [verses], rhyme, emotion shift, each [words], catchy chorus, imagery.
视觉Visual描述[场景]的视觉风格:色调、构图、光线、主体、氛围,给 3 个变体,用于 AI 绘图提示。Describe [scene] visual style: color, composition, light, subject, mood, 3 variants, for AI image prompt.

六、研究类(6 条)

VI. Research (6)

场景Scene提示词模板Prompt
综述Review综述[主题]:[年份]以来的主要进展,分几方向,每方向给代表工作+结论,最后给趋势+空白,标注来源。Review [topic]: main progress since [year], in directions, each gives representative work + conclusion, finally trends + gaps, mark sources.
对比Compare对比 A 和 B:维度[列表],每维度谁优谁劣+原因,基于事实,最后给适用场景建议,用表格。Compare A and B: dimensions [list], each who's better + why, based on facts, finally use-case advice, in table.
溯源Trace追溯[概念/方法]的起源:谁提出何时,发展脉络,关键节点,当前状态,标注来源,区分事实和推测。Trace [concept/method] origin: who proposed when, development, key nodes, current state, mark sources, distinguish fact and speculation.
批判Critique批判这个观点/研究:论证是否成立,证据是否充分,逻辑有无漏洞,反例,最后给修正版,标注来源。Critique this view/research: is argument sound, evidence sufficient, logic gaps, counterexamples, finally revised version, mark sources.
方法Method解释[研究方法]:原理、步骤、适用、局限、常见错误,用例子,给使用建议,标注来源。Explain [research method]: principle, steps, applicable, limitations, common errors, with examples, usage advice, mark sources.
数据Data分析[数据]:关键指标、趋势、异常、可能原因,给可视化建议,标注数据来源,区分相关和因果。Analyze [data]: key metrics, trends, anomalies, possible causes, visualization advice, mark data source, distinguish correlation and causation.
💾这 50 条是模板不是答案。把方括号换成你的具体信息,提示质量立刻提升。存进自己的提示库,按场景分类,用时即取。These 50 are templates, not answers. Replace brackets with your specific info, prompt quality instantly improves. Save into your own prompt library, categorized by scenario, grab when needed.
⚠️用提示词库的原则:① 上下文越具体,输出越准;② 一次一个任务,别贪多;③ 关键输出一定验证;④ 模板是起点,按效果迭代。Principles for prompt library: ① more specific context = more accurate output; ② one task at a time; ③ always verify critical output; ④ templates are starting points, iterate by effect.
C16

30 个 AI 素养学习资源(真实可查)

30 AI Literacy Learning Resources (Verifiable)

从入门到进阶的 30 个真实资源,覆盖课程/书籍/工具/社区
30 real resources from beginner to advanced, covering courses/books/tools/communities
DeepLearning.AI Specialization 免费
吴恩达的 AI 入门系列,含提示工程、LLM 应用,有证书
Andrew Ng's AI intro series, incl. prompting, LLM apps, with cert
https://www.coursera.org/specializations/deep-learning
Google AI Essentials 免费
Google 官方 AI 基础课,2 小时入门,含提示工程实践
Google's official AI basics course, 2-hour intro, incl. prompting practice
https://grow.google/ai/
Microsoft AI Skills 免费
微软 Learn 平台 AI 模块,含 Copilot、Azure AI,可获证书
Microsoft Learn AI modules, incl. Copilot, Azure AI, cert available
https://learn.microsoft.com/training/ai
Khan Academy AI 免费
可汗学院 AI 基础,适合零基础,数学基础同步
Khan Academy AI basics, for zero-base, math foundation synced
https://www.khanacademy.org/computing
edX CS191 免费
哈佛计算思维课,理解 AI 底层逻辑,学术严谨
Harvard computational thinking, understand AI underlying logic, academically rigorous
https://www.edx.org/
fast.ai Practical Deep Learning 免费
实用深度学习,从做项目学原理,代码实战
Practical deep learning, learn principles from projects, code hands-on
https://course.fast.ai/
Stanford CS229 免费
斯坦福机器学习经典课,数学原理,进阶
Stanford classic ML course, math principles, advanced
https://cs229.stanford.edu/
AI4ALL 免费
AI 素养公益课程,面向非技术人群,通俗易懂
AI literacy public-interest course, for non-technical, plain language
https://ai4all.org/
Google AI Ethics 免费
Google AI 伦理课,偏见/公平/责任,案例教学
Google AI ethics course, bias/fairness/responsibility, case-based
https://ai.google/ethics/
MIT Sloan AI for Executives 部分免费
MIT 高管 AI 课,战略视角,商业应用
MIT exec AI course, strategic view, business application
https://mitsloan.mit.edu/ai
DeepLearning.AI Prompt Engineering 免费
专门教提示工程的短课,实操强
Dedicated prompt engineering short course, strong practice
https://www.deeplearning.ai/short-courses/
Anthropic Prompt Engineering Guide 免费
Anthropic 官方提示工程指南,系统性
Anthropic official prompt engineering guide, systematic
https://docs.anthropic.com/
OpenAI Cookbook 免费
OpenAI 官方代码示例,API 应用,开发者
OpenAI official code examples, API applications, developers
https://cookbook.openai.com/
LangChain Tutorials 免费
LLM 应用开发教程,RAG/Agent,工程实践
LLM app development tutorials, RAG/Agent, engineering practice
https://python.langchain.com/
Hugging Face Course 免费
NLP/多模态课程,含模型使用,社区活跃
NLP/multimodal course, incl. model usage, active community
https://huggingface.co/learn
DataCamp AI Track 部分免费
数据科学+AI 综合路径,项目制
Data science + AI comprehensive path, project-based
https://www.datacamp.com/
Coursera Generative AI 部分免费
多校合开的生成式 AI 课,前沿应用
Multi-school generative AI course, frontier applications
https://www.coursera.org/
YouTube 3Blue1Brown 免费
用动画讲 AI/数学原理,直观易懂
Explains AI/math principles with animation, intuitive
https://www.youtube.com/@3blue1brown
Two Minute Papers 免费
AI 新进展快速解说,保持跟进
Quick explainers of AI new progress, stay current
https://www.youtube.com/@TwoMinutePapers
AI Explained (Matt Wolfe) 免费
每周 AI 工具/趋势汇总,实用导向
Weekly AI tools/trends roundup, practical
https://www.youtube.com/@AIExplained
Sequoia AI Podcast 免费
AI 投资/产业深度访谈,战略视角
AI investment/industry in-depth interviews, strategic view
https://www.sequoiacap.com/podcast/
Latent Space Podcast 免费
LLM 技术深度播客,开发者
LLM technical deep podcast, developers
https://www.latent.space/
Hacker News AI 版块 免费
技术社区,AI 讨论,一手信息
Tech community, AI discussion, first-hand info
https://news.ycombinator.com/
arXiv cs.AI 免费
AI 论文库,最新研究,学术前沿
AI paper library, latest research, academic frontier
https://arxiv.org/list/cs.AI/recent
Papers with Code 免费
论文+代码+基准,可复现研究
Papers + code + benchmarks, reproducible research
https://paperswithcode.com/
Hugging Face Open LLM Leaderboard 免费
开源模型对比基准,选型参考
Open model comparison benchmarks, selection reference
https://huggingface.co/spaces/lmsys
Artificial Intelligence Act (EU) 免费
欧盟 AI 法案全文,监管/伦理依据
EU AI Act full text, regulation/ethics basis
https://artificialintelligenceact.eu/
NIST AI Risk Management Framework 免费
美国 AI 风险管理框架,企业合规
US AI risk management framework, enterprise compliance
https://www.nist.gov/itl/ai-risk-management-framework
Stanford AI Index Report 免费
年度 AI 发展数据报告,权威统计
Annual AI development data report, authoritative stats
https://aiindex.stanford.edu/
Our World in Data - AI 免费
AI 与社会的开放数据,可视化
AI and society open data, visualization
https://ourworldindata.org/
📚资源使用路径:① 入门先 Google AI Essentials(2 小时);② 提示工程看 DeepLearning.AI 短课;③ 原理看 3Blue1Brown + fast.ai;④ 伦理看 Anthropic + EU AI Act;⑤ 跟进看 AI Explained + arXiv。Resource path: ① start Google AI Essentials (2 hours); ② prompting see DeepLearning.AI short course; ③ principles see 3Blue1Brown + fast.ai; ④ ethics see Anthropic + EU AI Act; ⑤ stay current see AI Explained + arXiv.
C17

25 个行业:AI 素养的落地差异

25 Industries: How AI Literacy Differs in Practice

不同行业对 AI 素养的要求差异巨大,对照你的行业自查
AI literacy requirements differ hugely by industry; self-check against yours
行业Industry高优先 AI 能力High-priority AI skills最大风险Biggest risk入门重点Entry focus
教育Education用 AI 因材施教、出题、批改AI for personalized teaching, quiz, gradingAI 代学生作业,能力退化AI doing homework for students, ability atrophies提示工程 + 验证习惯Prompting + verification habit
医疗HealthcareAI 查文献、辅助诊断AI for literature, assisted diagnosisAI 误诊、责任不清AI misdiagnosis, unclear liability幻觉检测 + 数据隐私Hallucination detection + data privacy
金融FinanceAI 分析、风控、投研AI for analysis, risk, researchAI 幻觉导致决策错AI hallucination causing wrong decisions批判验证 + 合规Critical verification + compliance
法律LegalAI 查判例、起草文书AI for case law, drafting docsAI 引用不存在的判例AI citing non-existent cases引用验证 + 保密Citation verification + confidentiality
编程ProgrammingAI 写代码、debug、reviewAI for coding, debug, review盲信 AI 代码有 bugBlindly trusting buggy AI code代码验证 + 安全Code verification + security
设计DesignAI 出图、灵感、排版AI for imagery, inspiration, layout过度依赖,创意同质化Over-reliance, homogeneous creativity多模态 + 审美判断Multimodal + aesthetic judgment
营销MarketingAI 写文案、投放、分析AI for copy, ads, analysisAI 文案千篇一律AI copy all the same创意 + 数据Creativity + data
科研ResearchAI 查文献、分析数据AI for literature, data analysisAI 编造引用/数据AI fabricating citations/data溯源 + 批判Tracing + critique
制造ManufacturingAI 质检、预测维护AI for QC, predictive maintenanceAI 误判导致事故AI misjudgment causing accidents数据质量 + 安全Data quality + safety
农业AgricultureAI 遥感、预测产量AI for remote sensing, yield prediction数据不准导致损失Inaccurate data causing losses数据 + 本地化Data + localization
物流LogisticsAI 路径优化、需求预测AI for routing, demand predictionAI 过度优化导致异常AI over-optimization causing anomalies优化 + 容错Optimization + fault tolerance
媒体MediaAI 写稿、配音、生成视频AI for writing, voice, videoAI 假新闻、版权AI fake news, copyright事实核查 + 版权Fact-checking + copyright
政府GovernmentAI 决策支持、公共服务AI for decision support, public serviceAI 偏见影响公平AI bias affecting fairness伦理 + 透明Ethics + transparency
零售RetailAI 推荐、库存、客服AI for recommendation, inventory, CSAI 推荐茧房AI recommendation bubble数据 + 个性化Data + personalization
房地产Real estateAI 估值、带看、营销AI for valuation, tour, marketingAI 估值不准Inaccurate AI valuation数据 + 本地Data + local
旅游TourismAI 行程、客服、个性化AI for itinerary, CS, personalizationAI 行程不靠谱Unreliable AI itineraries多模态 + 本地Multimodal + local
餐饮FoodAI 配方、供应链、营销AI for recipes, supply chain, marketingAI 配方不安全Unsafe AI recipes数据 + 安全Data + safety
游戏GamingAI 生成内容、NPC、平衡AI for content gen, NPC, balanceAI 内容质量差Poor AI content quality多模态 + 设计Multimodal + design
音乐MusicAI 作曲、混音、翻唱AI for composition, mixing, coverAI 版权纠纷AI copyright disputes多模态 + 版权Multimodal + copyright
体育SportsAI 训练分析、伤病预测AI for training analysis, injury predictionAI 误判伤运动员AI misjudgment harming athletes数据 + 谨慎Data + caution
能源EnergyAI 预测、优化、监控AI for prediction, optimization, monitoringAI 误判导致安全事故AI misjudgment causing safety incidents数据 + 安全Data + safety
汽车AutomotiveAI 自动驾驶、制造AI for autonomous driving, manufacturingAI 误判导致事故AI misjudgment causing accidents安全 + 伦理Safety + ethics
建筑ArchitectureAI 设计、BIM、估量AI for design, BIM, estimationAI 设计不安全Unsafe AI designs多模态 + 规范Multimodal + codes
保险InsuranceAI 理赔、核保、反欺诈AI for claims, underwriting, fraudAI 偏见拒绝赔付AI bias denying claims数据 + 公平Data + fairness
电信TelecomAI 网络优化、客服AI for network optimization, CSAI 误判导致服务中断AI misjudgment causing outages数据 + 容错Data + fault tolerance
🏭你的行业在列表里吗?如果不在,用同样的框架分析:你的高优先 AI 能力是什么、最大风险是什么、入门重点是什么。AI 素养不是通用一套,是按行业定制的。Is your industry in the list? If not, use the same framework: what's your high-priority AI skill, biggest risk, entry focus. AI literacy isn't one-size-fits-all; it's customized by industry.
C18

40 个 AI 素养自检项(逐项打分)

40 AI Literacy Self-Check Items (Score Each)

把 40 个能力项逐项打 0-2 分,算出你的总分和等级
Score 40 capability items 0-2, compute your total and level
维度Dimension自检项Self-check item0 分02 分2
理解Understanding我能解释 LLM 工作原理I can explain LLM how it works完全不能Cannot at all能向非技术人讲清Can explain to non-technical
理解Understanding我理解 token/上下文/温度的含义I understand token/context/temperature不知道Don't know能说明影响Can describe their effects
理解Understanding我知道 AI 的『记忆』是假的I know AI's 'memory' is fake以为是真记忆Thinks it's real memory知道是检索/上下文Knows it's retrieval/context
理解Understanding我能说出 AI 不会做什么I can state what AI can't do说不出Can't能列 5 项以上Can list 5+ items
提示Prompting我会给角色+背景+任务+格式I give role+context+task+format只打字Just types提示结构完整Complete prompt structure
提示Prompting我会用 few-shot 给例子I use few-shot examples不会Don't熟练用例子Skillful with examples
提示Prompting我会让 AI 分步思考I make AI think step-by-step不会Don't用 CoT 提升质量Use CoT to improve quality
提示Prompting我有自己的提示模板库I have my own prompt library没有None有 20+ 模板Have 20+ templates
验证Verification我会交叉验证关键信息I cross-verify key info直接相信Believe directly关键都验证Verify all critical
验证Verification我会打开 AI 给的引用I open AI's citations从不Never每次验证Verify every time
验证Verification我能识别 AI 的自信幻觉I can spot confident hallucinations识别不了Can't能识别多数Can spot most
验证Verification我会用反事实测试模型I test models with counterfactuals不会Don't会设计测试Can design tests
验证Verification我区分事实/推断/观点I separate fact/inference/opinion不区分Don't清晰区分Clearly separate
验证Verification我会多模型对比I compare multiple models只用一个Only one关键问题多对比Compare for key questions
隐私Privacy我知道免费版会训练数据I know free versions train on data不知道Don't know清楚区分Clearly distinguish
隐私Privacy我不把机密喂给 AII don't feed AI secrets有时喂Sometimes feed从不喂Never feed
隐私Privacy我会匿名化再输入I anonymize before input不会Don't习惯匿名化Habitual anonymization
隐私Privacy我理解数据流向I understand data flow不理解Don't understand能说出流向Can describe the flow
隐私Privacy我会用本地/商用处理敏感I use local/commercial for sensitive不会Don't按敏感度选Choose by sensitivity
伦理Ethics我了解 AI 的常见偏见I know AI's common biases不了解Don't know能举例Can give examples
伦理Ethics我会用反例测试偏见I test bias with counterexamples不会Don't会多角度测试Multi-angle testing
伦理Ethics我知道 AI 内容的版权归属I know AI content copyright不清楚Unclear清楚规则Clear on rules
伦理Ethics我理解深度伪装的识别I understand deepfake recognition不懂Don't能识别Can recognize
伦理Ethics我遵守 AI 使用的红线I follow AI usage red lines常违反Often violate从不违反Never violate
应用Application我能用 AI 完成完整任务I can use AI for complete tasks只会单点Single points端到端完成End-to-end
应用Application我会用多模态 AII use multimodal AI只用文字Text only会用图/音/视频Use image/audio/video
应用Application我理解 AI Agent 的能力I understand AI Agent capabilities不了解Don't know能设计简单 AgentCan design simple Agent
应用Application我能搭建 RAG 检索I can build RAG retrieval不会Don't理解原理Understand the principle
应用Application我会用 AI 提效日常工作I use AI for daily work efficiency不常用Rarely天天用Daily use
应用Application我能教别人用 AII can teach others to use AI不会Don't能带人入门Can onboard people
批判Critical我对 AI 输出保持怀疑I stay skeptical of AI output完全相信Fully believe默认怀疑Default skeptical
批判Critical我会追问 AI 的依据I ask AI for its basis不追问Don't ask每次追问Ask every time
批判Critical我能指出 AI 论证的漏洞I can spot AI argument gaps不能Can't能找出多数Can find most
批判Critical我不让 AI 替我做重大判断I don't let AI make major decisions有时让Sometimes从不让Never let
批判Critical我持续跟进 AI 新进展I keep up with AI progress不跟进Don't定期跟进Regularly follow
批判Critical我能向别人解释 AI 的局限I can explain AI's limits to others不能Can't能清晰解释Can explain clearly
批判Critical我有自己的 AI 使用原则I have my own AI usage principles没有None有明确原则Clear principles
批判Critical我评估 AI 工具的优劣I evaluate AI tools' pros/cons不评估Don't会对比选型Compare and choose
批判Critical我警惕 AI 的过度承诺I'm wary of AI over-promises轻信Believe easily保持警惕Stay wary
批判Critical我能把 AI 纳入工作流I can integrate AI into workflows不会Don't熟练整合Skilled integration
📊评分规则:每项 0(不会)/1(部分)/2(熟练)。40 项满分 80 分。等级:<30=L2,30-45=L3,45-60=L4,60-70=L5,>70=L6。把你的分数对照 10 节的七级水平表。Scoring: each 0/1/2. 40 items = max 80. Levels: <30=L2, 30-45=L3, 45-60=L4, 60-70=L5, >70=L6. Compare with the 7-level table in section 10.
🎯自评的意义不是打分,而是找到『哪一维最弱』。多数人最弱的是『验证』和『批判』——这两项正是区分 L2 和 L4 的关键。优先补这两维。The point of self-assessment isn't scoring but finding your weakest dimension. Most people's weakest is 'verification' and 'critical' — exactly what separates L2 from L4. Prioritize these two dimensions.
C19

20 条 AI 素养金句(记住这 20 句就够用)

20 AI Literacy Golden Lines (20 Sentences Are Enough)

20 条浓缩的 AI 素养原则,记住就够日常使用
20 condensed AI literacy principles, enough for daily use
金句(中)Golden line (ZH)金句(英)Golden line (EN)解释Explanation
AI 是镜子,不是答案AI is a mirror, not the answerAI 反映你的提问水平;问得差,答得差AI reflects your questioning level; bad question, bad answer
先怀疑,再使用Doubt first, then use默认 AI 可能错,验证后才用Assume AI may be wrong, verify before use
提示就是沟通Prompting is communication把 AI 当聪明但没上下文的同事Treat AI as a smart colleague without context
上下文决定一切Context determines everything给足背景,AI 才能精准Give enough background for AI to be precise
AI 负责初稿,你负责灵魂AI drafts, you give the soul重复交给 AI,判断留给自己Give repetition to AI, keep judgment for yourself
关键信息必验证Verify critical info数字、引用、结论,交叉核对Numbers, citations, conclusions — cross-check
别喂机密Don't feed secrets敏感数据用本地/商用,别上云端Sensitive data local/commercial, not cloud
AI 没有立场,但有权力AI has no stance, but has power它能放大偏见,要主动纠正It can amplify bias, actively correct it
持续验证,持续怀疑Keep verifying, keep doubtingAI 在变,你的警惕也要变AI changes, your vigilance must too
别被效率绑架Don't be hijacked by efficiencyAI 省时间,但别省思考AI saves time, but don't skip thinking
多模型对比Compare multiple models关键问题问 2-3 个模型Key questions to 2-3 models
AI 是杠杆,不是替代AI is leverage, not replacement放大你的能力,不替代你的判断Amplify your ability, not replace your judgment
建立提示库Build a prompt library好提示值得积累和复用Good prompts worth accumulating and reusing
教是最好的学Teaching is the best learning能教别人用 AI,才是真懂Can teach others = truly understand
警惕过度承诺Watch over-promisesAI 说『保证』时,最该怀疑When AI says 'guarantee', doubt most
数据质量决定质量Data quality determines quality垃圾进,垃圾出,AI 也一样Garbage in, garbage out, AI too
保留人工检查点Keep human checkpoints关键操作(删/付/发)必须人确认Critical ops (delete/pay/send) need human confirm
跟随,但别盲从Follow, but don't blindly了解新工具,但别追每个热点Know new tools, but don't chase every trend
AI 素养是长期主义AI literacy is long-termism不是一阵风,是底层能力Not a fad, a foundational ability
用得对,比用得多重要Using right > using much质量优先,别贪多Quality first, don't be greedy
💡这 20 句浓缩了整页的精华。建议截图存手机,每天看 1 句。用 30 天,AI 素养内化成直觉。These 20 sentences condense the whole page's essence. Screenshot to phone, read 1 daily. In 30 days, AI literacy becomes intuition.
C20

60 个 AI 常见问题与解答(FAQ 扩充版)

60 Common AI Questions & Answers (Extended FAQ)

15 条主 FAQ 之外的 60 个高频疑问,覆盖使用/验证/隐私/进阶
60 high-frequency questions beyond the 15 main FAQs, covering usage/verification/privacy/advanced

A. 使用类(20 条)

A. Usage (20)

问题Question解答Answer
AI 能帮我写论文吗?Can AI write my paper?能帮你搭结构、出初稿、润色,但别让它写全文——同质化、查重不过、能力不增长。正确做法:你定观点,AI 出框架和素材,你填充血肉。It can build structure, first drafts, polish; but don't let it write the whole thing — homogenized, fails plagiarism, no growth. Right way: you set the thesis, AI gives framework + material, you fill the flesh.
AI 会取代老师吗?Will AI replace teachers?短期不会。AI 是『助教』,能批改、答疑、个性化,但教育的核心是『点燃』和『榜样』,这是 AI 做不到的。未来是『会用 AI 的老师』取代『不会用的』。Short-term no. AI is a 'teaching assistant' — grades, answers, personalizes; but education's core is 'igniting' and 'role model', which AI can't do. Future: teachers who use AI replace those who don't.
AI 能帮我做研究吗?Can AI help with research?能帮你查文献、总结、分析数据、找盲点,但别让它『保证』结论。关键发现一定要自己验证。AI 是加速器,不是替代者。It can search literature, summarize, analyze data, find blind spots; but don't let it 'guarantee' conclusions. Verify key findings yourself. AI accelerates, doesn't replace.
AI 能帮我学编程吗?Can AI help me learn programming?非常能。让 AI 当结对编程伙伴:你写,AI 解释;你卡住,AI 给提示(不是答案);你写完,AI review。这是目前最有效的学编程方式之一。Very. Make AI a pair-programming partner: you write, AI explains; you're stuck, AI gives hints (not answers); you finish, AI reviews. One of the most effective ways to learn programming now.
AI 能帮我翻译吗?Can AI translate?能,日常翻译很好,专业/文学翻译要人工校对。翻译后用『母语者』视角检查:地道吗?术语对吗?语气对吗?Yes, daily translation is good; professional/literary needs human proofread. After translating, check from a 'native speaker' view: idiomatic? terms right? tone right?
AI 能帮我做 PPT 吗?Can AI make PPT?能出大纲、每页要点、讲稿,但排版和设计要自己做。AI 出『内容』,你做『形式』。It can outline, per-slide points, scripts; but layout and design you do. AI gives 'content', you do 'form'.
AI 能帮我写代码吗?Can AI write code?能,日常代码很靠谱,复杂逻辑要验证。关键:别盲信,跑测试、读逻辑、查边界。AI 写的代码也要 review。Yes, daily code is reliable; complex logic verify. Key: don't blindly trust, run tests, read logic, check edges. AI code also needs review.
AI 能帮我做数据分析吗?Can AI do data analysis?能帮你写分析代码、解释结果、找趋势,但数据本身的质量你负责。垃圾数据进,垃圾结论出。It can write analysis code, explain results, find trends; but you're responsible for data quality. Garbage in, garbage out.
AI 能帮我做设计吗?Can AI do design?能出图、灵感、配色,但『审美判断』和『品牌一致性』是你。AI 是素材库,你是设计师。It can generate images, inspiration, color schemes; but 'aesthetic judgment' and 'brand consistency' are yours. AI is a material library, you're the designer.
AI 能帮我写商业计划吗?Can AI write a business plan?能出框架、市场规模估算、竞品分析,但核心假设(你的独特价值、数据验证)你自己。AI 不能替你做商业判断。It can give framework, market size estimates, competitor analysis; but core assumptions (your unique value, data validation) are yours. AI can't make business judgments for you.
AI 能帮我做投资分析吗?Can AI do investment analysis?能帮你整理数据、算指标、找模式,但投资决策你自己。AI 不承诺收益、不预测未来。记住:历史数据不代表未来。It can organize data, compute metrics, find patterns; but investment decisions are yours. AI doesn't promise returns, doesn't predict the future. Remember: past data ≠ future.
AI 能帮我写简历吗?Can AI write a resume?能帮你改写、优化关键词、突出亮点,但真实经历你自己。别编造——背调会露馅。It can help rewrite, optimize keywords, highlight; but real experience is yours. Don't fabricate — background checks expose it.
AI 能帮我做面试准备吗?Can AI prep me for interviews?非常能。让 AI 当面试官,模拟提问、追问、压力测试,你回答,AI 给反馈。这是最有效的面试准备方式之一。Very. Make AI an interviewer, mock questions, follow-ups, stress tests, you answer, AI gives feedback. One of the most effective interview prep methods.
AI 能帮我做决策吗?Can AI make decisions?能给正反对论、风险、盲点,帮你『看得更全』,但决策是你。AI 没有你的处境、价值观、责任。It can give pro/con, risks, blind spots, helping you 'see more completely'; but the decision is yours. AI has neither your circumstances, values, nor responsibility.
AI 能帮我管理时间吗?Can AI manage my time?能帮你排日程、设优先级、估算工作量,但『执行』和『专注』是你。工具再强,执行还是靠你。It can schedule, set priorities, estimate workload; but 'execution' and 'focus' are yours. No matter how strong the tool, execution depends on you.
AI 能帮我学语言吗?Can AI teach me a language?能当陪练:对话、纠错、解释、出题。但『坚持』和『输入量』是你。AI 是陪练,不是替练。It can be a sparring partner: conversation, correction, explanation, quiz. But 'persistence' and 'input volume' are yours. AI is a sparring partner, not a substitute trainer.
AI 能帮我写小说吗?Can AI write a novel?能帮你脑暴、出结构、写片段,但『灵魂』和『连贯性』是你。纯 AI 小说同质化、没深度。人机协作最好。It can brainstorm, structure, write fragments; but 'soul' and 'coherence' are yours. Pure AI novels are homogenized, shallow. Human-AI collaboration is best.
AI 能帮我做视频吗?Can AI make videos?能出脚本、分镜、配音、生成画面,但剪辑节奏和叙事是你。AI 是素材和草稿,你是导演。It can script, storyboard, voice, generate visuals; but editing rhythm and narrative are yours. AI is material and draft, you're the director.
AI 能帮我做音乐吗?Can AI make music?能作曲、编曲、生成,但『情感』和『品味』是你。AI 是乐器,你是音乐人。It can compose, arrange, generate; but 'emotion' and 'taste' are yours. AI is the instrument, you're the musician.
AI 能帮我做数学题吗?Can AI solve math problems?能解题、讲步骤,但别让它『直接给答案』。让 AI 当老师,引导你自己想出来,这才是学习。It can solve and explain steps; but don't let it 'give the answer directly'. Make AI a teacher, guiding you to figure it out yourself — that's learning.

B. 验证类(15 条)

B. Verification (15)

问题Question解答Answer
怎么验证 AI 说的数字?How to verify AI's numbers?交叉验证:查原始来源、对比多个独立来源、看量级是否合理。关键数字(金额、日期、比例)一定要查。Cross-verify: check original source, compare multiple independent sources, check if magnitude is reasonable. Key numbers (amount, date, ratio) must be checked.
怎么验证 AI 的引用?How to verify AI's citations?逐个打开。AI 可能编造『看起来真实』的引用(作者、标题、期刊都对,但论文不存在)。用 Google Scholar / arXiv 搜标题。Open each one. AI may fabricate 'plausibly real' citations (author, title, journal all right, but the paper doesn't exist). Search the title on Google Scholar / arXiv.
怎么验证 AI 的事实?How to verify AI's facts?用『反向搜索』:搜『[事实] 是不是真的』、搜『[事实] 辟谣』。如果找不到支持来源,警惕。Use 'reverse search': search '[fact] is it true', search '[fact] debunk'. If you can't find supporting sources, be wary.
怎么验证 AI 的代码?How to verify AI's code?跑测试、读逻辑、查边界条件、用 linter。别盲信『能跑』——能跑不代表正确。Run tests, read logic, check edge conditions, use linter. Don't blindly trust 'it runs' — running doesn't mean correct.
怎么验证 AI 的建议?How to verify AI's advice?对照你的具体情境。AI 的建议是『一般性』的,你的情况是『特定』的。问自己:这条建议在我这里成立吗?Compare with your specific situation. AI's advice is 'general', your situation is 'specific'. Ask: does this advice hold in my case?
怎么验证 AI 的翻译?How to verify AI's translation?用母语者视角读:地道吗?术语对吗?语气对吗?专业领域找同行核对。Read from a native speaker's view: idiomatic? terms right? tone right? For professional fields, get a peer to check.
怎么验证 AI 的总结?How to verify AI's summary?对照原文。总结是否遗漏关键信息?是否曲解?用『5 个 W』检查:谁、什么、何时、何地、为什么。Compare with the original. Does the summary miss key info? Misinterpret? Check with '5 Ws': who, what, when, where, why.
怎么验证 AI 的分析?How to verify AI's analysis?检查逻辑链:前提成立吗?推理有效吗?结论必要吗?有没有其他解释?Check the logic chain: do premises hold? Is reasoning valid? Is the conclusion necessary? Are there other explanations?
怎么验证 AI 的预测?How to verify AI's predictions?预测最难验证。看依据:基于什么数据?假设是什么?置信度多少?如果依据不透明,别信。Predictions are hardest to verify. Check the basis: based on what data? What assumptions? What confidence? If the basis is opaque, don't trust.
怎么验证 AI 的引用是否存在?How to verify AI's citation exists?搜完整标题(不是关键词)。如果搜不到,大概率是编的。学术引用必须 100% 验证。Search the full title (not keywords). If not found, likely fabricated. Academic citations must be 100% verified.
怎么验证 AI 的数据来源?How to verify AI's data source?问 AI『数据从哪来』。如果它说不出具体来源,或来源模糊,警惕。可信数据应有明确出处。Ask AI 'where does the data come from'. If it can't give a specific source, or the source is vague, be wary. Credible data should have a clear source.
怎么验证 AI 的逻辑?How to verify AI's logic?找漏洞:有没有以偏概全?有没有因果倒置?有没有忽略反例?逻辑错,结论全错。Find gaps: any overgeneralization? Any reversed causation? Any ignored counterexamples? If logic is wrong, the whole conclusion is wrong.
怎么验证 AI 的时效性?How to verify AI's timeliness?问『这是哪年的信息』。AI 的训练有截止日期,时效性强的信息(价格、政策、技术)要查最新。Ask 'what year is this info'. AI's training has a cutoff; time-sensitive info (price, policy, tech) check for the latest.
怎么验证 AI 的完整性?How to verify AI's completeness?问『还有没有别的角度』。AI 可能只给一边。主动追问:反方观点是什么?Ask 'are there other angles'. AI may only give one side. Proactively ask: what's the counterargument?
怎么验证 AI 的可靠性?How to verify AI's reliability?用『反事实测试』:问 AI 一个不存在的事(虚构的论文/事件),看它会不会编。会编的模型,要更警惕。Use 'counterfactual test': ask AI something that doesn't exist (a fake paper/event), see if it fabricates. If it fabricates, be more wary.

C. 隐私与安全类(15 条)

C. Privacy & Security (15)

问题Question解答Answer
免费版 AI 会保存我的数据吗?Do free AI versions save my data?大多数免费版会用于训练。敏感信息别用免费版。商用版或本地部署更安全。Most free versions use data for training. Don't use free versions for sensitive info. Commercial or local deployment is safer.
我可以把公司机密喂给 AI 吗?Can I feed AI company secrets?不能。除非用的是企业版/本地部署,且有数据协议。默认假设:喂给公有 AI 的数据可能被训练。No. Unless using enterprise/local deployment with a data agreement. Default assumption: data fed to public AI may be trained on.
AI 会记住我吗?Will AI remember me?单次对话内会(上下文),跨对话多数不记(除非产品有记忆功能)。别依赖它『记得』你。Within a single conversation, yes (context); across conversations, most don't (unless the product has memory). Don't rely on it 'remembering' you.
怎么保护我的隐私?How to protect my privacy?① 匿名化(化名、去标识);② 用商用/本地;③ 别喂身份证件、密码、财务;④ 了解数据协议。① Anonymize (alias, de-identify); ② use commercial/local; ③ don't feed ID, passwords, finances; ④ understand the data agreement.
AI 生成的内容我能商用吗?Can I commercially use AI-generated content?看平台和地区规则。多数地区:AI 生成内容版权模糊,部分不能注册版权。商用前查清楚。Depends on platform and regional rules. In most regions: AI-generated content copyright is fuzzy, some can't be registered. Check before commercial use.
AI 会泄露我的信息吗?Will AI leak my info?理论上可能(训练、存储、攻击)。降低风险:匿名化、用可信平台、敏感数据本地处理。Theoretically possible (training, storage, attacks). Reduce risk: anonymize, use trusted platforms, process sensitive data locally.
我能用 AI 处理客户数据吗?Can I process customer data with AI?看数据协议和法规(如 GDPR)。多数情况:需匿名化 + 用合规平台 + 客户知情。Depends on data agreement and regulations (e.g. GDPR). Most cases: need anonymization + compliant platform + customer consent.
AI 会判断我的意图吗?Will AI judge my intent?不会。AI 没有价值观判断,它按训练数据回应。你的责任是确保使用合法合规。No. AI has no value judgment; it responds per training data. Your responsibility is to ensure legal and compliant use.
怎么用 AI 不违反法律?How to use AI without breaking the law?① 别用 AI 造假(证件、签名);② 别侵犯版权;③ 别做歧视性内容;④ 遵守行业法规;⑤ 标注 AI 生成(如要求)。① Don't use AI to forge (ID, signatures); ② don't infringe copyright; ③ no discriminatory content; ④ follow industry regulations; ⑤ label AI-generated (if required).
AI 会帮我做违法的事吗?Will AI help me do illegal things?正规 AI 会拒绝明显违法的请求。但『灰色地带』要自己负责。别用 AI 绕开法律。Legitimate AI refuses clearly illegal requests. But 'gray areas' are your responsibility. Don't use AI to bypass the law.
我能用 AI 生成虚假内容吗?Can I use AI to generate false content?技术上能,但别用于欺骗。深度伪造(fake 名人、假新闻)是法律风险 + 道德问题。Technically yes, but don't use for deception. Deepfakes (fake celebrities, fake news) are legal risk + moral issue.
AI 会歧视我吗?Will AI discriminate against me?AI 可能带训练数据的偏见(性别、种族、地域)。遇到偏见,指出并要求换角度。AI may carry training-data bias (gender, race, region). When you encounter bias, point it out and ask for another angle.
我能用 AI 处理医疗数据吗?Can I process medical data with AI?敏感数据。需匿名化 + 合规平台 + 严格协议。AI 不能做诊断,只能辅助。Sensitive data. Need anonymization + compliant platform + strict agreement. AI can't diagnose, only assist.
AI 会记住我说的密码吗?Will AI remember passwords I say?对话内可能在上下文里,但别把密码告诉任何 AI。密码只在你和系统之间。Within a conversation it may be in context, but don't tell any AI your passwords. Passwords stay only between you and the system.
怎么用 AI 最安全?How to use AI most safely?① 最小必要(只喂需要的);② 匿名化;③ 商用/本地;④ 了解平台协议;⑤ 敏感操作人工确认。① Minimum necessary (only feed what's needed); ② anonymize; ③ commercial/local; ④ understand platform terms; ⑤ human confirm for sensitive ops.

D. 进阶类(10 条)

D. Advanced (10)

问题Question解答Answer
什么是 RAG?What is RAG?Retrieval-Augmented Generation:先从知识库检索相关资料,再让 LLM 基于资料回答。解决 LLM 知识过时和幻觉问题。Retrieval-Augmented Generation: first retrieve relevant material from a knowledge base, then let the LLM answer based on it. Solves LLM outdated knowledge and hallucination.
什么是 AI Agent?What is an AI Agent?能自主规划、调用工具、执行多步任务的 AI 系统。如:自动调研+写报告、自动操作电脑。An AI system that can autonomously plan, call tools, execute multi-step tasks. E.g., auto research + write report, auto operate a computer.
什么是微调?What is fine-tuning?用特定数据继续训练基础模型,让它更懂特定领域。如:用法律数据微调,让它更懂法律。Continue training a base model with specific data so it understands a specific domain better. E.g., fine-tune with legal data so it understands law better.
什么是多模态?What is multimodal?AI 能处理多种模态(文字、图像、音频、视频)的能力。如:看图回答、听音转文字。AI's ability to process multiple modalities (text, image, audio, video). E.g., answer from an image, transcribe audio.
什么是思维链(CoT)?What is Chain-of-Thought (CoT)?让 AI 一步步推理再给答案,提升复杂任务准确性。提示里写『一步步思考』。Make AI reason step-by-step before giving the answer, improving accuracy on complex tasks. Write 'think step by step' in the prompt.
什么是上下文窗口?What is the context window?AI 一次能『看到』的 token 数量上限。超过会被截断。长文档要分段或压缩。The token limit AI can 'see' at once. Exceeding gets truncated. Long documents must be split or compressed.
什么是温度(temperature)?What is temperature?控制 AI 输出的随机性。低温=确定/保守,高温=创意/多变。要准确用低温度,要创意用高温度。Controls randomness of AI output. Low = deterministic/conservative, high = creative/varied. Use low for accuracy, high for creativity.
什么是 embedding?What is embedding?把文字转成数字向量,让 AI 能比较语义相似度。RAG 的核心:检索就是找『向量最接近』的资料。Convert text into numeric vectors so AI can compare semantic similarity. The core of RAG: retrieval is finding the 'closest vector' material.
什么是大模型 vs 小模型?What is large vs small model?大模型(GPT/Claude)能力强但贵;小模型(7B-13B)便宜但弱。按任务选:要能力用大,要成本用小。Large models (GPT/Claude) are powerful but expensive; small (7B-13B) are cheap but weaker. Choose by task: capability → large, cost → small.
什么是 AI 对齐?What is AI alignment?让 AI 的行为符合人类意图和价值观的技术方向。避免 AI 做有害的事。是当前 AI 安全的核心研究。The technical direction of making AI's behavior align with human intent and values. Avoiding AI doing harm. A core current AI safety research area.
📖这 60 个问题是使用 AI 时最常遇到的。建议存下来,遇到问题先查这里,再问 AI。These 60 are the most common when using AI. Suggest saving them; when a problem arises, check here first, then ask AI.
C21

40 个主流 AI 工具对比(2026)

40 Mainstream AI Tools Compared (2026)

按用途分类的 40 个真实工具,帮助选型
40 real tools by use case, helping selection
类别Category工具Tool特点Feature适合Best for费用Cost
通用 LLMGeneral LLMChatGPTChatGPT综合能力最强,生态大Strongest overall, big ecosystem日常通用Daily general免费/订阅Free/sub
通用 LLMGeneral LLMClaudeClaude长上下文、推理强、安全Long context, strong reasoning, safe长文档/复杂推理Long docs/complex reasoning免费/订阅Free/sub
通用 LLMGeneral LLMGeminiGemini多模态强、Google 生态Strong multimodal, Google ecosystem多模态/Google 用户Multimodal/Google users免费/订阅Free/sub
通用 LLMGeneral LLMDeepSeekDeepSeek开源、性价比高Open-source, good value开发者/性价比Developers/value免费Free
通用 LLMGeneral LLMLlamaLlama开源、可本地部署Open-source, local deploy本地/隐私敏感Local/privacy-sensitive免费Free
通用 LLMGeneral LLMQwenQwen开源、中文强Open-source, strong Chinese中文场景Chinese scenarios免费Free
通用 LLMGeneral LLMMistralMistral开源、高效Open-source, efficient欧洲/高效Europe/efficient免费Free
写作WritingJasperJasper营销文案强Strong marketing copy营销/品牌Marketing/brand订阅Sub
写作WritingGrammarlyGrammarly英文语法/风格English grammar/style英文写作English writing免费/订阅Free/sub
写作WritingNotion AINotion AI笔记+AI 一体Notes + AI integrated知识管理Knowledge mgmt订阅Sub
写作WritingQuillBotQuillBot改写/同义替换Rewriting/synonyms改写/降重Rewrite/paraphrase免费/订阅Free/sub
翻译TranslationDeepLDeepL专业翻译质量高High professional translation quality专业翻译Professional translation免费/订阅Free/sub
翻译TranslationGoogle TranslateGoogle Translate多语言、快Many languages, fast日常翻译Daily translation免费Free
翻译TranslationMicrosoft TranslatorMicrosoft Translator实时翻译、TeamsReal-time, Teams会议/TeamsMeetings/Teams免费Free
编程CodingGitHub CopilotGitHub Copilot代码补全、集成 VS CodeCode completion, VS Code integrated编程Programming订阅Sub
编程CodingCursorCursorAI 编辑器、整项目理解AI editor, whole-project understanding重度编程Heavy programming订阅Sub
编程CodingCodeiumCodeium免费代码补全、开源Free code completion, open-source免费编程Free programming免费Free
编程CodingReplitReplit在线 IDE + AIOnline IDE + AI快速原型Quick prototypes免费/订阅Free/sub
图像ImageMidjourneyMidjourney艺术感最强Strongest artistry艺术/设计Art/design订阅Sub
图像ImageDALL-EDALL-EChatGPT 内置、理解好Built into ChatGPT, good understanding通用图像General image免费/订阅Free/sub
图像ImageStable DiffusionStable Diffusion开源、可本地Open-source, local本地/自定义Local/custom免费Free
图像ImageAdobe FireflyAdobe Firefly商用安全、Adobe 生态Commercially safe, Adobe ecosystem商用设计Commercial design订阅Sub
音频AudioElevenLabsElevenLabs语音合成最真Most realistic voice synthesis配音/有声Voiceover/audio免费/订阅Free/sub
音频AudioWhisperWhisper语音转文字、开源Speech-to-text, open-source转写/字幕Transcription/subtitles免费Free
音频AudioSunoSunoAI 作曲、生成歌曲AI composition, generate songs音乐制作Music production免费/订阅Free/sub
音频AudioMurfMurf配音、多语言Voiceover, multilingual企业配音Enterprise voiceover订阅Sub
视频VideoRunwayRunway文生视频、编辑Text-to-video, editing视频创作Video creation订阅Sub
视频VideoPikaPika短视频、特效Short video, effects短视频/特效Short video/effects订阅Sub
视频VideoSynthesiaSynthesia数字人视频、企业AI avatar video, enterprise企业培训Enterprise training订阅Sub
视频VideoCapCutCapCut剪辑 + AI 字幕Editing + AI subtitles短视频剪辑Short video editing免费Free
研究ResearchElicitElicit学术文献分析Academic literature analysis科研Research免费/订阅Free/sub
研究ResearchSciteScite引用分析、支持/反驳Citation analysis, supports/refutes学术Academic免费/订阅Free/sub
研究ResearchPerplexityPerplexity搜索 + 引用Search + citations研究/问答Research/Q&A免费/订阅Free/sub
搜索SearchPerplexityPerplexityAI 搜索、带来源AI search, with sources信息检索Info retrieval免费/订阅Free/sub
搜索SearchYou.comYou.comAI 搜索、隐私AI search, privacy隐私搜索Privacy search免费Free
搜索SearchGoogle AI OverviewGoogle AI Overview搜索 + AI 摘要Search + AI summary日常搜索Daily search免费Free
生产力ProductivityNotion AINotion AI笔记 + AINotes + AI知识管理Knowledge mgmt订阅Sub
生产力ProductivityMicrosoft 365 CopilotMicrosoft 365 CopilotOffice 集成Office integratedOffice 用户Office users订阅Sub
生产力ProductivityZapier AIZapier AI自动化 + AIAutomation + AI工作流自动化Workflow automation订阅Sub
🧭选型原则:① 先明确用途(写作/编程/图像/研究);② 再看费用(免费够不够);③ 最后看隐私(敏感用本地/商用)。别追『最强』,追『最合适』。Selection principle: ① clarify use (writing/coding/image/research); ② then cost (is free enough); ③ finally privacy (sensitive → local/commercial). Don't chase 'strongest', chase 'most suitable'.
⚠️工具迭代快(半年一代)。本表是 2026 年快照,用前确认工具还在、没换名、没改价。Tools iterate fast (a new generation every 6 months). This table is a 2026 snapshot; confirm the tool still exists, hasn't renamed, hasn't changed price before using.
C22

30 条 AI 素养误区与正解(对照表)

30 AI Literacy Myths vs Truths (Comparison)

30 个关于 AI 的常见误解,逐条对照正解,破除认知障碍
30 common AI misconceptions, each matched with the truth, clearing cognitive barriers
误区Myth正解Truth
AI 是百科全书,什么都知道AI is an encyclopedia, knows everythingAI 是概率模型,不是知识库。它会编造、会过时、会遗漏。AI is a probabilistic model, not a knowledge base. It fabricates, goes stale, and omits.
AI 回答越长越对Longer AI answers are more correct长度≠正确。简洁准确的回答可能比冗长的更对。Length ≠ correctness. A concise, accurate answer may be more right than a verbose one.
AI 不会错,错了是用户笨AI doesn't err; errors are the user's faultAI 会错,且会自信地错。错的责任在使用前没验证,不在模型。AI errs, and confidently so. The responsibility for the error is in not verifying before use, not in the model.
提示越短越好Shorter prompts are better提示要『完整』不是『短』。该给的上下文要给够。Prompts should be 'complete', not 'short'. Give the context that's needed.
AI 能帮我思考AI can think for meAI 能帮你『整理』和『加速』,但『思考』和『判断』是你。AI helps you 'organize' and 'accelerate', but 'thinking' and 'judging' are yours.
用 AI = 作弊Using AI = cheating取决于用途。辅助学习不是作弊,代写作业是。关键看『学习是否发生』。Depends on use. Assisting learning isn't cheating; doing homework for you is. The key is 'whether learning happened'.
AI 懂我的意思AI understands my meaningAI 按字面 + 统计推测,经常误解。说清楚比『它该懂』更可靠。AI reads literally + statistically, often misreading. Being explicit is more reliable than 'it should get it'.
免费 AI 和付费 AI 一样好Free and paid AI are equally good通常付费更强(更长上下文、更准、新功能优先)。按需求选。Usually paid is stronger (longer context, more accurate, new features first). Choose by need.
AI 能记住所有对话AI remembers all conversations多数 AI 跨对话不记(除非有记忆功能)。别依赖它『记得』。Most AI don't remember across conversations (unless there's a memory feature). Don't rely on it 'remembering'.
AI 生成的一定有版权AI-generated content is always copyrighted多数地区 AI 生成内容版权模糊,部分不能注册。商用前查规则。In most regions, AI-generated content copyright is fuzzy; some can't be registered. Check rules before commercial use.
AI 不会歧视AI doesn't discriminateAI 带训练数据偏见(性别/种族/地域)。要主动测试和纠正。AI carries training-data bias (gender/race/region). Actively test and correct.
AI 能预测未来AI can predict the futureAI 基于历史数据推测,不能预测真正的新颖事件。历史≠未来。AI infers from historical data; it can't predict truly novel events. Past ≠ future.
AI 越贵越好The more expensive, the better AI按任务选。简单任务便宜模型够;复杂任务才用强的。Choose by task. Simple tasks are fine with cheap models; complex ones need the strong ones.
AI 写的代码可以直接用AI-written code can be used as-is要跑测试、读逻辑、查边界。能跑≠正确。Run tests, read logic, check edges. Running ≠ correct.
AI 翻译比人好AI translation is better than humans日常 AI 很好,专业/文学要人校对。术语、语气、文化要人把关。Daily, AI is good; professional/literary needs human proofread. Terms, tone, culture need human gatekeeping.
AI 能帮我做所有事AI can do everything for meAI 是杠杆,不是全能。判断、责任、创造是你。AI is leverage, not omnipotent. Judgment, responsibility, creation are yours.
用 AI 会让我变笨Using AI makes me dumber取决于怎么用。辅助思考不变笨,替代思考会。Depends on how. Assisting thinking doesn't make you dumber; replacing thinking does.
AI 不会幻觉,是用户看错AI doesn't hallucinate; the user misreadAI 会幻觉(自信地编)。关键输出必须验证。AI hallucinates (confidently fabricates). Critical output must be verified.
大模型越大越好The bigger the model, the better大模型强但贵、慢。小模型在特定任务可能更好。Large models are powerful but expensive and slow. Small models may be better for specific tasks.
AI 能理解情感AI understands emotionsAI 识别情绪信号,但不『体验』情感。别把它当情感伙伴的唯一。AI detects emotional signals but doesn't 'experience' emotion. Don't make it your only emotional companion.
提示词有标准答案There's a standard answer for prompts没有万能提示。好提示是『具体 + 完整 + 迭代』。No universal prompt. Good prompts are 'specific + complete + iterative'.
AI 越新越好The newer AI, the better新模型有新能力,也有新坑。稳定 + 适合 > 追新。Newer models bring new capabilities and new pitfalls. Stable + suitable > chasing new.
AI 能取代我的职业AI will replace my professionAI 取代『任务』,不取代『职业』。会用 AI 的人更安全。AI replaces 'tasks', not 'professions'. People who use AI are safer.
AI 不需要监管AI doesn't need regulationAI 有风险(偏见/隐私/滥用),需要合理监管。AI has risks (bias/privacy/misuse) and needs reasonable regulation.
本地 AI 一定比云端安全Local AI is always safer than cloud本地有隐私优势,但模型可能弱、成本高。按敏感度选。Local has a privacy advantage, but the model may be weaker and cost more. Choose by sensitivity.
AI 回答就是事实AI answers are factsAI 回答是『最可能的文字』,不是『事实』。要验证。AI answers are 'most likely text', not 'facts'. Verify.
一次对话就能学会You can learn in one conversationAI 是工具,学习靠『用 + 练 + 复盘』。一次对话不够。AI is a tool; learning requires 'using + practicing + reviewing'. One conversation isn't enough.
AI 不能做创意AI can't be creativeAI 能做创意(组合、变体、灵感),但『原创灵魂』是你。AI can create (combination, variants, inspiration), but the 'original soul' is yours.
AI 永远在线不会累AI is always online and never tiresAI 没『疲劳』,但有上下文限制和速率限制。AI has no 'fatigue', but has context and rate limits.
AI 能给我的人生答案AI can give me life answersAI 给分析、给选项、给盲点,但『你的人生』只能你答。AI gives analysis, options, blind spots; but 'your life' only you can answer.
🚨这 30 个误区里,最危险的是:①『AI 不会错』(让你不验证);②『AI 能取代我』(让你不思考);③『AI 懂我』(让你不沟通清楚)。这三个认知一旦建立,AI 素养就废了。Of these 30 myths, the most dangerous are: ① 'AI doesn't err' (makes you not verify); ② 'AI can replace me' (makes you not think); ③ 'AI understands me' (makes you not communicate clearly). Once these cognitions take root, AI literacy is dead.
✅破除误解的方法:每遇到一个『AI 好像很全能』的念头,问自己三个问题——① 它真的对吗?② 我验证了吗?③ 如果错了,后果谁担?问完,误区自然破。Method to break myths: whenever you have the thought 'AI seems all-powerful', ask yourself three questions — ① Is it really right? ② Did I verify? ③ If wrong, who bears the consequence? After asking, the myth breaks naturally.
C23

10 个真实案例:AI 素养改变职业

10 Real Cases: AI Literacy Changing Careers

10 个普通人因 AI 素养提升而职业跃迁的真实案例,可复制
10 real cases of ordinary people whose careers advanced through higher AI literacy; replicable
案例Case背景BackgroundAI 素养动作AI literacy action结果Result
案例 1Case 135 岁行政,被优化35 admin, laid off自学 AI,用 AI 做简历+作品集,转型数据助理Self-taught AI, used AI for resume + portfolio, transitioned to data assistant3 个月入职数据助理,薪资持平Hired as data assistant in 3 months, salary maintained
案例 2Case 228 岁教师,想提升效率28 teacher, wants efficiency用 AI 出题、批改、个性化辅导,节省 2/3 批改时间Used AI for quiz, grading, personalized tutoring, saved 2/3 grading time能带更多学生,获校『效率先锋』Could take more students, won school's 'efficiency pioneer'
案例 3Case 340 岁设计师,创意瓶颈40 designer, creative bottleneck用 AI 出灵感、变体、配色,自己把关审美Used AI for inspiration, variants, color schemes, self-gated aesthetics产出翻倍,接更多项目Output doubled, took more projects
案例 4Case 425 岁应届生,面试焦虑25 fresh grad, interview anxiety用 AI 模拟面试 50 次,AI 追问+反馈Used AI to mock interviews 50 times, AI follow-up + feedback拿到 3 个 offerGot 3 offers
案例 5Case 550 岁创业者,不懂技术50 entrepreneur, not tech-savvy用 AI 写商业计划、分析市场、做产品原型Used AI to write business plan, analyze market, build product prototype产品 2 周上线,融资成功Product launched in 2 weeks, funding successful
案例 6Case 632 岁律师,文书繁重32 lawyer, heavy paperwork用 AI 起草文书、查判例、整理证据Used AI to draft documents, search case law, organize evidence文书时间减 60%,专注核心Document time cut 60%, focused on core
案例 7Case 722 岁学生,英语差22 student, poor English用 AI 当语言陪练,每天 30 分钟对话Used AI as language sparring partner, 30 min daily conversation一年后英语流利,拿到海外实习English fluent after a year, got overseas internship
案例 8Case 845 岁销售,不会写方案45 sales, can't write proposals用 AI 写方案、分析客户、准备谈判Used AI to write proposals, analyze clients, prepare negotiations成单率提升 40%Close rate up 40%
案例 9Case 930 岁自媒体,内容枯竭30 content creator, content dried up用 AI 脑暴、出结构、写初稿,自己加灵魂Used AI to brainstorm, structure, draft, self-added soul更新频率翻倍,涨粉 3 倍Update frequency doubled, followers tripled
案例 10Case 1038 岁 HR,招聘低效38 HR, inefficient hiring用 AI 筛简历、写 JD、面试辅助Used AI to screen resumes, write JDs, assist interviews招聘周期从 6 周缩到 2 周Hiring cycle from 6 weeks to 2 weeks
💡这些案例的共同点:不是『AI 替代了他们』,而是『他们用 AI 放大了自己』。AI 素养 = 把 AI 变成你的杠杆,而不是你的替身。The common thread in these cases: it wasn't 'AI replaced them', but 'they used AI to amplify themselves'. AI literacy = making AI your leverage, not your substitute.
🎯每个案例都可复制。找你的『痛点』(效率/创意/语言/文书),选一个 AI 工具,每天 30 分钟,30 天看变化。Each case is replicable. Find your 'pain point' (efficiency/creativity/language/documents), pick one AI tool, 30 minutes daily, 30 days to see the change.
C24

一页总结:AI 素养速记卡

One-Page Summary: AI Literacy Quick-Reference Card

把整页浓缩成一张速记卡,打印贴墙,每天扫一眼
Condense the whole page into a quick-reference card, print and stick on the wall, glance daily
维度Dimension核心Core口诀Mnemonic
认知CognitionAI 是概率模型,不是知识库AI is a probabilistic model, not a knowledge base先怀疑,再使用Doubt first, then use
提示Prompting完整 > 短;迭代 > 一次Complete > short; iterate > one-shot上下文决定一切Context determines everything
验证Verification关键必查:数字/引用/结论Key things must check: numbers/citations/conclusions交叉验证 + 反事实Cross-verify + counterfactual
隐私Privacy敏感不喂公有;匿名化Don't feed public with sensitive; anonymize最小必要 + 本地商用Minimum necessary + local/commercial
伦理Ethics主动测偏见;标注 AI 生成Actively test bias; label AI-generatedAI 有权力,要纠正AI has power, must correct
应用ApplicationAI 初稿,你灵魂AI drafts, you soul重复给 AI,判断留自己Repetition to AI, judgment to yourself
进阶AdvancedRAG/Agent/多模态/微调RAG/Agent/multimodal/fine-tune了解原理,按任务选Know principles, choose by task
长期Long-termAI 素养是底层能力,不是一阵风AI literacy is a foundational ability, not a fad用得对 > 用得多Using right > using much
🎓AI 素养不是终点,是起点。今天你懂了原理、会了验证、建了红线,明天就能把 AI 真正变成你的杠杆。2026 年,AI 素养 = 新的读写能力。AI literacy isn't the endpoint, it's the starting point. Today you understand the principles, can verify, and set red lines; tomorrow you can truly make AI your leverage. In 2026, AI literacy = the new literacy.
📌把这张卡存手机壁纸。每天看一次『口诀』列。30 天后,AI 素养内化成直觉,你就不再用『AI 全能』或『AI 不行』的极端思维,而是『AI 能干 X,我不能干 Y,我验证 Z』的精确思维。Save this card as your phone wallpaper. Read the 'mnemonic' column daily. After 30 days, AI literacy becomes intuition, and you no longer think in the extreme 'AI is all-powerful' or 'AI is useless', but in the precise 'AI can do X, I can't do Y, I verify Z'.
C25

50 条 AI 素养行动清单(按周执行)

50 AI Literacy Action Items (By Week)

50 个可立即执行的具体动作,按周分组,直接照做
50 immediately actionable concrete actions, grouped by week, do them directly
周Week动作Action检验标准Success metric
W1W1每天用 AI 做 1 件日常事Do 1 daily thing with AI daily当天用满,不是装Used it that day, not just installed
W1W1给每个提示加角色+背景+格式Add role+context+format to every prompt提示有 3 要素Prompt has 3 elements
W1W1存下 3 个你常用的提示模板Save 3 commonly used prompt templates有 3 个模板Have 3 templates
W1W1识别 1 个 AI 幻觉(反事实测试)Spot 1 AI hallucination (counterfactual test)测试了不存在的引用Tested a non-existent citation
W1W1了解你在用的 AI 的数据政策Learn the data policy of the AI you use能说出是否训练数据Can state if it trains on data
W1W1把机密/敏感信息从提示里去掉Remove secrets/sensitive info from prompts提示无敏感信息Prompt has no sensitive info
W1W1问 AI 一个它该不知道的问题Ask AI something it shouldn't know看到它怎么编Saw how it fabricates
W1W1记录 1 个 AI 帮你的时刻Record 1 moment AI helped you有记录Have a record
W2W2每个关键输出都验证一次Verify every key output once关键信息 100% 验证Key info 100% verified
W2W2打开 3 个 AI 给的引用Open 3 AI-given citations3 个都验证All 3 verified
W2W2用 2 个模型对比答同一问题Compare 2 models answering the same question看到差异Saw the difference
W2W2让 AI 分步思考解 1 个难题Make AI think step-by-step for 1 hard problem答案更准More accurate answer
W2W2给 AI 2-3 个例子(few-shot)Give AI 2-3 examples (few-shot)输出更贴合Output more on-target
W2W2建一个你的提示库文件夹Build a prompt library folder有 10+ 模板Have 10+ templates
W2W2测试 1 个 AI 的偏见(反例)Test 1 AI's bias (counterexample)指出 1 个偏见Pointed out 1 bias
W2W2用 AI 帮你写 1 篇长文的初稿Use AI to draft 1 long piece初稿+你改灵魂Draft + you shaped the soul
W2W2了解上下文窗口对你的影响Understand how context window affects you能分段长文档Can split long docs
W2W2把 1 个重复任务交给 AIDelegate 1 repetitive task to AI省了 30 分钟+Saved 30 min+
W3W3读 1 篇 LLM 原理科普Read 1 LLM principles explainer能讲清 token/上下文Can explain token/context
W3W3理解温度/采样对输出的影响Understand how temperature/sampling affects output按任务调温度Adjust temperature by task
W3W3了解 RAG 是什么、解决什么问题Understand what RAG is and what it solves能讲清原理Can explain the principle
W3W3了解 AI Agent 能做什么Understand what AI Agent can do能设计 1 个简单 AgentCan design 1 simple Agent
W3W3用多模态 AI 做 1 件事Do 1 thing with multimodal AI用了非文字模态Used a non-text modality
W3W3建立你的 AI 红线清单Build your AI red-line list有 5+ 条红线Have 5+ red lines
W3W3向 1 个人解释 AI 原理Explain AI principles to 1 person对方听懂了The other understood
W3W3评估 2 个 AI 工具的优劣Evaluate the pros/cons of 2 AI tools有对比结论Have a comparison conclusion
W3W3把 1 个 AI 工作流写下来Write down 1 AI workflow有步骤文档Have a step doc
W3W3了解 AI 版权的基本规则Understand the basic rules of AI copyright知道能不能商用Know if it can be commercial
W4W4用多模态+Agent 做 1 个完整任务Do 1 complete task with multimodal+Agent端到端完成Completed end-to-end
W4W4把你的提示库整理成分类Organize your prompt library by category分类清晰Categories clear
W4W4建立你的验证习惯 SOPBuild your verification habit SOP有验证步骤Have verification steps
W4W4用 AI 帮你复盘 1 个项目/决定Use AI to review 1 project/decision有复盘文档Have a review doc
W4W4教 1 个人用 AI(入门)Teach 1 person to use AI (intro)对方会用基础The other can use basics
W4W4制定你的下 30 天 AI 计划Make your next 30-day AI plan有计划Have a plan
W4W4跟踪 1 个 AI 新进展Track 1 AI new progress了解 1 个新东西Learned 1 new thing
W4W4更新你的 AI 红线(按新认知)Update your AI red-lines (by new understanding)红线迭代了Red lines iterated
W4W4把你最好的 5 个提示标为精华Mark your best 5 prompts as essentials有 5 个精华Have 5 essentials
W4W4用 AI 帮你规划 1 个学习目标Use AI to plan 1 learning goal有计划Have a plan
W4W4把 AI 素养自评做 1 次(40 项)Do 1 AI literacy self-assessment (40 items)有分数Have a score
W4W4找到你最弱的 1 个维度,定改进计划Find your weakest 1 dimension, make improvement plan有改进目标Have an improvement goal
持续Ongoing每天 30 分钟用 AIUse AI 30 min daily每天都用Used daily
持续Ongoing每周复盘 1 次 AI 使用Review AI usage 1x weekly有复盘Have review
持续Ongoing每月更新 1 次提示库Update prompt library 1x monthly库在更新Library updating
持续Ongoing每季度重做 1 次自评Redo self-assessment quarterly分数在升Score rising
持续Ongoing跟进 AI 新工具/新能力Follow AI new tools/capabilities了解趋势Aware of trends
持续Ongoing保持先验证的默认Keep verify-first as default从不盲信Never blindly trust
持续Ongoing把 AI 当杠杆,不当替身Treat AI as leverage, not substitute你主导You lead
持续Ongoing分享你的 AI 经验给 1 个人Share your AI experience with 1 person有分享Shared
持续Ongoing持续学习,保持好奇心Keep learning, keep curiosity持续投入Continuous investment
持续Ongoing把 AI 素养变成肌肉记忆Make AI literacy muscle memory自动化了Automated
✅50 个动作,按周做。做完一周,你就比 90% 的人更会用 AI。关键是持续,不是一次。50 actions, do them by week. After one week, you use AI better than 90% of people. The key is continuous, not one-time.
C26

20 个 AI 素养学习路径(按目标)

20 AI Literacy Learning Paths (By Goal)

按 20 种不同目标,各给一条学习路径,对号入座
20 different goals, each with a learning path; pick yours
目标Goal核心路径Core path关键资源Key resources时间Time
日常提效Daily efficiency提示工程 + 验证 + 场景应用Prompting + verification + scenario applicationDeepLearning.AI 短课 + 提示词库DeepLearning.AI short course + prompt library2 周2 weeks
职场转型Career transitionAI 技能 + 作品集 + 面试AI skills + portfolio + interviewsCoursera + 案例库Coursera + case library3 个月3 months
编程入门Programming introAI 结对编程 + 项目 + 验证AI pair programming + projects + verificationGitHub Copilot + Replit + 测试GitHub Copilot + Replit + tests2 个月2 months
内容创作Content creation多模态 + 创意 + 版权Multimodal + creativity + copyrightMidjourney + Runway + 版权规则Midjourney + Runway + copyright rules1 个月1 month
科研提效Research efficiencyRAG + 文献 + 批判RAG + literature + critiqueElicit + arXiv + 批判Elicit + arXiv + critique1 个月1 month
数据分析Data analysisAI 写代码 + 统计 + 验证AI writes code + stats + verification数据科学 + 测试 + 验证Data science + tests + verification2 个月2 months
语言学习Language learningAI 陪练 + 输入 + 输出AI sparring + input + outputAI 对话 + 30 天计划AI conversation + 30-day plan3 个月3 months
创业EntrepreneurshipAI 全栈 + 产品 + 市场AI full-stack + product + marketAI 工具 + 商业计划 + 原型AI tools + business plan + prototype3 个月3 months
教育EducationAI 因材施教 + 出题 + 批改AI personalized teaching + quiz + gradingAI 教学 + 验证 + 伦理AI teaching + verification + ethics1 个月1 month
设计Design多模态 + 审美 + 品牌Multimodal + aesthetics + brandMidjourney + 审美训练Midjourney + aesthetic training1 个月1 month
写作Writing提示 + 结构 + 润色Prompting + structure + polish提示库 + 结构 + 润色Prompt library + structure + polish2 周2 weeks
翻译TranslationAI 翻译 + 校对 + 术语AI translation + proofread + termsDeepL + 术语库DeepL + glossary2 周2 weeks
营销MarketingAI 文案 + 投放 + 分析AI copy + ads + analysisAI 文案 + 数据AI copy + data1 个月1 month
法律LegalAI 查判例 + 起草 + 保密AI search case law + draft + confidentialityAI 法律 + 保密 + 验证AI legal + confidentiality + verification1 个月1 month
金融FinanceAI 分析 + 风控 + 合规AI analysis + risk + complianceAI 金融 + 合规AI finance + compliance2 个月2 months
医疗HealthcareAI 查文献 + 辅助 + 隐私AI literature + assist + privacyAI 医疗 + 隐私AI medical + privacy2 个月2 months
自媒体Self-mediaAI 内容 + 涨粉 + 变现AI content + followers + monetizeAI 内容 + 平台规则AI content + platform rules2 个月2 months
家庭教育Family educationAI 辅导 + 习惯 + 伦理AI tutoring + habit + ethicsAI 教育 + 亲子AI education + parent-child1 个月1 month
银发学习Silver learningAI 入门 + 实用 + 耐心AI intro + practical + patienceAI 入门 + 大字体AI intro + large font1 个月1 month
学术严谨Academic rigor验证 + 溯源 + 批判Verification + tracing + critique学术工具 + 批判Academic tools + critique2 个月2 months
🧭对号入座:找到你的目标,走那条路径。大多数人的路径是日常提效(2 周)再进阶。别贪多,先跑通 1 条。Pick your goal, follow that path. Most people's path is daily efficiency (2 weeks) then advance. Don't be greedy, first run through 1 path.
🎓AI 素养的终点不是学会,而是用好。这 20 条路径,每条的终点都是:你能用 AI 解决你领域的真实问题,并且验证结果。The endpoint of AI literacy isn't learning, it's using well. Each of these 20 paths ends at: you can use AI to solve real problems in your field, and verify the results.
C27

40 个 AI 术语中英对照(扫盲表)

40 AI Terms: Chinese-English Glossary

40 个高频 AI 术语,中英对照 + 一句话解释,扫盲用
40 high-frequency AI terms, bilingual + one-line definition, for beginners
术语Term一句话解释One-line definition
大语言模型Large Language Model (LLM)用海量文本训练的、能理解和生成语言的人工智能模型An AI model trained on massive text that understands and generates language
tokentokenAI 处理文字的最小单位(中文约 1-2 字 = 1 token)The smallest unit AI processes text in (Chinese ~1-2 chars = 1 token)
上下文窗口Context WindowAI 一次能记住的文字量上限The max text AI can remember at once
温度Temperature控制 AI 输出随机性的参数(低=确定,高=创意)A parameter controlling output randomness (low=deterministic, high=creative)
幻觉HallucinationAI 自信地编造不存在的事实AI confidently fabricating non-existent facts
提示工程Prompt Engineering设计提示词让 AI 输出更好的技术Designing prompts to get better AI output
few-shotfew-shot在提示里给 2-3 个例子引导 AI 的格式Giving 2-3 examples in the prompt to guide AI's format
思维链Chain-of-Thought (CoT)让 AI 一步步推理再给答案的方法Making AI reason step-by-step before answering
微调Fine-tuning用特定数据继续训练基础模型的方法Continuing to train a base model with specific data
RAGRAG先检索资料再生成回答的技术(解决知识过时)Retrieve material first, then generate (solves outdated knowledge)
embeddingembedding把文字转成数字向量表示语义的技术Converting text into numeric vectors to represent semantics
AgentAI Agent能自主规划、调用工具、执行多步任务的 AI 系统An AI system that plans, calls tools, executes multi-step tasks
多模态MultimodalAI 能处理文字/图像/音频/视频多种模态的能力AI's ability to process text/image/audio/video
对齐Alignment让 AI 行为符合人类意图和价值观的技术Making AI behavior align with human intent and values
推理Inference用已训练好的模型生成输出的过程(区别于训练)Using a trained model to generate output (vs training)
训练Training用数据让模型学习参数的过程Using data to let the model learn parameters
参数量Parameter Count模型里可学习数字的数量(衡量规模)The number of learnable numbers in the model (measures scale)
开源Open Source模型权重公开、可下载本地用的模型Model weights are public, downloadable to run locally
闭源Closed Source模型权重不公开、只能 API 调用的模型Model weights not public, only via API
APIAPI调用 AI 模型的编程接口The programming interface to call AI models
token 限制Token Limit一次输入/输出的 token 上限The token limit for one input/output
速率限制Rate Limit单位时间内能调用 AI 的次数上限The max calls per unit time
系统提示System Prompt设定 AI 角色和规则的底层提示The underlying prompt setting AI's role and rules
角色设定Role-playing在提示里给 AI 指定一个专业角色Assigning AI a professional role in the prompt
输出格式Output Format要求 AI 用特定格式(表格/JSON/列表)Requiring AI to use a specific format (table/JSON/list)
流式输出StreamingAI 边生成边显示(不用等完)AI displays as it generates (no need to wait)
停止词Stop Word让 AI 在特定词处停止输出的设置A setting to make AI stop at a specific word
采样Sampling从概率分布里选下一个 token 的方法A method to pick the next token from a probability distribution
top-ptop-p只从概率最高的 p 比例词里采样的方法Sampling only from the top p% most probable words
top-ktop-k只从概率最高的 k 个词里采样的方法Sampling only from the top k most probable words
重复惩罚Repetition Penalty降低 AI 重复词概率的参数A parameter lowering AI's repetition probability
知识截止日期Knowledge CutoffAI 训练数据的最后时间(之后的不知道)The last time of AI's training data (doesn't know after)
幻觉率Hallucination RateAI 编造内容占输出的比例(衡量可靠性)The ratio of AI's fabricated content to output (measures reliability)
基准测试Benchmark用标准题目测模型能力的评测Testing model ability with standard questions
MMLUMMLU测模型多学科知识的大基准A big benchmark testing multi-discipline knowledge
GPT-4oGPT-4oOpenAI 的多模态模型(能听能说)OpenAI's multimodal model (can listen and speak)
ClaudeClaudeAnthropic 的模型(长上下文、安全)Anthropic's model (long context, safe)
GeminiGeminiGoogle 的模型(多模态强)Google's model (strong multimodal)
LlamaLlamaMeta 的开源模型(可本地部署)Meta's open-source model (local deployable)
DeepSeekDeepSeek深度求索的开源模型(性价比高)DeepSeek's open-source model (good value)
QwenQwen阿里的开源模型(中文强)Alibaba's open-source model (strong Chinese)
GPTGPTOpenAI 的生成式预训练 Transformer 系列OpenAI's Generative Pre-trained Transformer series
TransformerTransformer当前大模型的基础架构The base architecture of current large models
📖这张表是扫盲用。不用全记,用到哪个查哪个。30 天后,你读 AI 文章就不会被术语卡住。This table is for basics. Don't memorize all; look up when you use one. After 30 days, reading AI articles won't be blocked by terms.
C28

15 条 AI 使用红线(绝不碰)

15 AI Usage Red Lines (Never Touch)

15 条绝对不能碰的 AI 使用红线,碰了就是真风险
15 absolute AI usage red lines; touching them = real risk
红线Red line为什么危险Why dangerous正确做法Right way
用 AI 伪造证件/签名/身份Using AI to forge ID/signatures/identity违法(伪造文件罪)Illegal (forgery crime)只用 AI 做合法内容Only use AI for legal content
用 AI 生成深度伪造骗人Using AI to generate deepfakes to deceive违法 + 道德(侵犯人格)Illegal + immoral (violates personality)不生成假真人Don't generate fake real people
把他人隐私喂给 AIFeeding others' privacy to AI侵犯隐私Infringes privacy匿名化后再用Anonymize before use
把公司机密喂给公有 AIFeeding company secrets to public AI泄露 + 违反保密协议Leak + breach of NDA用企业版/本地Use enterprise/local
用 AI 代学生作业并谎称自做Using AI to do student homework and claiming it's your own学术不端Academic misconduct用 AI 辅助,如实标注Use AI to assist, label honestly
用 AI 做医疗诊断并照做Using AI for medical diagnosis and following itAI 会误诊,害身体AI may misdiagnse, harming the bodyAI 查文献,医生下诊断AI looks up literature, doctor diagnoses
用 AI 做法律判断并据此行动Using AI for legal judgment and acting on itAI 引用可能不存在,害自己AI citations may not exist, harming youAI 辅助,律师确认AI assists, lawyer confirms
用 AI 做投资并全仓押上Using AI for investment and going all-inAI 不预测未来,可能亏光AI doesn't predict the future, may lose everythingAI 分析,自己决策,分散AI analyzes, you decide, diversify
用 AI 生成内容冒充真人Using AI-generated content impersonating a real person欺骗 + 侵权Deception + infringement标注 AI 生成Label as AI-generated
用 AI 写代码不测试直接上线Using AI code without testing, going live可能有大 bug,出事故May have big bugs, causing incidents跑测试 + 读逻辑 + 查边界Run tests + read logic + check edges
用 AI 查实时数据当准Using AI's real-time data as accurate训练有截止日期,可能过时Training has a cutoff, may be outdated先查权威来源,再 AI 整理Check authoritative sources first, then AI organizes
用 AI 做歧视性内容Using AI for discriminatory content违法 + 伤害Illegal + harmful主动测偏见,拒绝歧视Actively test bias, refuse discrimination
用 AI 绕过版权(洗稿)Using AI to bypass copyright (rewriting)侵权Infringement尊重原创,标注参考Respect originals, label references
让 AI 替你做重大人生决定Letting AI make major life decisions for youAI 没有你的处境和价值观AI has neither your circumstances nor valuesAI 给分析,你做决定AI analyzes, you decide
把密码/密钥告诉 AITelling AI your passwords/keys泄露,账号被盗Leak, account stolen密码只在你和系统之间Passwords only between you and the system
🚫这 15 条红线,碰任何一条都可能是真损失(法律/金钱/健康/账号)。把这张表存手机,用 AI 前扫一眼。红线不是限制,是保护。These 15 red lines — touching any may be real loss (legal/money/health/account). Save this table to your phone, glance before using AI. Red lines aren't restrictions, they're protection.
🛡️AI 素养的最后一层,是『知道边界』。会用 AI 的人很多,但知道『哪些绝不能做』的人,才是真正安全的人。这 15 条红线,就是你的安全底线。The last layer of AI literacy is 'knowing the boundaries'. Many can use AI, but those who know 'what absolutely must not be done' are truly safe. These 15 red lines are your safety baseline.
C29

20 个高频问题(一问一答,扫盲)

20 High-Frequency Questions (Q&A, Basics)

20 个最常见的 AI 问题,直接给答案,扫盲用
20 most common AI questions, direct answers, for basics
C30

一页速记卡(10 条核心,贴在桌前)

One-Page Cheat Card (10 Core, Stick on Your Desk)

10 条最核心的 AI 素养要点,贴在桌前,每天看
10 most core AI literacy points, stick on your desk, see them daily
核心Core一句话One-liner
11AI 是模式匹配,不是真理解AI is pattern matching, not real understanding
22关键输出必验证(引用/数字/代码)Verify key outputs (citations/numbers/code)
33给角色 + 背景 + 格式 + 约束Give role + context + format + constraints
44分步思考 + 少样本 + 自检Step-by-step + few-shot + self-check
55敏感信息绝不喂公有 AINever feed sensitive info to public AI
66AI 给初稿,你给灵魂 + 把关AI gives draft, you give soul + gatekeeping
77红线不碰(法律/健康/账号/隐私)Don't touch red lines (legal/health/account/privacy)
88用 AI 强化你,不是替代你Use AI to strengthen you, not replace you
99持续跟进新工具,底层不变Keep up with new tools, underlying unchanged
1010先验证,再使用,最后信任Verify first, then use, finally trust
🎓这 10 条,是 AI 素养的全部精华。贴在桌前,用 AI 前扫一眼。三个月后,你会变成一个『既会用 AI,又懂 AI,又不被 AI 骗』的人——这正是 AI 素养的终点。These 10 points are the entire essence of AI literacy. Stick them on your desk, glance before using AI. In three months, you'll become a person who 'can use AI, understands AI, and isn't fooled by AI' — that's exactly the endpoint of AI literacy.
31

AI 素养 10 条常识速查

10 AI Literacy Common-Sense Facts

📋 10 条 AI 素养常识(30 秒读完)

📋 10 AI Literacy Common Sense (Read in 30s)

常识Fact记住什么Remember
AI 会编AI hallucinates关键信息必须验证Verify key info
AI 无长期记忆No long-term memory重要信息要重复/存外部Repeat/save important info
AI 有截止日期Knowledge cutoff实时数据要另查Check real-time separately
AI 不是专家Not an expert高风险找真人Find real expert for high-risk
AI 有偏见AI has bias多源验证,别单信Multi-source verify
AI 快但浅Fast but shallow深度任务要迭代Iterate for deep tasks
AI 能辅助不能替代Assist not replace你主导决策You lead decisions
AI 数据会被用Data may be used敏感信息别喂Don't feed sensitive info
AI 工具在变Tools change学底层方法Learn underlying method
AI 是杠杆AI is leverage放大你的能力Amplify your ability
✅这 10 条常识,是 AI 素养的地基。记住它们,你就不会被 AI 的『看起来很对』骗到。These 10 common-sense facts are the foundation of AI literacy. Remember them and you won't be fooled by AI's 'looks-right'.