📚 UK Evidence Principles for AI in Education | 教育中应用人工智能：英国提出的决策原则
Artificial intelligence has moved from a novelty in the classroom to a routine part of how teachers plan lessons and how students revise. In response, the UK government has set out a set of evidence principles: a decision framework that asks school leaders, teachers and buyers to judge AI tools by the quality of the evidence behind them rather than by the polish of the sales pitch. This article explains what those principles mean, why they exist, and how they can be applied in real classrooms and in your own revision routine.
人工智能已经从课堂里的新鲜事物，变成了教师备课和学生复习中的日常工具。作为回应，英国政府提出了一套"证据原则"：这是一个决策框架，要求学校管理者、教师和采购者依据产品背后的证据质量来判断人工智能工具，而不是依据销售演示的精致程度。本文将解释这些原则的含义、它们为何存在，以及如何把它们应用到真实课堂和你自己的复习安排中。
1. Why Evidence Matters When Schools Adopt AI | 为什么学校采用人工智能时需要证据
Every week brings new AI products promising to cut marking time, personalise learning or raise attainment. The problem is not that AI fails; the problem is that tools are frequently adopted on the strength of a demonstration, a free trial or a colleague's enthusiasm, and the learning outcome is never actually measured. When that happens, money, teacher time and pupil attention are all spent on an intervention with an unknown return.
每周都有新的人工智能产品承诺缩短批改时间、实现个性化学习或提升成绩。问题不在于人工智能会失败，而在于许多工具是凭着一次演示、一次免费试用或同事的热情被采用的，而学习成效从未被真正测量。一旦如此，资金、教师时间和学生注意力就被投入到一个回报未知的干预措施上。
- Demo effect: a smooth product demonstration shows what a tool can do, not what pupils learn. | 演示效应：流畅的产品演示展示的是工具能做什么，而不是学生学到了什么。
- Novelty effect: early gains often fade once the tool stops being new. | 新奇效应：早期出现的提升往往在工具不再新鲜后消退。
- Opportunity cost: budget spent on an unevaluated tool is budget not spent on small-group tutoring, reading interventions or staff development. | 机会成本：花在未经评估的工具上的预算，就是没有花在小组辅导、阅读干预或教师发展上的预算。
- Equity risk: poorly planned AI adoption can widen the gap between pupils with reliable devices and connectivity at home and those without. | 公平风险：规划不当的人工智能应用，可能加大家中有稳定设备和网络的学生与其他学生之间的差距。
2. The UK Position: A Framework, Not an Approved List | 英国的立场：是框架，而不是"白名单"
The Department for Education (DfE) has published policy and guidance on generative AI in education, and has followed this with evidence principles designed to support decision-making about AI products in schools and colleges. The important point for teachers is what these principles are not: they are not a list of approved tools, and they do not certify any particular product. They are a set of questions that a buyer or a school leader should be able to answer before committing time, money and pupil data.
英国教育部（DfE）发布了关于教育领域生成式人工智能的政策与指南，并在此基础上提出了"证据原则"，用以支持中小学和学院在人工智能产品方面的决策。对教师而言，关键要理解这些原则不是什么：它们不是一份获批工具清单，也不为任何特定产品背书。它们是一组问题，采购者或学校管理者在投入时间、资金和学生数据之前，应当能够回答这些问题。
| Principle | 原则 | Core question | 核心问题 |
|---|---|
| Start with the problem | 从问题出发 | What educational need are we trying to solve, and how will we know if we have solved it? | 我们要解决什么教育需求，又如何知道它已被解决？ |
| Check the strength of evidence | 核查证据强度 | Is the claim supported by independent, well-designed research rather than a case study? | 该主张是否有独立、设计良好的研究支持，而不是仅靠案例？ |
| Consider relevance and context | 考虑相关性与情境 | Was the evidence generated with pupils, subjects and settings like ours? | 证据是否来自与我们相似的学生、学科和学校环境？ |
| Scrutinise claims | 审视宣传主张 | Who produced the evidence, and who benefits from the conclusion? | 证据由谁产生，结论又对谁有利？ |
| Protect pupils and data | 保护学生与数据 | Does the tool comply with data protection, safeguarding and accessibility duties? | 该工具是否符合数据保护、安全保障和无障碍方面的义务？ |
| Evaluate and monitor | 评估与监测 | How will impact be measured, and what would make us stop using it? | 如何衡量成效，什么情况下我们会停止使用？ |
Guidance is revised periodically, so always read the current version on gov.uk before quoting wording in an essay or a policy document.
指南会定期修订，因此在论文或学校政策文件中引用具体表述之前，务必查阅 gov.uk 上的最新版本。
3. Principle One: Start With the Educational Problem | 原则一：从教育问题出发
AI should be the answer to a question you have already asked, not the starting point. A useful problem statement is specific, measurable and bounded: for example, "Year 10 pupils in set 3 lose an average of two marks per question on multi-step algebra because they skip the planning step", rather than "we want to use AI".
人工智能应当是你已经提出的问题的答案，而不是起点。有用的问题陈述应当是具体、可测量、有边界的：例如"10年级3层班的学生在多步代数题上平均每题丢两分，因为他们跳过了规划步骤"，而不是"我们想用人工智能"。
A well-defined problem also allows you to compare AI against non-AI alternatives, which is where many procurement decisions go wrong. If targeted feedback from a teaching assistant produces the same gain at a lower cost, the AI tool has not earned its place.
定义清晰的问题还能让你把人工智能与"非人工智能"的替代方案进行比较，而许多采购决策正是在这里出错的。如果助教提供的针对性反馈能以更低的成本带来同样的提升，那么这个人工智能工具就没有赢得它的位置。
Attainment gap = Target outcome − Current baseline
Write the gap down as a number before you pilot anything. A tool that cannot plausibly move that number is not a solution to this problem, however impressive its features are.
在试点任何东西之前，把这个差距写成一个数字。一个不可能推动这个数字变化的工具，无论功能多么令人印象深刻，都不是这个问题的解决方案。
4. Principle Two: Judge the Strength and Type of Evidence | 原则二：判断证据的强度与类型
Not all evidence is equal. At the bottom sit testimonials, conference talks and vendor case studies; at the top sit randomised controlled trials and systematic reviews that pool many trials. The key question to ask is not simply "is there evidence?" but "evidence of what?". Engagement, enjoyment and time saved are outcomes worth measuring, but they are not attainment, and they do not always correlate with it.
并非所有证据都同等可靠。底层是用户证言、会议演讲和厂商案例；顶层是随机对照试验，以及汇总多项试验的系统综述。关键问题不只是"有证据吗"，而是"是什么的证据"。参与度、愉悦感和节省的时间都值得测量，但它们不等于学业成绩，也不总与成绩相关。
- Peer-reviewed meta-analysis: pools many studies, allows effect sizes to be compared. | 同行评审的元分析：汇总多项研究，可比较效应量。
- Randomised controlled trial: the strongest single-study design for causal claims. | 随机对照试验：检验因果关系最强的单项研究设计。
- Quasi-experiment: useful when randomisation is impractical, but vulnerable to selection bias. | 准实验：在无法随机分组时有用，但容易受选择偏差影响。
- Pre-post study with no control group: cannot separate the intervention from normal progress, maturation or testing effects. | 无对照组的前后测研究：无法把干预效果与正常进步、成长或测试效应区分开。
- Anecdote and testimonial: generates hypotheses, does not test them. | 轶事与证言：能产生假设，但不能检验假设。
5. Principle Three: Relevance and Context Matter | 原则三：相关性与情境至关重要
Evidence is always evidence-in-a-context. A study conducted with undergraduates using AI to summarise lecture notes tells you very little about whether a 14-year-old will learn more about the causes of the First World War. External validity - the extent to which findings transfer to your setting - has to be judged deliberately.
证据总是"特定情境中的证据"。一项以大学生用人工智能总结讲座笔记为对象的研究，几乎无法说明一名14岁的学生是否会因此更好地理解一战的起因。外部效度，也就是研究结论迁移到你所处情境的程度，必须被有意识地加以判断。
- Age and phase: primary, secondary, further and higher education behave differently. | 年龄与学段：小学、中学、继续教育和高等教育表现不同。
- Subject and task: essay feedback, translation drills and problem sets are not interchangeable. | 学科与任务：作文反馈、翻译练习和习题集不可互换。
- Pupil characteristics: SEND, EAL, prior attainment and reading age all moderate effects. | 学生特征：特殊教育需求、英语作为附加语言、先前成绩和阅读年龄都会调节效果。
- Implementation: a tool used for 10 minutes a week is a different intervention from one used daily. | 实施方式：每周使用10分钟的工具与每天使用的工具是两种不同的干预。
6. Principle Four: Scrutinise the Claims | 原则四：审视宣传主张
Much of the evidence circulating about AI in education is produced or funded by the companies selling the products. That does not automatically make it false, but it does mean it should be read with the incentives in mind. Look for the sample size, the comparison group, the outcome measure, the effect size and the confidence interval. If any of these is missing, treat the headline claim as marketing rather than evidence.
教育人工智能领域流传的许多证据，是由销售这些产品的公司制作或资助的。这并不自动意味着它是假的，但确实意味着阅读时要把利益动机考虑进去。要寻找样本量、对照组、结果测量指标、效应量和置信区间。如果这些信息缺失，就把那个醒目的结论当作营销，而不是证据。
- Check who wrote it, who paid for it and whether it was independently reviewed. | 核查谁撰写的、谁出资的、是否经过独立评审。
- Check whether the comparison was a genuine control or simply "before versus after". | 核查对照是真正的对照组，还是仅仅是"前后对比"。
- Check whether the reported outcome is learning or merely usage. | 核查报告的结果是学习成效，还是仅仅使用率。
- Be alert to the halo effect: enthusiastic early adopters can produce gains that do not survive scaling. | 警惕光环效应：热情的早期使用者可能带来无法在规模化后持续的效果。
7. Principle Five: Safety, Data Protection and Ethics | 原则五：安全、数据保护与伦理
In the UK, any use of pupil data must comply with the UK GDPR and the Data Protection Act 2018, and schools must consider the Information Commissioner's Office guidance and the DfE's product safety expectations for generative AI. Practical questions include where data is stored and processed, whether inputs are used to train models, what the minimum age of use is, and how outputs are moderated.
在英国，任何使用学生数据的做法都必须遵守英国《通用数据保护条例》（UK GDPR）和2018年《数据保护法》，学校还必须考虑信息专员办公室（ICO）的指南以及教育部对生成式人工智能产品安全性的期望。实际问题包括：数据在哪里存储和处理、输入内容是否被用于训练模型、最低使用年龄是多少，以及输出内容如何被审核。
- Data minimisation: do not upload names, SEN information or pastoral records into public models. | 数据最小化：不要把姓名、特殊教育需求信息或学生关怀记录上传到公开模型中。
- Transparency: pupils and parents should know when AI is being used to assess or support learning. | 透明度：当人工智能被用于评估或支持学习时，学生和家长应当知情。
- Bias and accuracy: models can produce confident but wrong output, and can reproduce stereotypes. | 偏见与准确性：模型可能给出自信但错误的答案，也可能复制刻板印象。
- Academic integrity: check current JCQ and awarding organisation guidance on AI use in coursework and assessments. | 学术诚信：查阅JCQ和考试局关于在课程作业与考试中使用人工智能的最新规定。
- Accessibility and equity: check that the tool works for pupils using assistive technology. | 无障碍与公平：核查工具对使用辅助技术的学生是否可用。
8. Principle Six: Plan the Evaluation Before You Buy | 原则六：在购买之前就规划好评估
The final principle is the one most often skipped. Evaluation should be designed at the same time as the pilot, not bolted on afterwards. That means agreeing a baseline measure, a comparison group where possible, a small number of pre-specified outcomes, and a review date at which the school is genuinely willing to stop.
最后一条原则是最常被跳过的。评估应当与试点同时设计，而不是事后补上。这意味着要确定基线测量指标、尽可能设置对照组、预先设定少量结果指标，以及一个学校真正愿意停止使用该工具的复核日期。
- Baseline: record prior attainment and the specific skill being targeted. | 基线：记录先前成绩以及被针对的具体技能。
- Comparison: use a parallel class, a delayed-start group or historical data with caution. | 对照：使用平行班、延迟开始组，或谨慎使用历史数据。
- Implementation fidelity: record how often the tool was actually used, by whom and for how long. | 实施保真度：记录工具实际被使用的频率、使用者和时长。
- Cost: calculate cost per pupil per year, including training and technical support. | 成本：计算每名学生每年的成本，包括培训和技术支持。
- Exit criteria: state in advance what result would lead you to discontinue the tool. | 退出标准：事先说明什么样的结果会导致停用该工具。
9. Evidence Hierarchy and Effect Size in Practice | 证据层级与效应量的实际运用
Two quantitative ideas make the principles operational. The first is the hierarchy of study designs; the second is the effect size, which expresses the size of a difference in standard deviation units so that results from different tests can be compared.
有两个定量概念能让这些原则变得可操作。第一个是研究设计层级；第二个是效应量，它以标准差为单位表达差异大小，从而使不同测试的结果可以相互比较。
Cohen's d = (M₁ − M₂) ÷ SD pooled
95% confidence interval = mean ± 1.96 × standard error
An effect size of about +0.4 is commonly treated as a benchmark for a worthwhile educational intervention, and the Education Endowment Foundation converts effect sizes into approximate "months of additional progress". Treat that conversion as a rough communication tool, not a precise law, and always check whether the confidence interval crosses zero - if it does, the study cannot rule out no effect at all.
效应量约为 +0.4 通常被视为一项值得投入的教育干预的基准，教育捐赠基金会（EEF）会把效应量换算成近似的"额外进步月数"。请把这个换算当作粗略的沟通工具，而不是精确的定律，并且始终检查置信区间是否跨越零——如果跨越零，这项研究就无法排除"完全没有效果"。
| Design | 设计 | Causal strength | 因果强度 | Typical use | 典型用途 |
|---|---|---|
| Systematic review / meta-analysis | 系统综述／元分析 | High | 高 | Choosing between approaches | 在多种方法间取舍 |
| Randomised controlled trial | 随机对照试验 | High | 高 | Testing a specific tool | 检验特定工具 |
| Quasi-experiment | 准实验 | Moderate | 中 | School-level pilots | 学校层面的试点 |
| Pre-post without control | 无对照前后测 | Low | 低 | Monitoring, not proving | 用于监测，而非证明 |
| Testimonial | 用户证言 | Very low | 极低 | Generating ideas | 产生想法 |
10. Applying the Principles: A Practical Checklist | 实践应用：一份可操作的清单
For school leaders, the principles translate into a small number of routines that can be built into existing processes rather than added on top of them. The checklist below can be dropped into a department meeting or a purchasing conversation.
对学校管理者而言，这些原则可以转化为少量常规做法，嵌入现有流程，而不是额外叠加。
Published by TutorHao | Exam Prep Revision Series | aleveler.com
更多咨询请联系16621398022（同微信）
或截图后用「扫一扫 → 相册」选择本图