要点总结
- What ChatGPT does and how it works. ChatGPT is a conversational service built around large language models that generate responses from patterns learned during training and subsequent development.
- Where generative AI adds value. Public organisations can use AI to prepare first drafts of meeting minutes, public guidance, standard correspondence and briefing structures.
- Risks that require active controls. Models can invent sources, misinterpret numerical tables or produce uneven results across languages and groups.
- 如何采购和治理人工智能。NIST人工智能风险管理框架及其生成式人工智能配置文件为绘制危害图、衡量绩效和管理风险提供了有用的结构。
- 公共机构和捐赠方资助的项目。从错误可逆且人工审查切实可行的狭窄用例开始。
In short: Practical guide to ChatGPT and generative AI: key decisions, risks, examples and lessons for ministries, public agencies and donor-funded project teams.
Free self-check: the governance self-audit gives a quick check of governance, controls and accountability, in about 10 minutes.
In short: Chatgpt and generative ai is best understood through practical decisions, institutional responsibilities and evidence of results. This guide explains the mechanisms, trade-offs and safeguards that determine whether policies and projects deliver lasting benefits.
ChatGPT and Generative AI: Capabilities, Limits and Governance is not a single solution or a universal recipe. Results depend on the context, the capabilities of implementing organisations and the people affected. The sections below distinguish what is established, what must be tested locally and what project teams can do before making irreversible commitments.
What ChatGPT does and how it works
语言模型与对话界面
ChatGPT is a conversational service built around large language models that generate responses from patterns learned during training and subsequent development. It can draft, classify, summarise and help with code, but plausible language is not proof of accuracy.
Practical example: Ask it to summarise a policy document, then verify each important statement against the actual policy. Keep a record of the prompt, source version and human reviewer.
有益的第一步是区分可观察到的状况与其推测的原因。团队应记录证据由谁提供、涵盖哪个时间段,以及哪些其他解释可能产生相同的规律。这样,在情况发生变化时,建议会更站得住脚。
在作出决定之前,先明确决策者、受影响的用户以及决策失误的代价。开展小规模试点或审查可比案例,通常比仅凭一次演示就采用一套永久性系统代价更低。
Multimodal work and tool use
Contemporary AI systems may work with text, images, documents and other inputs, depending on the product and permissions. Some can use connected tools to retrieve information or complete workflows, which changes both their usefulness and their security exposure.
Practical example: A procurement officer can extract requirements from a scanned tender, but should confirm quantities, dates and exceptions directly from the source before publishing.
实施应从基线和对现行工作流程的描述开始。没有这一基线,后来的改善可能只是反映了需求变化、临时增加的资金或报告方式的改变,而非干预措施本身的作用。
职责应以可操作的方式写明:由谁收集信息、由谁批准例外情况、由谁有权叫停不安全或无效的工作。答案应让员工、供应商和监督机构都能看到。
Where generative AI adds value
行政起草与分析
Public organisations can use AI to prepare first drafts of meeting minutes, public guidance, standard correspondence and briefing structures. Staff time saved is valuable only if the review process prevents fabricated citations and incorrect advice.
Practical example: Pilot AI on anonymised, low-risk correspondence and compare correction time, accuracy and accessibility with the existing process.
决策者需要同时考虑眼前的收益和后续的义务。资本支出只是成本的一部分;培训、维护、更新、无障碍和申诉处理,可能决定这项投资能否持续发挥作用。
可信的方案评估应包括一个成本较低的替代方案,以及预期需求或绩效未能实现的情景。公开假设,便可以在不掩盖最初不确定性的情况下修订计划。
Education, software and citizen services
AI can create learning exercises, explain code and assist help desks. It should not be treated as a substitute for teachers, qualified clinicians, caseworkers or formal eligibility decisions.
Practical example: A service desk can offer AI-assisted answers restricted to an approved knowledge base, with escalation to a human for uncertain or sensitive cases.
对证据的解读应把握适当的层面。某个设施、社区或机构取得的改善,未必适用于基础设施、人员配置、激励机制或风险暴露不同的其他地方。
项目经理可以通过列出先前干预措施取得成功的条件,并核查这些条件在本地是否存在,来检验其可移植性。如果条件不存在,应调整设计,而不是照搬表面结果。
Risks that require active controls
Hallucinations, bias and overconfidence
模型可能会虚构来源、误解数字表格或在不同语言和群体之间产生不均衡的结果。即使问题很简单,自信的回答也可能是错误的。
Practical example: Use a test set containing difficult local-language queries, conflicting documents and deliberately missing information. Score abstention and escalation as well as correctness.
利益相关方参与 在重大选择尚未敲定之前最为有用。磋商可以揭示行政数据中所缺失的运营约束、非正式做法和分配效应。
不仅要记录征询了多少人的意见,还要记录哪些关切改变了设计。在意见冲突时,用受影响人群能够理解的语言说明所作的权衡以及申诉或投诉机制。
Privacy, security and intellectual property
Sensitive data entered into an external system may be processed under terms that differ from an institution’s requirements. Uploaded documents can contain personal data, confidential bids or malicious instructions.
Practical example: Before a pilot, approve data classes, retention settings, access rights and incident response. Never upload identifiable patient records or sealed bids without a lawful and secure arrangement.
有益的第一步是区分可观察到的状况与其推测的原因。团队应记录证据由谁提供、涵盖哪个时间段,以及哪些其他解释可能产生相同的规律。这样,在情况发生变化时,建议会更站得住脚。
在作出决定之前,先明确决策者、受影响的用户以及决策失误的代价。开展小规模试点或审查可比案例,通常比仅凭一次演示就采用一套永久性系统代价更低。
决策清单:
- 明确问题及受影响人群。
- 确认基线及其局限。
- 至少比较两个可行方案。
- 为每项重大风险指定一名负责人。
- 明确可衡量的成果和审查日期。
How to procure and govern AI
Risk assessment and human accountability
The NIST AI Risk Management Framework and its generative AI profile offer useful structures for mapping harms, measuring performance and managing risks. Accountability remains with the deploying institution.
Practical example: Create an AI use-case register identifying owner, purpose, affected groups, permitted data, reviewer and shutdown procedure.
实施应从基线和对现行工作流程的描述开始。没有这一基线,后来的改善可能只是反映了需求变化、临时增加的资金或报告方式的改变,而非干预措施本身的作用。
职责应以可操作的方式写明:由谁收集信息、由谁批准例外情况、由谁有权叫停不安全或无效的工作。答案应让员工、供应商和监督机构都能看到。
Vendor evaluation and measurable testing
Procurement should examine interoperability, accessibility, information security, model updates, audit logs and exit options. Marketing demonstrations rarely represent day-to-day accuracy on local tasks.
Practical example: Require vendors to demonstrate performance on representative documents and provide a documented process for challenging or correcting outputs.
决策者需要同时考虑眼前的收益和后续的义务。资本支出只是成本的一部分;培训、维护、更新、无障碍和申诉处理,可能决定这项投资能否持续发挥作用。
可信的方案评估应包括一个成本较低的替代方案,以及预期需求或绩效未能实现的情景。公开假设,便可以在不掩盖最初不确定性的情况下修订计划。
公共机构与捐助方资助项目
Choosing a safe first pilot
Start with a narrow use case where errors are reversible and human review is practical. Define baseline cost and quality before adding automation, and do not scale simply because staff like the interface.
Practical example: A donor project can test AI-assisted project-report drafting using public data, measuring factual corrections and staff hours over a fixed trial.
对证据的解读应把握适当的层面。某个设施、社区或机构取得的改善,未必适用于基础设施、人员配置、激励机制或风险暴露不同的其他地方。
项目经理可以通过列出先前干预措施取得成功的条件,并核查这些条件在本地是否存在,来检验其可移植性。如果条件不存在,应调整设计,而不是照搬表面结果。
Institutional capability and oversight
Effective adoption needs staff literacy, procurement rules, records management and a route for citizens to challenge decisions. Training alone is insufficient if system ownership and budgets are unclear.
Practical example: Build an implementation plan covering policy, training, data governance, evaluation and independent assurance before production use.
利益相关方参与在重大选择确定之前最能发挥作用。征询意见可以揭示行政数据中没有反映的运营约束、非正式做法和分配效应。
不仅要记录征询了多少人的意见,还要记录哪些关切改变了设计。在意见冲突时,用受影响人群能够理解的语言说明所作的权衡以及申诉或投诉机制。
实施演练:组建一个跨职能小组,成员包括服务负责人、技术专家、财务代表,以及一位能够代表预期用户发声的人员。给他们设定一个现实的约束:资金有限、信息不完整,且须在一个月内作出决策。请每位参与者提出一项行动、一个替代方案,以及其首选方案可能失败的一个原因。商定最小的安全测试、其预算、足以支持扩大规模的证据,以及触发暂停的条件。在下一次审查时,将团队实际观察到的情况与假设和观点区分开来,并记录调整方向的理由。这项演练刻意比一般的意识提升研讨会要求更高,因为它检验的是决策权、证据质量、运营可行性以及应对不利发现的能力。它既可用于单个设施,也可用于多机构项目,而无需假定所有情境都面临相同的约束。
For ministries and project implementation units, the next step is to convert these principles into a short action plan. Identify one bottleneck, the evidence needed to verify it, the responsible official, the first corrective action and the date on which management will review the result. The Risalat Diagnostic Centre can support a structured initial discussion, while the programme categories provide routes for developing staff capability. Organisations requiring tailored support can request a proposal.
供进一步核查的部分原始来源:
- NIST Generative AI Profile
- NIST AI Risk Management Framework
- UNESCO Recommendation on the Ethics of AI
常见问题
What does ChatGPT and generative AI mean in practice?
这意味着依据可靠的信息、明确的职责和可衡量的成果,作出因地制宜的选择,而不是依赖某一种技术或政策。
How should an organisation begin work on ChatGPT and generative AI?
在投入大量资源之前,先建立基线,征询受影响群体的意见,比较可行方案,并指定决策负责人。
需要管理的主要风险有哪些?
常见风险包括数据薄弱、能力不足、隐性运营成本、影响不均以及纠正措施缺乏问责。
项目团队如何衡量成果?
使用少量基线和跟踪指标,尽可能独立核实记录,并在报告成绩的同时说明局限。
公共机构如何应用这些经验?
将这一方法纳入采购、预算、保障、报告和长期运营之中,明确负责人并定期开展管理审查。
观看
- 开放合同伙伴关系(Open Contracting Partnership) 开放合同伙伴关系(Open Contracting Partnership)
- 经济合作与发展组织(OECD) OECD
Know people who should read this? Become an Ambassador or write with Risalat.
自2011年以来始终是项目成功的合作伙伴,为政府部委、公共机构和项目团队提供诊断评估、咨询支持与应用型学习。



