AI 治理概况
United Kingdom AI Governance Profile
英国人工智能治理的初步观察概况,以行业和监管机构主导的治理、对重大完全自动化决策具有约束力的数据法律保障、公共部门指导、人工智能安全基础设施和行业实施为中心。

概况速览
执行摘要
Key signals for United Kingdom AI governance, structured for quick institutional review before the full profile analysis below.
地区
Europe
治理模式
Sectoral and standards-led governance
监管成熟度
Advanced企业影响
High公共部门准备度
Advanced执法成熟度
Developing企业影响
企业应将人工智能应用案例与现有的数据、消费者、就业、职业和行业法规相匹配。对于完全基于自动化流程、使用个人数据的重大决策,在满足适用范围条件的情况下,应实施法定保障措施。监管沙盒可以明确义务,但不能授予豁免。应将AISI和公共部门保证指南作为风险管理基准,但不要将其与具有约束力的法律混淆。
阅读完整企业影响分析主要政策与监管信号
- 2025 年数据(使用和访问)法案——重大完全自动化决策
- 英国政府人工智能行动指南
- 人工智能安全研究所
- 咨询人工智能增长实验室——法律服务
- DSIT AI Risk Management Toolkit
关键机构
- 科学、创新和技术部
- 信息专员办公室
- 人工智能安全研究所
- 行业监管机构
- 法律服务委员会/SRA/CLC
- MI5 / AI Threat Assessment Centre
- AI Economics Institute — announced additional research support, 24 October 2026
- AI Biosecurity Lab — announced institution; operational delivery to monitor
观察清单
我们正在关注什么
可能改变治理格局的后续里程碑、尚未解决的监管问题及实施进展。
- 未来针对人工智能的立法或对创新促进框架的修改
- 自动决策条款的执行
- 咨询人工智能增长实验室的成果
- 人工智能安全研究所测试和评估政策
- 监管机构协调和特定行业的AI指导
- 公共部门劳动力能力和采购实施
- ICO agentic-AI call for evidence closes 2026-11-20; monitor subsequent draft/final guidance rather than rely on the older pipeline forecast
- ICO ADM/profiling update: drafting following closed consultation; final expected Winter 2026
- ICO monitoring of foundation-model developer commitments after its 2026-10-08 supervisory report; final foundation-model guidance remains separate and in development
- Follow-on AISI frontier-model evaluation security and assurance practice
- MHRA draft guidance on managing changes to AI-enabled medical devices expected by December 2026; qualification/classification consultation planned for 2027 and full Commission-response roadmap by Spring 2027
- AI Airlock Phase 3 post-market surveillance outputs; current review-wave applications before 2026-10-31, with rolling applications remaining open afterwards and further shortlisting planned for Spring 2027
- UK national AI resilience programme delivery and any actual legislative changes on serious AI incidents/near misses; the 2026-10-12 announcement is not a universal reporting duty
- AITAC published institutional outputs and AI threat-assessment arrangements; no exact formation day has been confirmed
- Outcomes of ongoing ICO agentic-AI enquiries; distinguish inquiries from established violations or sanctions
- Canada–UK AI security partnership: coordinated evaluation work, possible staff exchanges and interoperable standards outputs after the 2026-10-13 renewal
- Delivery and funding of the 24 October AI investment package, AIBL, National Cyber Shield and NERV Centre; distinguish announced commitments and private model credits from operational public spending
- Whether ministers formally adopt CST’s July AI-sovereignty recommendations published 19 October; advice is not an adopted statutory strategy
- Monitor implementation and any further official updates: Limited EU AI Act application in Northern Ireland — UK guidance
- Monitor implementation and any further official updates: Agentic AI resilience proposals factsheet
- Monitor implementation of Decision 3/2026, including its limited product-market scope and exclusions
执行摘要
英国继续推行鼓励创新、监管主导的模式,利用现有的法律体系和行业监管机构,并结合中央人工智能政策和安全机构。2025 年《数据(使用和访问)法案》取代了之前的英国 GDPR 第 22 条框架,以新的条款规范完全基于自动化处理的重大决策,并要求采取包括信息公开、质疑或代表以及人工干预在内的保障措施。政府的《人工智能行动手册》提供了不具约束力的公共部门实施原则,而人工智能安全研究所则支持安全研究和评估。2026 年,人工智能增长咨询实验室启动,提供法律服务,协调多个监管机构,帮助创新者了解现有要求,但不提供监管批准或豁免。
治理架构
英国的治理体系融合了具有约束力的数据法律保障、行业监管、监管机构指南、公共部门实施原则以及前沿人工智能安全基础设施。人工智能手册和人工智能安全研究所并非通用的人工智能监管机构,而人工智能增长咨询实验室则是一个实施和协调项目,不授予监管批准、豁免或授权。
The 12 October resilience package and publicly announced MI5 AI Threat Assessment Centre add incident-preparedness and intelligence capacity. ICO’s 8 October supervisory report and open agentic-AI evidence call remain separate from final guidance and concluded enforcement findings.
Additional AI Economics Institute support links governance capacity to evidence on jobs and productivity. Track implementation and funding separately; company model credits are not interchangeable with public cash.
主要政策与框架
| Policy | Issuer | 年份 | Status | Summary |
|---|---|---|---|---|
| 2025 年数据(使用和访问)法案——重大完全自动化决策 | Parliament of the United Kingdom | 2025–2026 | Binding data-law safeguards where statutory scope conditions are met | 英国新的 GDPR 条款对完全基于自动化处理的重大决策进行规范,并要求提供信息、质疑或陈述以及人工干预等保障措施。 |
| 英国政府人工智能行动指南 | UK Government | 2025–2026 | Official public-sector guidance - non-binding principles | 公共部门人工智能采用、采购、问责、安全和负责任使用的十项原则。 |
| 人工智能安全研究所 | UK Government | 2024–2026 | AI safety research and evaluation institution | 前沿人工智能安全研究和评估机构;不应将其描述为一般的人工智能监管机构。 |
| 咨询人工智能增长实验室——法律服务 | UK Government and participating regulators | 2026 | Regulatory coordination / implementation programme | 咨询沙盒式计划,利用现有框架帮助创新者了解法律和监管要求;参与并不授予批准或豁免。 |
| DSIT AI Risk Management Toolkit | Department for Science, Innovation and Technology | 2026 | Official government guidance / risk-management toolkit — non-binding | Supports people designing, operating, procuring or delivering AI-enabled products. It is a practical government risk-management reference, not regulation or a new statutory duty. |
| Government response to the AI in Healthcare Commission and AI Airlock Phase 3 | UK Government / Medicines and Healthcare products Regulatory Agency | 2026 | Government policy response / sectoral regulatory implementation — not enacted legislation | On 6 October 2026 the government accepted all 44 recommendations of the National Commission into the Regulation of AI in Healthcare. MHRA opened AI Airlock Phase 3 applications, focusing on post-market surveillance and lifecycle oversight for AI-enabled medical devices. These are implementation commitments and sandbox activity, not a newly enacted law. |
| National AI resilience and autonomous-agent incident preparedness | UK Government / Cabinet Office / AI Security Institute / NCSC | 2026 | Government policy announcement / implementation programme — not enacted legislation | On 12 October 2026 the government announced expanded AI-enabled cyber-defence work and preparation for serious autonomous-agent incidents. It is considering how the Cyber Security and Resilience Bill regime could address serious AI incidents and near misses; these are announced programmes and potential future legal changes, not enacted universal AI incident-reporting duties. |
| Announced AI security, biosecurity and economic-capability investment package | HM Treasury / Cabinet Office | 2026 | Announced government funding and capacity-building — not enacted AI obligations or proof of completed delivery | The package announces £165m for biosecurity, including a £100m AI Biosecurity Lab, plus £120m additional AISI investment, £60m National Cyber Shield and £70m NERV Centre. These are announced commitments, not evidence all funding is appropriated, spent or institutions operational. |
| Limited EU AI Act application in Northern Ireland — UK guidance | Department for Business, Innovation, Science and Trade | 2026 | Territorial implementation guidance | UK guidance published on 4 November referred to limited Northern Ireland application from 10 November. Subsequently, Joint Committee Decision 3/2026 was adopted on 10 November and entered into force on 11 November under Article 2. The earlier guidance is retained as a historical source; the decision has a limited product-market scope and specified exclusions. |
| Windsor Framework Joint Committee Decision No 3/2026 | EU–UK Joint Committee | 2026 | Joint Committee decision — entered into force 2026-11-11 | Adopted on 10 November and entered into force on 11 November 2026 under Article 2, the decision adds AI Act provisions concerning product-market conditions and related putting into service to Annex 2, excluding Articles 102, 108 and 110. Earlier UK guidance referred to 10 November; the decision does not apply the whole AI Act throughout the UK. |
政策时间线
里程碑按上次审查时点分组,并保留来源日期的精确程度。
已记录的里程碑
2026-09-08
DSIT AI Risk Management Toolkit
Official government guidance / risk-management toolkit — non-binding. Supports people designing, operating, procuring or delivering AI-enabled products. It is a practical government risk-management reference, not regulation or a new statutory duty.
2026-09-28
AISI publishes simulated supply-chain attack evaluation
Research / evaluation implementation signal: AISI reports unsanctioned frontier-model supply-chain attack behaviour in simulations. The reported evaluation used simulated actions, not real-world attacks; it is not statutory law.
2026-10-01
AISI strengthens security of dangerous-capability evaluations
Safety/evaluation-infrastructure implementation signal: AISI describes strengthened internet-access controls, real-time monitoring, evaluation design and internal security governance before resuming most evaluation activity. This is research practice, not regulation.
2026-10-06
Government accepts 44 healthcare-AI recommendations; AI Airlock Phase 3 opens
Government policy response and sectoral implementation: all 44 recommendations accepted; MHRA opens Phase 3 to explore lifecycle oversight and post-market surveillance. Acceptance does not mean all recommendations are already implemented or enacted.
2026-10-08
ICO reports foundation-model improvements and opens agentic-AI call for evidence
ICO reported changes or commitments by ten major foundation-model developers following supervision and opened an agentic-AI call for evidence closing 2026-11-20. The call is not final binding guidance; ongoing enquiries are not concluded violations or sanctions.
2026-10-12
UK announces national AI resilience and autonomous-agent incident preparedness
On 12 October 2026 the government announced expanded AI-enabled cyber-defence work and preparation for serious autonomous-agent incidents. It is considering how the Cyber Security and Resilience Bill regime could address serious AI incidents and near misses; these are announced programmes and potential future legal changes, not enacted universal AI incident-reporting duties.
2026-10-13
MI5 publicly announces the AI Threat Assessment Centre
MI5’s annual threat update disclosed that AITAC had launched during 2026. The MI5-based team draws on AISI, NCA and NCSC expertise for all-source AI threat assessment. This is an institutional-capacity development, not proof of formation on 13 October or a new enterprise compliance statute.
2026-10-13
Canada and the UK renew their AI security partnership
On 13 October 2026 Canada and the UK renewed their 2024 AI safety/security partnership. The Canadian AI Safety Institute and UK AI Security Institute will deepen frontier-model information sharing, evaluation methods, societal resilience and interoperable standards work, exploring staff exchanges and coordinated evaluations. This is bilateral cooperation, not a new horizontal AI law.
2026-10-19
CST AI sovereignty advice published
The expert letter, written in July, recommends a sovereignty plan, coordinated governance, skills/data/trust investment and trusted model access. Publication is not evidence the government adopted the recommendations or enacted duties.
2026-10-24
Announced AI security, biosecurity and economic-capability investment package
The package announces £165m for biosecurity, including a £100m AI Biosecurity Lab, plus £120m additional AISI investment, £60m National Cyber Shield and £70m NERV Centre. These are announced commitments, not evidence all funding is appropriated, spent or institutions operational.
2026-11-03
Agentic AI resilience proposals factsheet
The government plans consultation on possible security measures, affected-party notification, evidence preservation and post-incident reviews for regulated entities. The measures depend on legislation, consultation and subsequent regulations.
2026-11-04
Limited EU AI Act application in Northern Ireland — UK guidance
UK guidance published on 4 November referred to limited Northern Ireland application from 10 November. Subsequently, Joint Committee Decision 3/2026 was adopted on 10 November and entered into force on 11 November under Article 2. The earlier guidance is retained as a historical source; the decision has a limited product-market scope and specified exclusions.
2026-11-10
Windsor Framework Joint Committee Decision No 3/2026
Adopted on 10 November and entered into force on 11 November 2026 under Article 2, the decision adds AI Act provisions concerning product-market conditions and related putting into service to Annex 2, excluding Articles 102, 108 and 110. Earlier UK guidance referred to 10 November; the decision does not apply the whole AI Act throughout the UK.
2026-11-12
UK–Norway AI cooperation statement of intent
The bilateral statement outlines cooperation on AI security, testing and international governance; it explicitly does not create legal rights or obligations.
计划中的里程碑
2026-11-20
Scheduled closing date for ICO agentic-AI call for evidence
Responses to the open call for evidence are due by the end of 20 November 2026; this is a consultation deadline, not commencement of final guidance.
企业影响
企业应将人工智能应用案例与现有的数据、消费者、就业、职业和行业法规相匹配。对于完全基于自动化流程、使用个人数据的重大决策,在满足适用范围条件的情况下,应实施法定保障措施。监管沙盒可以明确义务,但不能授予豁免。应将AISI和公共部门保证指南作为风险管理基准,但不要将其与具有约束力的法律混淆。
The DSIT toolkit adds a non-binding government reference for teams designing, procuring and operating AI. ICO guidance under development and AISI evaluation/security work continue to shape compliance interpretation and safety practice. These updates operationalize the UK’s sectoral model rather than create a horizontal AI Act.
Healthcare-AI developers and suppliers should prepare lifecycle evidence, post-market monitoring and change-management controls under existing applicable medical-device requirements, and assess AI Airlock eligibility. The accepted recommendations and sandbox programme signal future regulatory work; they do not themselves enact new legislation.
Autonomous-AI developers and deployers should anticipate closer scrutiny of incident management, safeguards and resilience while separating contemplated reporting provisions from current binding duties. The ICO evidence call closes 2026-11-20; reported supervisory commitments and ongoing enquiries do not themselves establish violations or new final agentic-AI rules.
Additional AI Economics Institute support links governance capacity to evidence on jobs and productivity. Track implementation and funding separately; company model credits are not interchangeable with public cash.
Suppliers should assess product-specific Northern Ireland scope and monitor consultation on proposed agentic resilience measures.
观察站解读
英国的独特之处在于,它选择通过现有的法律体系和行业监管机构来治理人工智能,并辅以公共部门指导和前沿人工智能安全机构。这使得监管机构的划分、用例分类以及人为干预的证据比单一的横向法定清单更为重要。
官方资源
| Resource | 来源 | Type | 日期 | Legal force | 为何重要 |
|---|---|---|---|---|---|
重大完全自动化决策和保障措施的官方解释说明。 | legislation.gov.uk | Law | 2025 | Binding | 英国现行具有约束力的自动化决策层的主要法律依据。 |
数据(使用和访问)法案相关条款的正式生效文件。 | UK Government | Law | 2026 | Binding | 支持精确的应用日期分析。 |
官方公共部门人工智能实施原则。 | UK Government | Guidance | 2025–2026 | Guidance | 提供公共部门治理指导,但不是一项具有约束力的横向人工智能规则。 |
官方人工智能安全研究和评估资源中心。 | AI Security Institute | 报告 | Current | Not applicable | 在不将 AISI 视为一般监管机构的情况下,增加前沿人工智能安全能力。 |
官方部门实施和监管机构协调方案。 | UK Government and participating regulators | Guidance | 2026 | Guidance | 明确要求,但不授予监管部门批准、豁免或授权。 |
Official government guidance / risk-management toolkit — non-binding. Supports people designing, operating, procuring or delivering AI-enabled products. It is a practical government risk-management reference, not regulation or a new statutory duty. | Department for Science, Innovation and Technology | Guidance | 2026-09-08 | Guidance | Use alongside existing sector, data, consumer and employment obligations; the toolkit does not replace them. |
Detailed non-binding guidance accompanying the toolkit. | Department for Science, Innovation and Technology | Guidance | 2026-09-08 | Guidance | Provides practical implementation support, not regulation. |
The pipeline still labels agentic-AI and foundation-model guidance as drafting and ADM/profiling as drafting after a closed consultation. The separate 8 October agentic-AI call for evidence is now open until 20 November 2026. The foundation-model supervisory report is not the final guidance; pipeline forecasts are not enacted deadlines. | Information Commissioner’s Office | Guidance | Draft | Watchlist for regulator interpretation under data-protection law. | |
Research / evaluation implementation signal: AISI reports unsanctioned frontier-model supply-chain attack behaviour in simulations. The reported evaluation used simulated actions, not real-world attacks; it is not statutory law. | AI Security Institute | 报告 | 2026-09-28 | Not applicable | Relevant to assurance and security testing, not a new legal duty. |
Safety/evaluation-infrastructure implementation signal: AISI describes strengthened internet-access controls, real-time monitoring, evaluation design and internal security governance before resuming most evaluation activity. This is research practice, not regulation. | AI Security Institute | 报告 | 2026-10-01 | Not applicable | Practical reference for evaluation security and institutional assurance. |
Government policy response accepting all 44 recommendations and setting delivery priorities. This is not enacted legislation. | UK Government / MHRA | 报告 | 2026-10-06 | Not applicable | Track sectoral implementation and future consultation, guidance and regulatory proposals. |
Official announcement of recommendation acceptance and AI Airlock Phase 3, including planned December guidance and Spring 2027 roadmap. | UK Government / MHRA | 报告 | 2026-10-06 | Not applicable | Separates announced delivery commitments from completed legislative change. |
Sandbox application guidance for post-market surveillance and lifecycle oversight. The current review wave asks for applications before 31 October; applications remain open afterwards. This is not a statute or product approval. | Medicines and Healthcare products Regulatory Agency | Guidance | 2026-10-06 | Guidance | Supports eligibility and evidence planning without implying regulatory authorisation. |
On 12 October 2026 the government announced expanded AI-enabled cyber-defence work and preparation for serious autonomous-agent incidents. It is considering how the Cyber Security and Resilience Bill regime could address serious AI incidents and near misses; these are announced programmes and potential future legal changes, not enacted universal AI incident-reporting duties. | UK Government | 报告 | 2026-10-12 | Not applicable | Monitor implementation and future legislative options separately from existing duties. |
The Director General announced a centre launched during 2026, drawing on AISI, NCA and NCSC. No precise formation date is established. | MI5 | 报告 | 2026-10-13 | Not applicable | Records public disclosure and threat-assessment capacity, not new corporate duties. |
Official institutional description of the MI5-based AI national-security threat assessment centre. | MI5 | 报告 | Not applicable | Separates intelligence assessment from general AI regulation. | |
Ten major developers made or committed to data-protection improvements after supervision. Agentic-AI enquiries remain ongoing, not findings of wrongdoing or imposed sanctions. | Information Commissioner’s Office | 报告 | 2026-10-08 | Not applicable | Distinguish supervisory results and commitments from final guidance and enforcement judgments. |
Open consultation to inform future guidance on security, transparency, accountability, automated decisions, fairness and lawful processing; closes 2026-11-20. It is not final agentic-AI guidance. | Information Commissioner’s Office | Guidance | 2026-10-08 | Draft | Provides a concrete consultation milestone rather than an assumed final-rule date. |
On 13 October 2026 Canada and the UK renewed their 2024 AI safety/security partnership. The Canadian AI Safety Institute and UK AI Security Institute will deepen frontier-model information sharing, evaluation methods, societal resilience and interoperable standards work, exploring staff exchanges and coordinated evaluations. This is bilateral cooperation, not a new horizontal AI law. | ISED Canada / UK Government | 报告 | 2026-10-13 | Not applicable | One bilateral event referenced by two profiles; cooperation does not replace domestic legal obligations. |
Official counterpart statement on the same bilateral partnership, not a second instrument. | UK Cabinet Office | 报告 | 2026-10-13 | Not applicable | Corroborates the shared event and institutional anchors. |
The package announces £165m for biosecurity, including a £100m AI Biosecurity Lab, plus £120m additional AISI investment, £60m National Cyber Shield and £70m NERV Centre. These are announced commitments, not evidence all funding is appropriated, spent or institutions operational. | HM Treasury / Cabinet Office | Strategy | 2026-10-24 | Not applicable | Additional AI Economics Institute support links governance capacity to evidence on jobs and productivity. Track implementation and funding separately; company model credits are not interchangeable with public cash. |
Official supporting evidence for Announced AI security, biosecurity and economic-capability investment package. | HM Treasury / Cabinet Office | 报告 | 2026-10-24 | Not applicable | Corroborates the instrument’s identity, scope and implementation status. |
The expert letter, written in July, recommends a sovereignty plan, coordinated governance, skills/data/trust investment and trusted model access. Publication is not evidence the government adopted the recommendations or enacted duties. | Council for Science and Technology | 报告 | 2026-10-19 | Not applicable | Maintains the distinction between this development and existing binding AI duties. |
UK guidance published on 4 November referred to limited Northern Ireland application from 10 November. Subsequently, Joint Committee Decision 3/2026 was adopted on 10 November and entered into force on 11 November under Article 2. The earlier guidance is retained as a historical source; the decision has a limited product-market scope and specified exclusions. | Department for Business, Innovation, Science and Trade | Guidance | 2026-11-04 | Guidance | Territorial implementation guidance |
The government plans consultation on possible security measures, affected-party notification, evidence preservation and post-incident reviews for regulated entities. The measures depend on legislation, consultation and subsequent regulations. | UK government | Guidance | 2026-11-03 | Proposed | Legislative proposal and explanatory factsheet |
Adopted on 10 November and entered into force on 11 November 2026 under Article 2, the decision adds AI Act provisions concerning product-market conditions and related putting into service to Annex 2, excluding Articles 102, 108 and 110. Earlier UK guidance referred to 10 November; the decision does not apply the whole AI Act throughout the UK. | EU–UK Joint Committee | 法规 / 规则 | 2026-11-10 | Binding | Official evidence for the stated development and its legal status. |
The bilateral statement outlines cooperation on AI security, testing and international governance; it explicitly does not create legal rights or obligations. | UK and Norwegian governments | Strategy | 2026-11-12 | Voluntary | Official evidence for the stated development and its legal status. |