Beijing Puts a Jobs Test Inside AI Project Review

Key points

• 国办发〔2026〕29号, signed by the General Office of the State Council on 16 October 2026 and published on 23 October, requires that AI application projects in market-access sectors capable of causing a large employment shock undergo an employment-risk assessment at the same time as project review.
• The same document encourages enterprises rolling out AI at scale to explore setting up a “失业缓冲期制度” — an unemployment buffer period, with advance notice to affected workers and redeployment support — and folds AI-related job risk into China’s existing mechanism for preventing large-scale unemployment.
• Korea’s cabinet approved amendments to its AI Basic Act and Data Industry Act on 13 October 2026 that would oblige AI training data built in government-funded projects to be linked to and provided through the state AI Hub platform, with commencement six months after promulgation, which the Korean legal press puts at about April 2027.

China has made the labour consequences of an AI deployment part of the paperwork for approving it. On 16 October 2026 the General Office of the State Council signed 国办发〔2026〕29号, the 《关于适应人工智能技术发展助力构建就业友好型发展方式的意见》 — Opinions on adapting to the development of artificial intelligence technology and helping to build an employment-friendly development model — and the central government portal published it on 23 October, under the formula 经国务院同意, with the agreement of the State Council. Xinhua’s dispatch carried a Beijing dateline of the same day; reproductions on the Ministry of Ecology and Environment’s site and on the Liaoning provincial government’s are dated 24 October. The document runs to 18 tasks across six areas. One of them changes what a project file has to contain.

Task seven is the one to read. For AI application-scenario projects in market-access sectors that could cause a sizeable employment shock, an employment-risk assessment is to be carried out at the same time as project review — “在评审时同步开展就业风险评估” (in Chinese; translation by AI & GPP) — with the relevant departments to coordinate transition and redeployment support. A further sentence extends the exercise to projects in other sectors, assessed in the light of actual conditions. The passage bans nothing, sets no threshold and names no penalty, and it does not say that the answer determines whether a project proceeds. What it does is put a question about jobs inside a process a regulated deployment must already pass through, and require that it be asked while the decision is still open.

What the employment test covers, and where it bites

The scope is graded rather than narrow. The firm limb attaches to 市场准入领域 — market-access sectors, where a licence, permit or project review stands between a company and operation — and within those only to projects judged capable of a sizeable employment shock. Beyond them the document still calls for an assessment, qualified by 结合实际, in the light of actual conditions, which leaves trigger and rigour to the reviewer.

Two further features matter. The trigger is a judgement about scale, and no numeric threshold appears, which leaves the reviewing body to decide what counts as large. And the duty to act on a bad answer falls on departments, not the applicant: the text pairs the assessment with an obligation on the relevant authorities to organise transition and redeployment support.

An employment test inside project review is a procedural duty, not a prohibition. It changes what a reviewer has in front of them, not what they are empowered to refuse.

Task six encourages firms rolling out AI at scale to explore an unemployment buffer period — 失业缓冲期制度 — under which affected workers are told in advance and offered support in moving to other work. Task thirteen turns to the other end of the employment relationship: AI use in human-resources management to be brought under rules, and employment discrimination from the misuse of algorithms dealt with under law, alongside fake recruitment and encroachments on privacy and the right to rest. Tasks sixteen to eighteen build the monitoring layer: cross-departmental sharing of AI industry and employment data, a number of AI employment research centres — “布局一批人工智能就业研究中心” — on the existing resources of universities and research institutes, and the inclusion of AI-driven job risk in the standing mechanism for heading off large-scale unemployment.

The industrial policy is unapologetic. The document deepens the “人工智能＋” — “AI Plus” — initiative, encourages eastern regions to site the core AI industry while guiding central and western ones to take on data labelling and smart-terminal manufacturing as local conditions allow, and names the 国家人工智能产业投资基金 and the 国家创业投资引导基金 — the national AI industry investment fund and the national venture capital guidance fund — among the government funds to be used. Beijing is not slowing AI adoption; it is pricing the labour externality into the approval process while spending to accelerate the rest — the same two-handed posture AI & GPP traced when model weights became a controlled export a week earlier.

What an 意见 obliges, and what it does not

Legal character decides how much binds. An 意见 from the General Office of the State Council is a normative policy document addressed to provincial governments and State Council departments. It is not a statute, not an administrative regulation, and it creates no directly enforceable duty on a private company. Its force is administrative: the implementation section assigns no lead ministry to any task, asking only that human resources and social security and development and reform authorities at every level strengthen overall coordination — in practice the Ministry of Human Resources and Social Security (MOHRSS) and the National Development and Reform Commission.

What that means for a company is specific. The assessment will bite only when a reviewing body applies it, through its own procedures; the buffer period is encouraged, not mandated, with no sanction for ignoring it. No commencement date, transition period or reporting deadline appears anywhere in the text. As of 25 October no implementing measure from MOHRSS or the NDRC had been published, the ministry’s own site carried no page or press release on the document, and the policy-interpretation slot on its gov.cn record was empty. The step exists on paper; nothing yet says how it is to be taken. That is the familiar sequence in Chinese AI governance, where practice runs ahead of or behind the instruments that supposedly drive it — as AI & GPP found when Beijing policed AI-faked soldiers faster than it wrote standards.

Seoul spent the same fortnight regulating the inputs

Three days before Beijing signed, Seoul moved on a different part of the problem. On 13 October 2026 the Ministry of Science and ICT announced that amendments to the AI Basic Act — the Framework Act on the Development of Artificial Intelligence and the Establishment of a Foundation for Trustworthiness, passed by the National Assembly on 26 December 2024 and in force since January 2026, 뉴시스 reported and 디지털데일리 confirmed on the eve of commencement — and to the Data Industry Act had passed the cabinet meeting, 국무회의, in a release reproduced by the Korea Development Institute’s Economic Information and Education Centre. The AI Basic Act amendment would oblige AI training data built or produced in government projects to be linked to and provided through the state AI Hub platform, with support and data-quality duties attached. The Data Industry Act amendment gives a statutory footing to the One-Window data-location service, with public institutions required to cooperate unless there is a special reason not to, and private holders merely asked.

A client note from the law firm Shin & Kim of 16 October, carried by 법률신문 three days later, records that the grounds on which an operator may withhold data — trade secrets, infringement of personal information or copyright, national security — are to be fixed by Presidential Decree (대통령령), and puts commencement six months after promulgation, “2027년 4월 예정”, April 2027. 파이낸셜뉴스 and 바이라인네트워크 report both points independently. 엄열, the ministry’s 인공지능정책실장 — head of its AI policy office — was quoted in the Shin & Kim note and by 바이라인네트워크 as tying AI competitiveness to both high-quality training data and a public able to find the data it wants (in Korean; translation by AI & GPP).

Precision is owed here too, and the record is incomplete. Every source opened says the same thing — the amendments 국무회의를 통과, passed the cabinet meeting — and none mentions the National Assembly. The phrase is ambiguous in Korean practice: the cabinet sits in the sequence both before a government bill goes to the Assembly and again when it resolves on promulgating a bill the Assembly has passed. 뉴시스 reported that the AI Basic Act itself took the second route, and the six-month commencement clock uniformly reported for these amendments, about April 2027, points the same way, since it implies a promulgation date already in view. The desk could not confirm which. Either way nothing binds a data holder today, and the ministry will revise the subordinate regulations in time for commencement.

Two capitals, two points of intervention

Side by side, the divergence is structural, not one of severity. Beijing regulates the labour consequence of adoption; Seoul regulates the input that makes adoption possible. Neither does what Brussels does: regulate the system itself by risk class.

The practical reading is that the Chinese duty is vaguer but live, the Korean one sharper but not yet in force. A company preparing a market-access application in China should assume a reviewer may ask for headcount effects and a redeployment plan, and that producing one late is worse than early. A company holding data built with Korean public money should be mapping which datasets would fall inside the AI Hub obligation and which exemption it would claim, since the grounds will be settled in a Presidential Decree rather than the statute.

Three things are worth watching. Whether MOHRSS or the NDRC supplies a threshold for “large employment shock” will decide whether task seven is a real constraint or a form of words. When the Korean amendments are promulgated sets when the six-month clock runs out, and so when the AI Hub duty becomes an obligation rather than a plan. And whether any other government copies the Chinese design will show whether labour-market exposure, which the World Bank put at 7 per cent of South Asian employment, is about to become a licensing question rather than a research finding.

Questions this answers

Does China now require an employment assessment before an AI project can be approved?
In regulated sectors, yes in policy terms. 国办发〔2026〕29号 directs that AI application projects in market-access sectors which could cause a large employment shock have an employment-risk assessment run at the same time as project review, with departments to coordinate redeployment support; a softer version is called for in other sectors as conditions warrant. It is an Opinions document rather than a statute, so it binds departments rather than companies directly; no threshold, deadline or penalty is stated, and the text does not say that the assessment’s outcome decides whether a project proceeds.

Is the Korean AI training-data obligation in force?
No. The Ministry of Science and ICT announced on 13 October 2026 that amendments to the AI Basic Act and the Data Industry Act had cleared the cabinet meeting; no source opened for this piece records where that step sits relative to the National Assembly. On every account commencement falls six months after promulgation, around April 2027, so nothing binds a data holder today.

What should a company deploying AI at scale in China do now?
Prepare the employment case before filing, not after. The document also asks enterprises rolling out AI at scale to explore a “失业缓冲期制度”, an unemployment buffer period with advance notice to affected workers and support in moving to other roles; that is encouraged rather than required, but it is the standard against which a reviewer coordinated by MOHRSS and the NDRC is likely to read a deployment plan.

Sources

1. 国务院办公厅关于适应人工智能技术发展助力构建就业友好型发展方式的意见（国办发〔2026〕29号）— 中国政府网, 成文日期 16 October 2026, 发布日期 23 October 2026 (in Chinese; translation by AI & GPP)
2. 国务院办公厅关于适应人工智能技术发展助力构建就业友好型发展方式的意见 — 生态环境部, reproducing 中国政府网, 24 October 2026 (in Chinese; translation by AI & GPP)
3. 国务院办公厅印发《关于适应人工智能技术发展助力构建就业友好型发展方式的意见》— 辽宁省人民政府, reproducing 新华社, 24 October 2026 (Xinhua dispatch datelined Beijing, 23 October 2026; in Chinese; translation by AI & GPP)
4. 국가 AI 데이터의 체계적 관리·활용을 위한 「AI기본법」, 「데이터산업법」 개정안 국무회의 통과 — 과학기술정보통신부, reproduced by KDI 경제정보센터, 13 October 2026 (in Korean; translation by AI & GPP)
5. 국가 AI 데이터의 체계적 관리·활용을 위한 「AI기본법」·「데이터산업법」 개정안 국무회의 통과 — 법률신문, 19 October 2026 (carries the Shin & Kim note at source 7; in Korean; translation by AI & GPP)
6. ‘AI 데이터의 확보·제공 확대’ AI기본법·데이터산업법 개정안 통과 — 파이낸셜뉴스, 13 October 2026 (in Korean; translation by AI & GPP)
7. 국가 AI 데이터의 체계적 관리·활용을 위한 「AI기본법」·「데이터산업법」 개정안 국무회의 통과 — 법무법인(유) 세종 (Shin & Kim) 뉴스레터, 16 October 2026 (in Korean; translation by AI & GPP)
8. 과기정통부, AI 학습데이터 공유 의무화 추진 — 바이라인네트워크, 13 October 2026 (in Korean; translation by AI & GPP)
9. ‘고영향 인공지능 규제’ AI기본법 국회 통과…2026년 1월 시행 — 뉴시스, 26 December 2024 (in Korean; translation by AI & GPP)

Claim ledger (piece 13)