Document No. Guobanfa [2026] 29, the General Office of the State Council's Opinions on Adapting to AI Development and Building an Employment-Friendly Development Model, is dated October 16, 2026, and was released publicly in late October. The document sets out 18 tasks across six areas. Its core message is not to stop companies from using AI, but to set the pace for how jobs are handled: technology may move fast, but the impact on jobs has to be counted first, and projects with a large displacement effect cannot launch first and file the paperwork later.
The tone the document sets first
The opinions state the direction plainly: keep employment first, let AI's positive effects — creating jobs, empowering traditional roles, improving job quality — come through, while slowing the displacement effect of AI adoption. For a company, that means a project proposal cannot rest on efficiency and cost alone; it also needs a page on jobs: which roles are replaced, which are reshaped, which are newly created, and where the affected people go. The document is addressed to governments and departments at all levels across China, not to a single industry. Human resources and social security departments and development and reform departments coordinate implementation, local authorities adapt it to local conditions, and major matters are reported through proper channels. Companies should read it as an early signal of the local rules and review criteria to come, not as the finished implementation rules themselves.
Who feels this line first
Three kinds of cases feel the constraint first. The first is labour-intensive industries: the opinions require a steady, cautious grip on the scope, intensity and pace of AI adoption in these industries, minimising the shock to workers, and prioritising rollout in posts facing labour shortages or hazardous conditions. The second is projects in market-access sectors: if an AI application project may cause a large employment shock, an employment-risk assessment must run alongside its review, and the relevant departments must coordinate transition and redeployment support; projects in other sectors assess risk as appropriate, so the intensity of assessment is not uniform. The third is companies using AI in human resources management: the opinions require such uses to be regulated, with unlawful conduct addressed under the law, including employment discrimination, false recruitment and infringements of privacy and rest rights caused by misused algorithms. Workers get matching arrangements: a consultation and coordination mechanism for rights under human-machine collaboration, with companies urged to protect pay, rest and leave, and occupational safety and health; flexible workers and people in new forms of employment are encouraged to join urban employee basic pension insurance, and the occupational-injury protection pilot for new forms of employment is to be expanded.
Four things companies can do now
First, run a job-impact inventory of AI projects in use and under construction, sorting roles into replaced, reshaped and newly created, with labour-intensive links listed separately; scope, intensity and pace belong in the project file, not just a launch timetable. Second, for market-access projects, prepare the employment-risk assessment and transition support plan before review; assessment cannot be backfilled after approval. Third, before large-scale rollout, explore an unemployment buffer period: tell affected workers in advance and provide transition and redeployment support, so the buffer is genuinely used for moving to new roles rather than delaying notice. Fourth, keep recruitment and management algorithms auditable: screening criteria, ranking rules and the scope of data collection must be traceable, and collection touching privacy or rest time must pass a legality check first; once an algorithm is found to be misused, the dispute will turn not on how advanced the technology is, but on its real effect on workers' rights. The document also has a job-creation side: faster development of new occupations and standards, expanding demand for roles such as digital management specialists, AI engineering technicians and AI trainers; a dedicated AI training programme, including the Skills Illuminate the Future training campaign; support for micro, small and medium enterprises to pursue low-cost, lightweight intelligent upgrades and build small, fast, light and precise digital products; and, as prescribed, access to special loans for stabilising and expanding employment for eligible companies.
Three boundaries that are easy to misread
First, this is an opinion from the General Office of the State Council, not a single statute with uniform penalties. Wording such as encouraged and guided does not carry the same force as the requirement for assessment alongside review; it would be wrong to read every clause as an immediate fine, and equally wrong to read the opinions as a non-binding suggestion, because once review criteria and local implementation catch up, project timelines will genuinely be affected. Second, not every AI project needs the same intensity of assessment: large-shock market-access projects are assessed in parallel with review, others as appropriate, so assessment must be tiered rather than one-size-fits-all. Third, the American form of regulation is different: California's SB 947 limits on discipline and dismissal, for example, are hard constraints at state-law level, with their scope and penalties written into statute, so its conclusions cannot simply be transplanted onto this Chinese opinion. For that contrast, see the site's California SB 947: automated systems cannot alone decide discipline and dismissal. Risk also sits in the monitoring loop: the opinions require AI employment risk to be folded into the mechanism for preventing and resolving large-scale unemployment risk, with coordinated contingency plans and joint responses to labour-relations risk. In other words, if workforce changes at individual companies erupt together, they stop being purely internal matters and enter the regional and sectoral early-warning picture.
What to watch next is how local authorities and departments turn as appropriate into workable review criteria: which market-access projects count as a large employment shock, how long a buffer period counts as reasonable, and what level of transition support counts as sufficient. There is no uniform answer yet, so the practical move for companies now is to have the job list, the assessment material and the notice process ready, rather than starting from scratch when the criteria arrive.
