How to Avoid Uneven Benefits and Risks in the AI Era: Boston University Scholars on Protecting Jobs and People-Centered Governance

(The author of this article, Wang Yan, is a senior researcher at the Boston University Global Development Policy Center, and Wu Yijun is a research assistant at the Boston University Global Development Policy Center)

On September 29, 2026, Trump met with heads of technology companies at the White House, and six companies signed the "Joint Commitment on Frontier Responsibility," proposing a four-tier arrangement of internal controls, internal oversight, external auditing, and board oversight. According to David Sacks's recollection, it was Zuckerberg and Jensen Huang who proposed the principles of strengthening auditing and board oversight. However, this agreement mainly focuses on risks such as cybersecurity and biosecurity, and does not address employment and income distribution.

AI governance must both prevent technology from spiraling out of control and respond to the issues of benefit distribution and cost-bearing in the application of technology. The former concerns whether systems can operate safely; the latter concerns whether improvements in production efficiency can improve the lives of most people. This article argues that AI governance needs to incorporate both technical safety and worker protection into its scope.

America's AI boom is occurring against a backdrop of highly concentrated wealth and polarization. In the second quarter of 2026, the wealthiest 1% of households held 32.5% of the nation's household net worth, while the less wealthy 50% of households held only 2.3% of net worth. According to Reuters, actions opposing data center construction in July of this year involved 42 states and 142 protests. Technology investment is expanding rapidly and capital returns are rising, yet residents are worried about jobs, electricity bills, inflation, and environmental costs, reflecting the asymmetry between benefits and risks. U.S. media call it "K-shaped growth."

On the one hand, the risks brought by AI are increasingly apparent: risk incidents related to AI and HuggingFace, and to PHP vulnerabilities, have flooded the internet. The risks brought by AI also include infringement on personal property, data, and the right to make autonomous decisions. After AI agents obtain access permissions to accounts, files, and software, they can continuously execute multiple operations. If the scope of authorization is too broad, or if the system misunderstands the user's goals, it may take actions beyond the person's wishes without the person even knowing.

Malicious instructions from external hackers may also abuse permissions granted by users. A controlled test published by security researcher Jason Martin in 2024 showed that an attacker hid instructions in a PDF, inducing an AI with computer-operating capabilities to execute operations deleting system files in a test environment. Anthropic has also acknowledged that malicious instructions in web pages and documents may hijack agents and cause them to deviate from user requirements. Such tests reveal weak points in security mechanisms.

Macro-financial risks also deserve attention. Zhou Xiaochuan, former governor of the People's Bank of China, pointed out at the 2026 Bund Summit that AI productivity gains mainly come from the application layer, while also warning that a large amount of investment is flowing into data centers, and that the degree to which future models and applications will depend on such computing power remains uncertain. Therefore, changes in technology pathways and asset return risks amid the investment boom need to be included in policy assessment together with employment shocks, to avoid excessively shifting the losses of technological transformation onto ordinary households.

On the other hand, when companies use AI to improve efficiency and reduce labor, how can workers, the unemployed, and retirees participate in sharing the benefits? In particular, as AI rapidly replaces grassroots jobs, how can young people who are entering the labor market find their first job and accumulate experience to grow? And how can workers whose efficiency has been greatly improved by AI change careers or cope with unemployment? Protecting the employment, income, and development opportunities of these groups should become a core focus of AI governance.

Acemoglu, David Autor, and Simon Johnson, winners of the 2024 Nobel Prize in Economics, proposed that AI's impact on employment and income distribution depends on the direction of R&D and application. They advocate developing "AI that benefits workers," letting technology enhance people's skills and help workers take on more complex tasks. Achieving this goal requires public policy support and also requires companies to change application approaches aimed at replacing labor with AI and cutting costs.

Stanford University economist Susan Athey pointed out in the 2025 Kenneth Arrow Lecture that machine learning can help identify workers whose incomes suffer long-term damage after unemployment, thereby improving support for job transitions. The degree to which different workers bear the impact of unemployment varies greatly, and older workers and those with lower levels of education in manufacturing especially need attention. This indicates that training and employment services should be designed for specific groups and their effects continuously evaluated.

The changes at the hiring entry point also need sufficient attention. Even if companies do not carry out large-scale layoffs, they may reduce new hiring when business expands. Shrinkage of entry-level positions will compress young people's opportunities to gain experience. A single company reducing labor input may be operationally reasonable, but if a large number of companies do so at the same time, it may create macro employment pressure. Whether the productivity dividend can be transformed into broad income growth requires support from labor protection and distribution systems.

China should also be alert to structural divergence. In August of this year, the value added of high-tech manufacturing grew by 16.7% year on year, while total retail sales of consumer goods grew by only 0.4%. The ASEAN+3 Macroeconomic Research Office (AMRO), in its preliminary assessment of China on September 22, pointed out that household consumption remains weak and recommended that policy support gradually shift toward public services, social security, and targeted support for household income. If most people's incomes cannot keep up with production capacity, insufficient consumption will in turn constrain growth, and "K-shaped" divergence may further intensify.

I participated in the writing of the World Bank's 1997 book "Reforming China's Pension System," and have followed the development of the social security and medical security systems ever since. In 2000, I also took part in writing "The Quality of Growth," responsible for Chapter 3, which focused on equalizing opportunities in education and health. Whether economic growth can improve the lives of most people has always been my concern. In the AI era, supporting innovation and cultivating patient capital should be combined even more with a people-centered approach and the promotion of common prosperity.

Social security financing also needs to adapt to the changes in income structure brought about by technological application. AI may raise productivity and broaden the tax base. But if corporate profits grow faster than total wages, pension financing that relies on payroll contributions may come under pressure. Enterprises reap new gains in the technological transition, and they should also bear the corresponding responsibilities for labor protection and public financing.

To this end, policies can be improved in four areas. In the short term, cushion the impact through unemployment protection, support for job transfers, and corporate responsibility for worker placement; in the long term, improve the distribution of technological gains and intergenerational opportunities through public services, social security financing reform, and investment in child development.

For key decisions and operations involving financial, legal, aviation, medical, infrastructure, military and other matters that bear on life safety and major public interests, AI should be prohibited from autonomous execution without sufficient verification and without effective oversight and takeover mechanisms. General-purpose models that have not completed sufficient verification must not be connected to execution systems, nor should their outputs be converted into actionable instructions without independent human verification. Auxiliary information retrieval, data organization and simulation analysis should be carried out in isolated environments with clear permissions. Automated systems that have undergone specialized verification should also be subject to applicable industry review and continuous oversight.

Necessary third-party audits should be carried out, and qualified personnel should be ensured the capacity for independent judgment, able to review, take over at any time and bear responsibility. Clicking confirm on a screen alone cannot constitute effective review. Key positions should also retain manual training and emergency drills, to prevent long-term reliance on AI from weakening human judgment and takeover capacity.

For operations involving bank and investment accounts, authorization for information reading, trading recommendations and actual execution should be separated. High-impact operations such as buying and selling securities, transferring funds and deleting important data should explain the consequences item by item and obtain the explicit confirmation of the person concerned. Systems should by default grant only the permissions necessary to complete the task, and set transaction limits, file backups, complete logs and emergency stop mechanisms. Uses whose risks cannot be effectively controlled should be prohibited from access.

Enterprises should establish responsibility tracing and accountability mechanisms, making clear the responsibility of the CEO and directly responsible persons for AI deployment, permission configuration and continuous oversight. Damage caused by improper deployment and absent oversight must not be blamed on algorithms. At the same time, retain entry channels for young people and talent development at different levels, support in-service and job-transfer training with public investment, develop human-machine collaborative services such as elderly care and childcare, and combine technological application with job creation.

Second, let workers and communities share the gains of technological progress.

In my previous article, "How to Avoid a Race to the Bottom: Staying People-Centered," I proposed that minimum wages should be appropriately raised in light of productivity, prices and regional development levels, so that more of the fruits of growth flow to ordinary families. In the AI era, workers' income growth, job training and protection of rights and interests should further be incorporated into corporate transition arrangements, to prevent cost-cutting and efficiency gains from turning into a race to the bottom that suppresses wages and benefits. For enterprises planning or currently building AI data centers, community consultation mechanisms can be established to make specific arrangements on public facilities, subsidies for residents' costs, local recruitment and skills training, and to disclose implementation, so that local residents can share investment returns and reasonably share resource and environmental costs.

Third, expand social security and medical insurance coverage and improve corporate financing responsibilities.

Establish a multi-pillar protection system suited to different forms of employment, providing flexible contributions and cross-regional portability for flexible workers. Supplementary individual pension accounts can allocate diversified equity assets under risk constraints. Strengthen support for young people and unemployed workers, and expand unemployment and work injury insurance coverage.

For enterprises that replace labor with AI on a large scale, an "agent social security tax" can be explored, namely a social security adjustment tax or special contribution, used for unemployment benefits, training and pension protection. This is a policy idea that requires further argumentation; the tax should not simply be based on the number of agents, but should combine profits, value added and changes in total wages, be coordinated with existing taxes and fees, reward job transfers, placement and job creation, and at the same time avoid double taxation and discouraging beneficial innovation.

The protection system should also further eliminate obstacles caused by urban-rural household registration and regional differences, raise pension benefits for elderly people in difficulty in rural areas, and improve primary-level medical care and the convenience of seeking medical treatment outside one's home region. Beyond basic medical insurance, supplementary savings accounts voluntarily funded by individuals and dedicated to compliant medical spending can be explored, with personal income tax preferences for contributions up to a certain amount and appropriate fiscal subsidies for low-income families.

Fourth, increase redistribution and support marriage and childbearing, child development and equal opportunity.

China's tax system as a whole still has room to improve in progressivity and redistribution; income and wealth adjustment mechanisms need to be improved so that groups with stronger ability to pay bear the corresponding public responsibility, and redistribution also needs to be delivered through public spending. For example, consideration could be given to drawing on U.S. experience to set up an "education fund account" for eligible children with a one-time contribution, which provides a case and idea for long-term asset accumulation for children. China can, in light of its own circumstances and conditions, explore child development fund accounts, with the government providing basic contributions, allowing families to add voluntarily, and tilting toward families in difficulty.

At the same time, prudently advance real estate tax reform, study the institutional conditions for an inheritance tax, set reasonable protections for ordinary families, devote more public resources to social security, childcare and child development, reduce opportunity inequality caused by differences in family wealth, and promote social mobility.

In the AI era, it is all the more important to protect jobs and protect workers, and safeguard their employment and development opportunities. Measuring technological progress should look not only at computing power, investment and efficiency, but also at job quality, pension protection and development opportunities, and at whether the vast majority of people can benefit from it. In rapid transformation, workers must not be left alone to bear the risks of technological progress; enterprises should bear the transition risks and placement responsibilities brought by technological application. The government also cannot stand by; through clear reward and punishment mechanisms, it should support enterprises that create jobs, train employees and properly place transferred workers, and hold enterprises legally accountable if they evade labor protection obligations and shift transition costs onto workers, so that technological progress truly serves human development and common prosperity.

Only when every worker can share the dividends of technology can technological efficiency be translated into sustained consumer capacity and common prosperity.

Exclusively first published on Yicai's Yicai Account; this article represents only the author's views.