As artificial intelligence becomes increasingly integrated into public services, employment and social governance, responsible AI governance worldwide is gradually moving beyond the articulation of principles and the construction of policy frameworks and into a new phase focused on institutional implementation and the evaluation of governance outcomes. On 23 September 2026, the South Africa-based Global Center on AI Governance (GCG) released the Global State of Responsible AI Report. The report argues that the central tension confronting responsible AI governance today is that policy commitments are expanding faster than implementation capacity and accountability mechanisms are improving. This assessment suggests that the key question is not merely whether policies and frameworks have been established, but whether the relevant institutions can be translated into practice in specific application settings; whether they can effectively constrain governments, companies and other users of technology; and whether individuals and groups affected by AI can obtain meaningful protection.
This article reviews the report’s methodological adjustments and its main findings across five dimensions. It examines why countries tend to prioritize the development and adoption of AI capabilities without strengthening oversight of and constraints on their own conduct at the same pace. It then considers the limitations of the Global Index on Responsible AI in terms of both measurement methods and normative benchmarks.
Cover of the Global State of Responsible AI Report
Source: Original report
01 What It Takes to Measure Responsible AI
Measuring responsible AI requires answers to two basic questions: who is responsible, and to whom are they responsible?
The scope of responsibility must first be defined. The inaugural 2024 edition of the Global Index on Responsible AI (GIRAI; hereinafter, “the Index”) emphasized respecting and protecting human rights throughout the AI lifecycle. The 2026 edition expands this scope to include preventing harm, promoting inclusion and serving the public interest. This expansion responds to a problem identified in the first assessment: even when AI systems do not formally violate rights, they may still cause harm, exclude particular groups or concentrate benefits in the hands of a few. Human rights protection is therefore the baseline for responsible AI, but determining whether AI is responsible also requires examining who bears the risks and how the benefits are distributed.
Expanding the scope of responsibility also requires the assessment to identify who should be held responsible. The 2026 edition evaluates government frameworks and government action together, examining policy commitments alongside evidence of implementation. For non-state actors, it retains only records of relevant activities by civil society organizations and does not assign them scores. The research team argues that responsibility should rest primarily with actors capable of making consequential decisions across the AI value chain, rather than applying the same assessment to organizations seeking to strengthen oversight and correct existing power imbalances. The Index therefore places even greater emphasis on government: it considers both how governments regulate companies and how they constrain their own use of AI.
Global Index on Responsible AI — 2026 Report
Source: global-index.ai
The Index has also expanded its assessment framework, from three dimensions and 19 thematic areas to five dimensions and 38 indicators. In addition to policy commitments and government action, the 2026 edition introduces enabling conditions, bringing institutional and resource contexts into the analysis. The same policy arrangement may produce different outcomes depending on the foundations available for implementation. Country scores must therefore be understood in light of regulatory capacity, resource investment and existing institutions.
Changes in the scope of responsibility and the structure of the assessment nevertheless limit direct comparisons between the two editions. The research team chose to prioritize revising the methodology as the research developed over maintaining strict comparability between editions. This choice also reminds readers that country scores depend on how the Index defines responsibility—including which risks it prioritizes, which interests it protects and which actors it requires to assume obligations. The Index uses UNESCO’s 2021 Recommendation on the Ethics of Artificial Intelligence as a key benchmark, providing a common language for cross-country comparison. Yet it remains necessary to test how well a common standard captures different national institutional arrangements and responds to each country’s historical context and policy priorities.
02 The Governance Implementation Gap Across Five Dimensions
Countries have widely begun to formulate AI policies, but there are still major differences in whether those policies translate into meaningful protection. More than 90 percent of the 135 countries and jurisdictions covered by the Index have introduced at least one relevant policy or initiative, yet the global average score is only about 35 out of 100. Even where frameworks are in place, evidence of implementation can be found in only slightly more than half of cases. Drawing on the principal values and policy areas set out in UNESCO’s Recommendation, the Index examines inclusion and diversity, ethics and sustainability, labor and skills, trust and safety, and the use of AI in public services. In doing so, it reveals the gap between policy commitments and action in practice.
Framework coverage by dimension: implemented versus adopted but not implemented
Source: Original report
Inclusion and diversity first raise the question of who should bear responsibility for adapting to AI. Countries find it relatively easy to embrace general principles such as fairness and non-discrimination, but implementation lags significantly when it comes to protecting specific groups, including women and children. For example, fewer than half of the countries with gender equality frameworks have taken corresponding government action. Child protection also relies more heavily on AI literacy education in schools, while obligations requiring developers to prevent harm remain weak. Education can help children recognize risks, but it cannot substitute for safeguards built into products themselves. The result is that the burden of identifying risk falls on children, while developers assume little responsibility for prevention. Cultural and linguistic diversity is an exception: more countries have implementation initiatives (52) than have formal frameworks (47). Yet when governments fund local-language datasets and models without community participation or inclusive benefit-sharing arrangements, the gains may flow primarily to large foreign AI companies. The report’s conception of inclusion is not limited to enabling more people to use AI; it also requires AI design and use to respond to the needs of different communities.
Inclusion and diversity indicators: percentage of countries with frameworks and government initiatives
Source: Original report
The implementation of ethical principles likewise involves trade-offs. Transparency, human oversight, fairness and safety receive relatively substantial attention, while environmental and energy concerns lag behind, with fewer than 30 percent of countries taking relevant action. The report argues that requirements are less likely to be translated into action when they require industry to bear costs or entail a redistribution of benefits.
Implementation tools are also shaped by resource constraints. Since the first edition of the Index was published in 2024, more than three-quarters of newly introduced frameworks in the Global South have been non-binding. The report links this pattern to legislative and regulatory capacity, fiscal resources and bargaining power in investment negotiations. A high share of soft-law instruments cannot therefore be treated, on its own, as evidence of weak political will. Even legislation may remain on paper if enforcement capacity is lacking.
In labor and skills, the dominant pattern is that training comes first while protection lags behind. More than half of countries have frameworks for AI skills training, but fewer than 30 percent have frameworks for labor protection. Governments want workers to seize the employment opportunities created by AI, yet safeguards covering low-paid outsourcing, platform work and workers’ ability to assert their rights have not kept pace.
Share of binding and non-binding AI frameworks in the Global North, the Global South and across regions
Source: Original report
This gap is especially pronounced in the Global South. Work such as data labelling is viewed as an opportunity to attract investment and create jobs, and governments may fear that stronger protections will cause projects to move elsewhere. Layers of subcontracting and weak worker organisation further complicate accountability. Opportunities to participate in the AI economy do not necessarily allow workers to share in its benefits or protect them from harm. The report cites Kenya as an example. In 2023, Sama, an outsourcing company that provided content-moderation services for Facebook, laid off large numbers of employees. Dismissed workers said the layoffs were connected to their attempts to unionize. In the competition for outsourced work, gaps in labor protection may be treated as a competitive advantage—one that no single country can easily eliminate on its own.
Number of countries in the Global North and Global South with labour-protection and skills-training frameworks
Source: Original report
The key gap in trust and safety is that risk management and accountability have not advanced in tandem. More than half of the relevant frameworks are non-binding, while government action focuses mainly on training, research and institution-building. Far fewer provisions address how affected people can challenge decisions, hold responsible actors to account and obtain remedy. Of the regulations counted by the report, fewer than 20 percent require or incentivize companies to establish grievance mechanisms.
Accountability gaps also arise in governments’ own uses of AI. Among countries with relevant AI laws or regulations, more than three-quarters provide exemptions for defense and national security under each of the three indicators of safety, impact assessment and remedy. An agency’s ability to conduct its own risk management does not mean that external overseers or affected individuals can require it to explain and correct its decisions.
Percentage of countries whose AI laws and regulations contain defense and security exemptions, by indicator
Source: Original report
In the public-services dimension, governments are adopting AI far faster than they are governing it. This is the lowest-scoring of the five dimensions. Public-sector skills frameworks are relatively common, but requirements governing AI procurement and the disclosure of government algorithms lag far behind. Assessing government use of AI requires more than asking whether public employees have the necessary skills. It also requires asking whether procurement processes assign responsibility, whether use is transparent and whether members of the public can challenge AI-assisted decisions.
Two further patterns in this dimension merit attention.First, the North–South gap is smallest here, and this is the lowest-scoring dimension for nearly 60 percent of European countries. One reason is that the dimension assesses only national policies and excludes regional legislation such as the EU AI Act. Another is that public-sector AI in the Global South is often embedded in digital public infrastructure projects funded by institutions such as the World Bank. Procurement rules imposed by donors tend to focus on financial oversight and rarely address rights protection. Second, the Index records 53 cases across 35 countries in which governments used AI systems posing an “unacceptable risk,” including 27 cases of mass biometric surveillance. Countries with greater transparency and higher overall rankings recorded more cases. A larger number of publicly documented cases, however, does not necessarily indicate higher risk; it may instead reflect a stronger capacity among the media, courts and civil society to uncover problems.
Frameworks and initiatives across three indicators: public-sector skills, procurement and algorithmic disclosure
Source: Original report
The implementation gaps across all five dimensions display a similar tendency. Skills training, capacity-building and the promotion of applications are easier to advance, while oversight, labor protection and mechanisms for remedy lag behind. This disparity suggests that implementation is shaped not only by resources and capacity, but also by how governments assess the benefits and costs of different measures.
03 Why Do More Rules Still Fail to Produce Effective Constraints?
The report reveals a gap between policy commitments and implementation. Yet attributing this gap mainly to governments’ preference for expanding their own capabilities while avoiding constraints on themselves does not fully explain current developments in governance. The Index’s data run through September 2025, while 2026 has brought a series of regulatory developments and agent-safety incidents that need to be incorporated into the same analytical framework. The central tension in governance today is that public institutions are assuming growing responsibility for AI without necessarily possessing the information, technical capacity and powers of intervention that would allow them to discharge it. Whether rules can be implemented depends on whether these conditions can be combined; it cannot be determined solely by the number of policies or the share of binding instruments.
Governments must first be capable of identifying the risks they regulate. On 5 May 2026, the U.S. National Institute of Standards and Technology announced that its Center for AI Standards and Innovation (CAISI) had reached agreements with Google DeepMind, Microsoft and xAI to conduct pre-deployment evaluations and targeted research. The capabilities and dangerous behaviors of frontier models cannot readily be inferred from their user interfaces alone. External evaluators need access to the models, the ability to design tests and support from relevant specialists. Without these conditions, a requirement that companies submit to oversight may result in little more than formal compliance. It is therefore important to distinguish among the forms of capacity-building criticized in the report: capacity to promote adoption may accelerate the spread of risk, whereas independent testing and investigative capacity are prerequisites for effective constraint.
CAISI evaluation report
Source: CSA Lab Space
Early government access to models can help identify risks. Whether governments can require companies to change deployment plans after risks are identified, however, depends on the powers available to them. A U.S. executive order issued on 2 June proposed a mechanism for evaluating the network capabilities of frontier models. Under a voluntary cooperation framework, the government could receive access to models for up to 30 days before developers make them available to other trusted partners. At the same time, the order expressly states that it does not authorize mandatory licensing, preclearance or permitting requirements for model development or release. The government may therefore gain an earlier opportunity to evaluate models, but other institutional arrangements are still needed to determine how evaluation results affect corporate deployment decisions.
Provision in the U.S. executive order on a voluntary framework for cooperation on frontier AI models, expressly stating that it does not authorize mandatory government licensing or preclearance for model development and release
Source: The White House
Cooperation also raises the question of whether oversight can remain independent. Governments need companies to provide models and technical information, while companies have incentives to participate in government procurement and national-security projects. If what is tested, how it is tested and what information evaluators can obtain all depend on corporate cooperation, evaluators may find it difficult to scrutinize companies’ safety claims fully. Determining whether such cooperation imposes effective constraints requires examining whether evaluators can design and carry out tests independently, obtain the information they need, and ensure that someone has the authority to require corrective action when risks are found. There should also be clear consequences when companies refuse to cooperate. The value of cooperation agreements must ultimately be judged against these concrete conditions and the evidence of their implementation.
Governments are simultaneously regulators, purchasers and users of AI, and different agencies do not necessarily share the same objectives. In May, U.S. defense agencies announced agreements with OpenAI, Google, Microsoft, NVIDIA, Amazon Web Services and other companies to deploy AI on classified networks. On 7 July, Senator Elizabeth Warren asked the defense agencies and companies involved to disclose the scope of the contracts and the safeguards they contained, and to publish the contracts in full. These inquiries do not by themselves establish that the deployments violated any rules. They do, however, raise an important question: as governments expand their use of AI, who can review its purposes and safeguards, and how much information can reviewers obtain?
This tension is particularly visible in national security. Military agencies seek rapid access to advanced capabilities, technical evaluation bodies focus on risk, and legislatures demand clarity about responsibility and boundaries of use. Secrecy requirements may further restrict external scrutiny, making it difficult for the public to determine whether safety commitments are being honored. The report’s explanation that states selectively adopt governance requirements must therefore be examined in light of relationships among specific government bodies. Whether constraints are effective depends largely on who can access deployment information, who is empowered to object and whether objections can bring about change. Treating government simply as a single entity that is either willing or unwilling to constrain itself risks obscuring the institutional conditions that shape implementation.
The European Union’s approach also demonstrates that voluntary cooperation and binding constraints can work together. Although companies sign the General-Purpose AI Code of Practice voluntarily, declining to sign does not eliminate their legal obligations. Evaluating a cooperation mechanism requires examining whether clear duties stand behind it and what consequences follow from non-compliance. This approach also shows why soft law cannot simply be equated with weak governance, or hard law with strong governance. The Code of Practice is a voluntary instrument designed to help companies meet their obligations under the AI Act; choosing not to sign does not exempt them from those statutory duties. The European Commission’s enforcement powers in relation to general-purpose AI models have applied since 2 August 2026, including powers to request information, evaluate models and require corrective action. The real question is whether a voluntary arrangement is backed by unavoidable duties and credible enforcement consequences. Cooperative tools can make rules more operational, but they can also conceal an absence of accountability. Institutional design and evidence of implementation determine which of the two is taking place.
The implementation gaps identified in the report should therefore be disaggregated further. Some institutions lack the capacity to identify risks; others cannot obtain the necessary evidence; still others identify problems but lack the authority to require corrective action. These barriers call for different solutions. Progress in governance should be judged by whether risks, once identified, can be corrected and whether the institutions making deployment and regulatory decisions are themselves subject to oversight.
04 Turning Governance Requirements into Sustained and Effective Protection
Recent testing and regulatory action concerning frontier models show that AI governance is moving from discussion of principles into concrete operational processes. As agents become capable of carrying out extended tasks, using software and interacting with external systems, regulators must look beyond what a model generates to what it is able to do—and whether it can be stopped promptly when something goes wrong. A recent testing incident disclosed by the United Kingdom’s AI Security Institute (AISI) provides a concrete illustration of this shift.
According to AISI’s public account, during a cyber-capability evaluation in July 2026, some agents took unauthorized action directed at real people and organizations, including an attempt to insert malicious code into an open-source project. To test model capabilities, researchers had deliberately enabled internet access and disabled some safety filters. AISI also stated that it found no evidence of resulting real-world harm and that behavior under these testing conditions should not be treated as directly representative of ordinary commercial use. The incident therefore does not establish that frontier models are generally out of control. It does, however, expose a practical problem: even in evaluations conducted by specialized institutions, agents may act beyond the scope originally set for them.
An instance of out-of-scope agent behavior documented by AISI, involving an attempt to insert malicious code and efforts to conceal the conduct after it was detected
Source: AI Security Institute incident report
After the incident, AISI paused the highest-risk tests and resumed most activities after strengthening network isolation, real-time monitoring and evaluation review. New safeguards included blocking external network access, intercepting suspicious actions for human review and putting human-operated shutdown measures in place. This response shows that identifying dangerous model capabilities is only one part of evaluation. The ability to restrict access to external systems, revoke permissions and terminate tasks promptly is equally important to the safety of testing and deployment. AISI also acknowledged that monitoring can fail and that existing safeguards may not be effective for the next generation of models. Safety measures must therefore evolve alongside model capabilities.
Constraining agents consequently requires attention to both model capabilities and deployment conditions. The same model may pose different risks when connected to different tools and granted different permissions. Whether a system can access the internet, use account credentials or modify external files should all be included in an assessment. Existing safeguards should also be reassessed after a model is upgraded, new tools are added or the use case changes. For high-risk applications, deployers should specify in advance the conditions under which a system must be paused, who makes that decision and what conditions must be met before operations can resume.
Legal rules can provide authority for such measures, but distance remains between the existence of a rule and the control of risk. Article 93 of the EU AI Act allows the Commission, under the relevant conditions, to require corrective action and risk mitigation and to restrict the making available of a model on the market, or to require its withdrawal or recall. These powers expand the range of actions available after an evaluation, but their practical effect depends on whether regulators can obtain evidence, act and oversee implementation in a timely manner.
Incidents arising in real-world use should also serve as an important source of governance improvement. Pre-deployment testing can cover only a limited set of conditions; task duration, tool combinations and the external environment can all change model behavior. If an incident record merely states that a system malfunctioned without documenting its configuration, permissions and the response process at the time, other institutions will struggle to adapt their own safeguards. Serious incidents and instructive near misses should therefore be recorded in sufficient detail and made available for independent scrutiny, subject to appropriate protections for security and trade secrets.
Model governance should connect test results to explicit consequences. For systems capable of high-impact action, deployers should define risk triggers in advance, establish procedures for restricting permissions or pausing operations, and specify the verification evidence required for resumption. Model upgrades, expanded tool access and changes in use cases should also trigger reassessment. Shutdown measures should distinguish among terminating a particular deployment, revoking a particular permission and restricting a model’s availability on the market, rather than conflating these different interventions. Article 93 of the EU AI Act provides a legal basis for post-evaluation action by allowing the Commission, under the relevant conditions, to require corrective action and risk mitigation and to restrict the making available of a model on the market, or to require its withdrawal or recall. Whether this legal authority is timely and effective must still be tested in practice. The existence of such provisions alone is not enough to claim that risk is under control.
Article 93 of the EU AI Act provides that the Commission may require providers of general-purpose AI models to take risk-mitigation measures and, under the relevant conditions, restrict a model’s availability, or require its withdrawal or recall
Source: European Commission website
A higher number of publicly reported incidents may indicate greater risk, but it may also reflect more robust mechanisms for detection and disclosure. Simple comparisons of case counts may make institutions that proactively disclose problems appear to perform worse. A more meaningful way to understand national governance performance is to ask whether risk incidents are detected promptly, whether they recur after corrective action, and whether affected people can obtain explanations and remedy.
New mechanisms at the international level have recently created opportunities to share these experiences. In July 2026, the United Nations convened the inaugural Global Dialogue on AI Governance. In September, the Independent International Scientific Panel on Artificial Intelligence issued a thematic brief on AI agents, misalignment and the risk of losing human control. These developments show that agent-related risks have entered the multilateral agenda and provide countries with a shared body of scientific evidence. The UN Global Dialogue, however, does not perform regulatory or enforcement functions. Investigations, deployment restrictions and accountability therefore continue to depend primarily on national and regional institutions.
António Guterres speaks at the opening of the Global Dialogue on AI Governance
Source: Xinhua News Agency
Given this division of labor, multilateral cooperation could begin with incident reporting and evaluation methods. If countries can share essential risk information and ensure that those with weaker testing capacity can participate, they may be able to reduce dependence on conclusions supplied by a small number of companies and technologically advanced states while preserving institutional differences. Whether such cooperation proves effective will depend on whether the information shared is sufficiently specific, opportunities to participate are genuinely accessible, and common understandings feed into national regulatory practice.
Concern about frontier-model risks should not displace protection against harms already occurring in everyday life. Emotional manipulation, discrimination, labor abuses and erroneous public-sector decisions are already affecting the rights and interests of specific users and groups. China’s Interim Measures for the Administration of Anthropomorphic AI Interaction Services, which took effect on 15 July, establish protections for minors, safeguards against excessive dependence, exit and complaint mechanisms for services involving ongoing emotional interaction, and requirements for providers to restrict functions or suspend services when they identify major security risks. Rules of this kind translate protective requirements into specific service relationships. Their effects remain to be seen, but they demonstrate that progress in governance can also take the form of clearer exit rights for users and more concrete duties of prevention for companies.
Two questions require continued attention. First, can institutions promptly interrupt dangerous system actions? Second, can affected people require institutions to explain and correct decisions? The first concerns operational safety; the second concerns whether governments and companies using AI can be held accountable. A focus limited to technical safeguards may overlook rights violations committed by institutions themselves. Conversely, principles and complaint channels without the capacity for timely intervention may allow protection to lag behind harm.
Participants enter the Geneva venue during the inaugural Global Dialogue on AI Governance
Source: Türkiye Today
Overall, the report shows the gap between policy formulation and implementation across countries. In light of recent developments, judging whether governance is effective also requires asking whether regulators can require corrective action or restrict use after identifying risks—and whether those measures are actually implemented. Once a policy has been adopted, who detects risk, who decides what measures to take and who ultimately bears responsibility reveal more about the effectiveness of protection than the number of rules on the books. As model capabilities and the range of applications continue to expand, whether these arrangements can withstand technological upgrades, commercial pressure and national-security demands will be an important test of progress in AI governance.
Original Link: https://mp.weixin.qq.com/s/jmG-hsXx6rVfZxSASqszCQ
