A safe, manageable future for artificial intelligence may depend on cooperation between the United States and China – even as each increasingly casts the other as the chief source of risk, PBS reports. Washington and Beijing together control the talent, data, computing power and commercial ecosystems that will shape AI’s capabilities; yet growing strategic rivalry, export controls and competing visions for governance are driving a split that could undermine global efforts to manage the technology’s dangers.
The standoff is not abstract: policymakers worry about military applications, misuse by nonstate actors, and the rapid deployment of powerful models without agreed safety standards. If the two superpowers fail to find a way to cooperate, experts warn, the world risks fragmented rules, an accelerated arms race in AI-enabled systems, and weaker mechanisms for preventing catastrophic accidents.
This article examines how mutual suspicion is reshaping AI diplomacy, the concrete barriers to U.S.-China collaboration, and the narrow pathways that could preserve a shared framework for global AI safety.
Mutual Mistrust Between the United States and China Is Undermining Global AI Safety, Prioritize Transparency for Frontier Models and Targeted Export Control Reforms
Mutual mistrust between the United States and China is increasingly shaping – and constraining – the international architecture meant to keep advanced artificial intelligence safe. Officials and industry leaders in both capitals have adopted secrecy and competitive postures that make coordinated, pre-deployment safety checks harder to arrange, leaving global researchers and regulators working at cross-purposes. Experts warn that without a baseline of shared transparency on high-risk, frontier models – including pre-release safety evaluations, red-team findings, and machine-readable model cards – the world will face fragmented safety standards and higher chances of surprise failures. Journalism and policy reviews now point to a pragmatic middle path: multilateral disclosure of safety test results and independent audits carried out under agreed confidentiality rules to preserve legitimate national-security concerns while enabling collective risk assessment.
- Pre-release safety summaries: standard model cards and impact statements for frontier systems.
- Independent audits: accredited third-party red teams with protected reporting channels.
- Shared benchmarks: open, international test suites for robustness and misuse risk.
- Time-limited data sharing: secure windows for cross-border safety research.
Targeted export control reforms are equally urgent: sweeping bans on technology flows can push research into opaque channels and undermine the very cooperation needed to detect and mitigate systemic AI risks. Policymakers should pivot from blanket restrictions to calibrated controls that focus on choke-point hardware, verifiable licensing for sensitive tools, and carve-outs for bona fide safety research. This approach would preserve options for international verification and timely incident response while still limiting malicious access. Observers say a mix of narrow export rules, cross-border research waivers, and coordinated sanctions on demonstrable bad actors could reduce strategic mistrust without stifling the normative and technical work required to keep advanced AI systems aligned with public interest.
| Control focus | Recommended reform |
|---|---|
| Specialized chips | Targeted licensing with audit trails |
| Model weights | Controlled disclosure + certified safety review |
| Research exchange | Visa exemptions for accredited safety teams |
Concrete Confidence Building Measures Can Bridge the Gap: Joint Red Team Exercises, Standardized Incident Reporting and Reciprocal Third Party Audits
Capable, verifiable drills between U.S. and Chinese teams could convert suspicion into measurable safety gains. Observers and technical leads on both sides have proposed a program of mutual red-team exercises that use shared threat scenarios, independent referees and pre-agreed transparency gates so results are credible to outside experts. Practical steps being discussed by policy circles include:
- Shared scenarios-aligned test cases that reflect dual-use risks rather than national priorities;
- Transparent metrics-common benchmarks for model behaviour, failure modes and remediation;
- Observer access-trusted third-party monitors and rotating delegations to witness live tests;
- Data-use agreements-clear rules for what can be recorded, published and kept confidential.
Journalists and analysts say these measures, if institutionalized, would create a public record of capability testing and reduce the incentive to conceal hazardous advances.
Equally important are standardized incident reports and reciprocal audits that make cross-border comparison routine and fast. A compact incident template-what happened, timeline, affected models, mitigation steps-paired with a commitment to publish redacted summaries within fixed windows would change incentives. Independent audits, conducted by vetted firms on a reciprocal basis, would verify adherence to patching, monitoring and access controls. Suggested operational elements include:
- Standardized fields-uniform taxonomy for categorizing misbehaviour;
- Rapid disclosure-clear timelines for notifying counterparts and an agreed public summary;
- Reciprocal audits-annual third-party reviews with the right to follow up on corrective action.
| Measure | Purpose | Cadence |
|---|---|---|
| Incident Template | Faster, comparable reports | Immediate |
| Third-Party Audit | Independent verification | Annual |
| Joint Playbooks | Coordinated response steps | Quarterly |
Analysts caution that these technical fixes will only stick if accompanied by political guarantees and enforcement mechanisms that both governments trust; otherwise they remain voluntary checklists rather than durable institutions.
Build a Multilateral Enforcement Framework Combining Binding Norms on Dual Use Research, Investment Screening and Independent Verification While Protecting Scientific Collaboration
Policymakers in Washington and Beijing are quietly converging on a hard truth: a global AI safety architecture cannot succeed if it forces scientific communities into a zero-sum posture. Experts at recent diplomatic back-channels argued for a calibrated package that treats technological risk as a shared problem, not merely a national security threat. The proposal centers on binding norms for dual-use research-clear prohibitions, mandatory risk assessments and licensing for sensitive projects-backed by coordinated export controls and aligned investment screening to prevent rapid decoupling that would fragment research ecosystems. Absent a multilateral framework, analysts warn, unilateral measures will incentivize secrecy, accelerate strategic competition and undermine the very transparency necessary to detect misuse.
Practical enforcement would combine four mutually reinforcing tools delivered through an international secretariat and peer-review mechanisms:
- Binding norms on dual-use research – standardized risk categorizations and mandatory reporting for high-risk work;
- Harmonized investment screening – shared criteria to flag acquisitions that threaten critical AI capabilities or supply chains;
- Independent verification – regular audits, on-site inspections and a rapid fact-finding mechanism to resolve disputes;
- Protected channels for scientific collaboration – safe-harbor agreements and vetted consortia to preserve cross-border research on benign and beneficial projects.
| Mechanism | Primary Purpose |
|---|---|
| Binding Norms | Prevent misuse while standardizing oversight |
| Investment Screening | Shield critical capabilities without blanket bans |
| Independent Verification | Build confidence and deter violations |
| Protected Collaboration | Maintain scientific progress and trust |
Experts emphasize that the framework must be politically balanced-robust enough to constrain malign actors, flexible enough to allow responsible cooperation-and backed by transparent reporting to reassure publics on both sides of the Pacific.
In Summary
As Washington and Beijing circle one another with mutual suspicion, the question is no longer whether AI will transform economies and militaries, but whether the world can agree on the rules that will govern that transformation. Experts warn that without sustained cooperation between the two largest AI powers, efforts to set global standards, manage risks and respond to crises will be piecemeal at best and dangerous at worst.
Finding common ground will require political will, transparent channels for technical exchange, and third‑party mechanisms to build trust – steps that run up against deep strategic rivalry and domestic pressures in both capitals. Yet the alternative is clear: a fragmented landscape of competing standards, accelerated arms‑race dynamics and weakened global institutions at a moment when coordinated oversight is most needed.
The path forward remains uncertain, but the choices are not. If global AI safety is to be more than a slogan, U.S. and Chinese policymakers must reckon with the reality that their actions affect not just each other, but the entire world. Time is short; the consequences of inaction will be felt far beyond their borders.
