When Donald Trump and Xi Jinping agreed on the latest APEC summit to set a flooring beneath spiraling China-U.S. commerce relations, in addition they agreed to contemplate cooperation on synthetic intelligence (AI) within the 12 months forward. With Trump and Xi planning an alternate of visits in 2026, now could be the time for a “sensible agenda” to establish which parts of synthetic intelligence make sense for Washington to debate with its prime strategic rival.
Each side agree that AI will form their near-term safety planning, driving an intense competitors over compute, cloud infrastructure, frontier AI fashions, and the power wanted to gas them. The Trump administration’s AI Motion Plan captures that focus. It’s skeptical of multilateral efforts and cautious of technological cooperation with Beijing on superior methods.
China’s new World AI Governance Motion Plan presents itself as a multilateral different. It helps Beijing’s ambition to form international requirements and develop its technological affect by proposing to create a brand new worldwide group primarily based in Shanghai.
If Washington and Beijing strategy AI as nothing greater than a zero-sum contest, they could miss the small however essential areas for cooperation on protocols, evaluations, and verification that may set a “sensible agenda” for future use of AI. These quiet, technical steps can preserve a strong expertise from slipping past both aspect’s management or into the palms of criminals or terrorists. Those self same steps may function the primary strikes towards a deeper, risk-reducing relationship.
Trump himself has already cracked the door open on some promising areas for cooperation – most notably on the intersection of AI and biotechnology. In his 2025 United Nations Normal Meeting speech, he referred to as for a world effort to finish “the event of organic weapons as soon as and for all” and proposed utilizing AI to assist confirm compliance. His AI Motion Plan equally warns that AI and artificial biology “might create new pathways for malicious actors to synthesize dangerous pathogens and different biomolecules” and proposes worldwide screening tips to cut back these dangers.
To observe up, Washington will want a brief, concrete menu of points that decrease shared safety dangers with out handing Beijing a strategic edge. On the identical time, the US can set up a strategy of steady professional alternate free from ups and downs within the bilateral relationship – a lot because the U.S. and Soviet Union allowed professional arms management talks to proceed all through the Chilly Struggle.
How Geopolitical Rivals Have Managed Harmful Applied sciences Earlier than
Historical past suggests that when rivals handle harmful applied sciences, they normally begin with tightly certain, low-risk measures that later allow deeper cooperation as confidence slowly builds. Through the Chilly Struggle, Washington and Moscow constructed slender agreements on nuclear testing, incident reporting, and disaster hotlines lengthy earlier than there was something like belief. They traded restricted technical info, arrange verification rituals, and created habits of communication that helped each side keep away from worst case misunderstandings and accidents. These talks continued in an insulated professional channel and weren’t lower off as a automobile to reveal political displeasure. None of that ended the arms race. But it surely made the arms race much less more likely to finish the world.
Cryptography provides the same sample. The U.S. authorities ran the Superior Encryption Normal (AES) as an open worldwide competitors. Researchers all over the world examined, attacked, and improved candidate algorithms. That transparency strengthened civilian safety instruments whereas delicate navy methods remained labeled. The strategy additionally inspired vast international adoption and customary implementation requirements. It reveals how modest early steps can develop into extra sturdy types of technical collaboration. In each instances, cooperation grew first the place sharing posed little danger, the threats had been international, and each side feared the identical disasters.
AI is now getting into that very same territory. Neither Washington nor Beijing will wish to share secrets and techniques about AI functions that would permit the opposite to establish nuclear-capable submarines or determine the destiny of Taiwan. However each are conscious that AI poses unknown dangers and far stays to be realized. Establishing processes and practices for accountable AI analysis and use might allow Washington to construct the identical type of slender flooring it insisted on in earlier expertise races.
The problem is to search out these few slender openings for AI the place restricted cooperation with Beijing can preserve a harmful expertise from outrunning each international locations’ skill to manage it.
Shared Issues About Superior AI Techniques
Whilst Washington warns about China’s ambitions in synthetic intelligence, official paperwork on each side more and more flag overlapping safety risks. Current U.S. technique papers and govt actions, and China’s nationwide plans and regulatory steerage, all stress dangers from unsafe or uncontrollable methods and from malicious misuse at scale. The language differs – and Beijing places extra weight on info and social-stability harms – however the underlying fears considerably converge. Lack of management, decreasing obstacles for malicious actors, misalignment and failure to repair accidents prime the listing of widespread issues.
Three areas benefit particular focus: harmful AI capabilities, testing in opposition to crucial design dangers, and stopping assaults involving misuse and deception prime the listing of overlapping safety risks. U.S. technique papers, govt actions, and technical requirements, and China’s nationwide plans, AI rules, and rising security frameworks, all deal with these three clusters of dangers.
First are AI’s harmful capabilities. Each governments fear that superior fashions might decrease the barrier for outsiders to plan refined cyber operations or design organic and chemical weapons, and that extra autonomous methods might behave unpredictably as soon as deployed.
Second, accountable design is crucial to having highly effective methods that behave constantly and resist manipulation. Washington and Beijing might agree on rigorous testing necessities earlier than brittle AI is woven into crucial infrastructure or monetary markets. And so they might require investigations – as within the case of airline accidents – if something goes improper.
The third danger cluster is deception and opacity. Regulators and tech corporations warn about methods that may impersonate people, develop new pathogens, or flood the knowledge area with artificial media. Watermarking, labeling, and different disclosure necessities for AI-generated content material are being adopted, whilst enforcement particulars stay unsettled.
It’s no accident that the areas the place each governments fear most – harmful capabilities, testing, and deception – map immediately onto the three cooperation lanes of protocols, evaluations, and verification. Researchers have already begun to map this rising widespread floor, highlighting why a deal with these three areas provides a number of the most real looking beginning factors for cooperation between Washington and Beijing.
Three Lanes for Sensible, Slender, Sensible Cooperation
Geopolitical rivalry guidelines out sweeping AI accords, however Washington can stabilize AI competitors with China by means of technical cooperation to construct the shared science, testing procedures, and early confidence that previous rivals have relied on. Shared protocols, analysis strategies, and verification instruments are among the many most promising – and least dangerous – beginning factors for cooperation between geopolitical rivals on AI.
Security frameworks and finest practices give each side a shared vocabulary for accountable improvement. Testing and analysis strategies assist them perceive whether or not superior methods behave safely and reliably exterior the lab and might help make sure that deadly accidents don’t recur. And verification mechanisms supply methods to examine that claims a few system’s safeguards or capabilities are literally true.
None of this requires sharing mannequin weights, proprietary information, or something near navy functions. However these modest steps can begin the sluggish work of co-developing strategies, evaluating notes, creating agreed procedures, and trusting one another’s primary measurements – precisely the type of early scientific cooperation that helped previous rivals handle shared dangers. Taken collectively, they decrease the probabilities that highly effective methods fail unpredictably or are misused in ways in which neither authorities can absolutely management.
Lane One: Protocols and Greatest Practices
The primary and least delicate lane is codified protocols and security frameworks – the broad, non-binding playbooks that define how highly effective AI methods ought to be designed, examined, monitored, and halted if one thing goes improper. Governments and corporations are already gravitating towards these sorts of paperwork, and lots of labs now publish their very own security frameworks as a sign of accountable follow.
Constructing on that momentum wouldn’t require sharing mannequin weights or navy functions. It might imply quiet technical work: agreeing on a primary glossary of security phrases, sketching out what any credible framework ought to cowl, and creating easy templates or case research that make expectations simpler to check throughout establishments. One caveat is that standard-setting can typically be used to tilt the enjoying area on the multilateral stage. Even so, the general alternative is obvious.
A future Trump–Xi assembly might propel a joint dedication to publish and periodically replace nationwide security frameworks that cowl a couple of shared parts – pre-deployment testing, incident response, and primary transparency about high-risk makes use of. These early steps assist set up a shared understanding of the dangers that superior methods create and a course of for addressing accidents – foundations that any deeper cooperation will finally depend upon.
Lane Two: Evaluations
The second space for potential cooperation entails testing and investigation protocols and mannequin evaluations. Mannequin evaluations reveal what superior methods can do, how reliably they behave, and the place failure might create national-security dangers. Even within the present local weather, Washington and Beijing might publicly establish dependable evaluations as a shared precedence and examine a small variety of technical strategies that don’t expose mannequin internals.
In follow, this might contain small professional delegations exchanging brief technical notes and displays, and sometimes working the identical analysis procedures in parallel on their very own methods to allow them to examine solely high-level outcomes relatively than any underlying information or fashions. This might embody evaluating strategies for recognizing benchmark contamination, which happens when a mannequin has already encountered elements of the take a look at and appears safer and extra succesful than it truly is. It might additionally contain bettering multilingual take a look at suites, so evaluations cowl greater than English and catch dangers that seem solely in different languages. A 3rd space is creating and sharing easy checks to see whether or not an analysis rating truly predicts how a system behaves exterior the lab, the place fashions typically act much less reliably.
Governments might additionally decide to supporting parallel analysis packages that strengthen and share enhancements to those checks, permitting universities on each side to develop higher instruments with none joint entry to delicate methods. On the diplomatic stage, that interprets to modest commitments: naming dependable evaluations as a shared precedence, asking nationwide AI institutes to check contamination-detection strategies and multilingual checks, and inspiring parallel analysis grants that push universities and trade towards higher instruments.
Most significantly, Washington and Beijing can agree on processes for AI testing and analysis, even when they preserve the content material of their methods secret. Agreements might talk about acceptable error charges and, most significantly, remediation. After a deadly airline crash, worldwide requirements require investigation and remediation in order that the underlying security flaws don’t recur. Equally with AI, the most important powers ought to agree now on the method by which they’ll take a look at and remediate errors that can inevitably come up as this new expertise is deployed into more and more refined and probably harmful methods.
None of this requires shared red-teaming or cooperation in high-risk domains. It merely provides a restricted path for every authorities to higher perceive the dangers in its personal nation’s fashions whereas conserving political and safety issues manageable. Even the act of recognizing bettering the science of AI evaluations as a respectable aim has worth in a relationship the place security is commonly overshadowed by competitors.
Lane Three: Verification
The third space the place the US and China would possibly be capable of collaborate now could be verification, doubtless essentially the most delicate of the three pathways. Verification doesn’t ask what a system can do; it asks whether or not the claims made about that system are literally true. Verifying mannequin capabilities is probably going important for any future settlement that hopes to be trusted.
There may be worth in merely naming verification analysis as a shared precedence, which might push trade, universities, and requirements our bodies to develop higher methods for mannequin identification, safe coaching information, and different constructing blocks of credible oversight.
A second step is cooperation on verifiable audits of public or low-risk fashions. All sides might run the identical audit procedures by itself methods and share solely the strategy and aggregated findings. That will assist confirm whether or not the safeguards they declare to make use of are literally in place and functioning to a minimal agreed commonplace, with out revealing delicate information or mannequin internals.
A ultimate space, and one with clear shared incentives on each side, is content material provenance. C2PA-style requirements might help affirm when textual content or photos had been generated or altered by AI, which issues as a result of each governments wish to forestall third events from utilizing artificial media to destabilize their very own nations. Even right here, nevertheless, each side could hesitate if enhancements in provenance seem to reinforce the opposite’s capabilities. U.S. officers, for instance, will probably be cautious of sharing instruments that could possibly be repurposed to strengthen Beijing’s management over info flows inside China. That’s the reason any early work on verification ought to stay tightly scoped, centered on output-level instruments, and handled as groundwork for the extra demanding cooperation that will come later.
In a leader-level assembly, the factor that’s most believable within the close to time period is for each side to endorse verification analysis as a shared aim and to ask their consultants to launch small pilots on verifiable audits of public fashions and primary output provenance, together with narrowly scoped pilot packages that stress-test confidentiality-preserving verification instruments.
A extra bold goal would possibly mix a deal with essentially the most harmful capabilities with an curiosity in verification, as foreshadowed by Trump’s U.N. speech on verification of high-risk organic weapons functions of AI. Some finest practices to cut back dangers from AI and biotechnology had been publicized on the December 2025 U.N. Organic Weapons Conference conferences, the place Trump’s Undersecretary for Non-Proliferation inspired worldwide AI cooperation to stop abuses of biotechnology.
Constructing a Flooring Beneath a Sharper Rivalry
As synthetic intelligence turns into extra succesful, the prices of getting AI safety improper will develop quicker than the advantages of conserving each concept to ourselves. Washington and Beijing are already locked right into a long-term competitors over chips, information, and fashions. That competitors won’t vanish, and the following administration is unlikely to embrace broad new types of expertise cooperation with China. The query is whether or not the US insists on competitors with none shared guardrails, or whether or not it’s keen to form a skinny layer of primary safety practices that serve its personal pursuits even when relations bitter.
Beginning with low-risk, narrowly technical cooperation is essentially the most real looking method to do this. Work on protocols and finest practices, analysis strategies, and early verification instruments won’t resolve deeper political disagreements. However it could possibly create a small set of shared expectations about how highly effective methods ought to be managed, how severe incidents are dealt with, and what sorts of failures each side agree are just too harmful to disregard. Biotechnology is a crucial place to begin.
This proposal just isn’t a grand cut price. It’s a flooring. In a future disaster involving superior AI, American officers will wish to know that at the very least a number of the language, instruments, and habits for managing these dangers had been constructed prematurely. Treating AI as nothing however a race dangers leaving each side to improvise at the hours of darkness.
Constructing a skinny flooring of shared follow is how rivals have dealt with harmful applied sciences earlier than, and AI ought to be no exception.
