Concordia AI

Stated Theory of Change

Concordia claims that the Western narrative -- "China doesn't care about AI safety" -- is both false and dangerous. False because China has binding AI regulations, pre-deployment testing requirements, national safety standards, and a growing safety research community. Dangerous because the narrative gives Western companies and governments cover to avoid safety measures ("why bother if China won't?").

Concordia's mechanism: reduce information asymmetry by translating Chinese AI safety developments for Western audiences and vice versa, while directly improving safety practices in Chinese labs through consulting, and shaping Chinese national standards from within.

The causal chain: more accurate Western understanding of Chinese safety efforts --> less "race to the bottom" pressure in Western AI policy --> more space for both countries to implement safety measures --> cooperation on shared catastrophic risks (biosecurity, loss of control) --> reduced global AI risk.

Revealed Theory of Change

Actions largely align with stated theory, but with important nuances:

What they actually do most: Produce English-language reports, host high-profile convenings, and consult for Chinese AI labs on safety compliance. The consulting revenue (EUR 6.12M) is the financial engine. The reports and convenings are both genuine public goods and marketing for the consulting practice.

Where actions diverge from stated theory: The stated theory emphasizes honest brokering between two ecosystems. But several patterns suggest the brokering is systematically tilted:

1. Concordia consistently emphasizes convergence over divergence. Brian Tse repeatedly says Chinese and Western safety approaches are "95% similar." The 5% that differs -- political censorship encoded in safety standards, development-first hierarchy, absence of voluntary safety commitment compliance -- receives much less emphasis in Concordia's public communications than in independent assessments (Carnegie, China Media Project, AI Frontiers).

2. The consulting business creates selection pressure. When your revenue comes from Chinese AI labs, you have incentives to (a) present those labs' safety efforts favorably to Western audiences, (b) avoid publishing findings that embarrass your clients, and (c) frame the Chinese safety landscape in ways that make the consulting services seem more valuable.

3. The standards committee participation creates a different kind of value than the public communications suggest. Concordia helps write Chinese AI safety standards -- which is real influence -- but those standards include requirements for "core socialist values" compliance that are fundamentally about political control, not catastrophic risk reduction. Concordia does not publicly acknowledge this tension.

Key Assumptions

Assumption 1: Information asymmetry about China's AI safety efforts is causing worse Western policy.

• Evidence for: Brian Tse's point that many Western policy conversations are "explicitly premised on assumptions like 'China will never slow down their AI development' or even 'China doesn't care about AI safety'" (Nathan Labenz, Cognitive Revolution). These assumptions do appear to shape US policy.
• Evidence against: Even with perfect information, geopolitical competition would likely still dominate. The ITIF/PolicyArena argument is that cooperation is structurally undesirable regardless of China's safety posture.
• Testable? Partially. If Concordia's reports demonstrably changed specific policy decisions, that would be evidence. No such causal chain has been documented.
• If wrong: Concordia's information-bridge work is valuable but not high-leverage. Safety improvement would need to come from other mechanisms.

Assumption 2: Chinese AI safety efforts are genuine and not purely performative.

• Evidence for: Third Plenum decision classifying AI safety as public safety concern. Politburo study session on AI. 31+ research groups publishing safety papers. Shanghai AI Lab's Frontier AI Risk Management Framework with technically credible content. National emergency response plan including AI risk.
• Evidence against: "Failing to develop is the greatest threat to security." Zero Chinese companies have fulfilled Seoul AI Safety Commitments. CnAISDA has "little substantive action." Chinese safety standards encode political censorship. Regulatory compliance is the primary driver, not intrinsic x-risk concern.
• Testable? Yes -- watch for binding regulations on CBRN/loss-of-control testing, not just content-focused compliance.
• If wrong: Concordia's core narrative collapses and its consulting business becomes complicit in safety-washing.

Assumption 3: US-China cooperation on AI safety is both feasible and valuable.

• Evidence for: Bletchley Declaration signed by both. Joint statement on human control of nuclear weapons. IDAIS consensus statements. Chinese scientists (Yao) signing joint safety papers. Both countries acknowledge CBRN and loss-of-control risks.
• Evidence against: Last intergovernmental AI discussion was May 2024. BAAI placed on US Entity List (March 2025). Trump administration skeptical of multilateral AI governance. Export controls creating resentment. Structural mistrust on both sides.
• Testable? Yes -- will there be a resumed US-China AI safety dialogue before 2027?
• If wrong: Concordia's convening work becomes a series of nice conferences with no policy impact, while the cooperation window closes.

Strengths

1. Unique positioning. No other organization occupies Concordia's exact niche: Beijing-based, EA-rooted, bilingual, with genuine access to both Chinese policy institutions and Western AI safety networks. Brian Tse's personal network spans GovAI/Oxford, PAI, OpenAI, BAAI, Tsinghua, Carnegie, and the UN. This is irreplaceable.

2. Real policy influence in China. Sitting on 3-4 national standards committees as a social enterprise (not a state lab) is remarkable. Co-publishing the Frontier AI Risk Management Framework with Shanghai AI Lab gives Concordia quasi-official influence on how China approaches frontier risk. This is more concrete policy impact than most Western AI safety orgs achieve.

3. Self-sustaining business model. EUR 6.12M in consulting revenue with zero external funding means Concordia is not dependent on philanthropic cycles or funder preferences. This is unusual in the AI safety space and provides financial resilience.

4. Technically credible outputs. The AI Risk Monitoring Platform, Framework v1.0, and technical report evaluating 20+ models across 7 risk domains represent genuine technical contributions. Jack Clark's endorsement of convergence with Western evaluations validates the technical approach.

5. Speed and efficiency. Growing from 5 to 14 staff while generating 6M+ EUR revenue, attending every major global AI summit, and producing multiple 100+ page reports -- this is a remarkably productive small team.

Weaknesses and Risks

1. The incentive problem is structural and unaddressed. Concordia's revenue depends on Chinese AI labs continuing to pay for safety consulting. Concordia's public reputation depends on presenting Chinese AI safety favorably. No governance mechanism exists to separate these interests. This is not a hypothetical concern -- it predicts exactly the pattern of emphasis we observe (convergence highlighted, divergence downplayed).

2. The "AI safety" translation problem. When Concordia says Chinese and Western AI safety approaches are converging, this smooths over a fundamental difference: Chinese "AI safety" (anquan) encompasses both catastrophic risk concerns AND political security concerns. The 31 official safety risks put "core socialist values" violations at the top. Concordia's public communications do not adequately convey this distinction, which could mislead Western audiences about the nature of Chinese "safety" work.

3. Operating under CCP jurisdiction with no editorial independence protections. As a Beijing-registered company, Concordia cannot publish findings that seriously embarrass the Chinese government. This constraint is never acknowledged in their public materials. Consumers of Concordia's analysis should assume that negative findings about Chinese AI safety are systematically underrepresented.

4. Founder dependency. Brian Tse IS Concordia AI. His personal network, reputation, and relationships are the organization's core asset. If Tse departed, became politically compromised, or lost access to Chinese policy institutions, the organization's value proposition would largely collapse.

5. No external governance. No board, no advisory committee, no published conflict-of-interest policy. For an organization claiming to serve as an honest broker between two ecosystems, this is a significant credibility gap.

6. Declining geopolitical environment. US-China relations are trending toward competition and decoupling. BAAI's Entity List designation, Trump administration skepticism of multilateral governance, export control escalation -- all reduce the space for Concordia's cooperation agenda. Concordia's work may be most needed exactly when it is least politically feasible.

Cross-References

• GovAI/Oxford Martin (AIGI): Brian Tse is a GovAI Policy Affiliate. Kwan Yee Ng is an Oxford Martin research affiliate. Concordia co-hosted events with AIGI at Paris AI Action Summit and WAIC. This is their primary Western institutional anchor.
• Carnegie Endowment: Concordia collaborates extensively with Carnegie -- co-convening workshops, cross-referencing reports. Carnegie provides the more nuanced/critical external assessment of Chinese AI safety that Concordia's own reporting tends to soften.
• Shanghai AI Lab: Co-published the Frontier AI Risk Management Framework. Shanghai AI Lab is the key state-affiliated partner that gives Concordia's technical work quasi-official status in China.
• FAR.AI: Co-hosted AI Safety Social and workshops at ICLR 2025. Nathan Labenz compared Concordia to FAR.AI in the podcast.
• Beijing AISI (CnAISDA): Concordia was invited to CnAISDA closed-door seminars. CnAISDA is the new Chinese government-backed AI safety network. Concordia's relationship with CnAISDA could significantly enhance or constrain its role depending on how CnAISDA evolves.
• Partnership on AI: Concordia became a formal member in 2025. Brian Tse was formerly PAI Senior Advisor.

What Would Change This Assessment

Upward revisions if:

• Concordia publicly names its consulting clients and publishes a conflict-of-interest policy
• Concordia produces a report that explicitly acknowledges the divergence between Chinese political "safety" and catastrophic risk "safety"
• A major Chinese lab demonstrably changes its safety practices based on Concordia's consulting (with evidence)
• The Frontier AI Risk Management Framework is adopted as binding regulation by a Chinese government body
• The US-China AI safety dialogue resumes, with Concordia playing a documented role in facilitating it

Downward revisions if:

• Concordia's reporting is shown to systematically omit negative findings about paying clients
• Chinese government restricts Concordia's international publishing or forces editorial changes
• Brian Tse departs or is politically compromised
• The AI Risk Monitoring Platform produces findings unfavorable to Chinese models and the findings are subsequently softened or retracted
• Consulting revenue declines and Concordia pivots to a purely advocacy/marketing role for Chinese AI labs

Self-Critique

What sources should I have checked but didn't:

• The actual full State of AI Safety reports (100-150 page PDFs) -- reading these in full would allow assessment of how Concordia handles sensitive findings within the reports vs. in their public summaries
• Concordia's WeChat content (Chinese-language) -- the messaging to Chinese audiences may differ from English-language messaging, which would be highly informative
• Nature News articles citing Concordia (paywalled/blocked)

Where is this analysis potentially biased:

• I may be over-weighting the incentive conflict. Many think tanks and consultancies have similar structural tensions and still produce valuable work. The question is degree, not existence, of bias.
• I may be under-weighting the genuine difficulty of operating in China while maintaining international credibility. The constraints are real and Concordia navigates them better than any comparable organization.
• Western/English-language sources are over-represented in my evidence base, which may systematically under-represent Concordia's Chinese-language impact.

What would a thoughtful person who disagrees say: "You're holding Concordia to an impossible standard. They're the only org doing this work from inside China. Of course they can't be openly critical of the CCP -- that would get them shut down. The fact that they exist at all, produce English-language intelligence, and influence Chinese safety standards from within is enormously valuable. Your 'incentive conflict' analysis is the kind of pure-logic criticism that ignores the messy reality of doing policy work in authoritarian environments."

What's my single weakest claim: That Concordia's emphasis on convergence (vs. divergence) between Chinese and Western AI safety is primarily driven by business incentives rather than genuine assessment. It could be that Tse genuinely believes the convergence narrative based on his insider knowledge, and the incentive alignment is coincidental rather than causal. I cannot distinguish these explanations from the available evidence.

What information would most change my view: Seeing Concordia produce a report or public statement that explicitly costs them something -- either with Chinese authorities or with consulting clients. For example, naming a specific Chinese lab that failed to meet safety commitments, or acknowledging that certain Chinese safety regulations primarily serve political control rather than catastrophic risk reduction. If they did this, it would significantly increase confidence in their independence.