🇭🇰SCMP Technology•Recentcollected in 20m
Global Poll: China’s AI Leads in Capability, Lags in Trust
💡Understand how geopolitical sentiment impacts global AI adoption and the trust gap facing Chinese models.
⚡ 30-Second TL;DR
What Changed
Public First survey covered 18,000 participants across 15 countries.
Why It Matters
The findings suggest that geopolitical perception significantly influences AI adoption, potentially creating a bifurcated global market. Companies relying on Chinese AI infrastructure may face challenges in Western markets due to these trust barriers.
What To Do Next
If building global AI products, prioritize transparency and data governance to mitigate trust-related adoption barriers in international markets.
Who should care:Founders & Product Leaders
🧠 Deep Insight
Web-grounded analysis with 16 cited sources.
🔑 Enhanced Key Takeaways
- •The Public First survey, conducted by a London-based research firm, involved over 18,000 participants across 15 countries and found that only the United States, Japan, India, and Vietnam still perceive America as the leading force in AI.
- •China's AI models face a significant trust deficit, ranking 10th with a negative net trust value of -8, in contrast to Japan (+22) and the US (+16).
- •Public attitudes in the United States show a noticeable decline in optimism regarding AI's societal benefits, dropping from 39% in 2024 to 31% in 2026, with increasing concerns about employment and information integrity.
- •China plans to invest approximately $295 billion over five years to establish a nationwide network of AI data centers, with an ambitious goal of sourcing at least 80% of the underlying technology, including AI chips, from domestic suppliers by 2028.
- •Chinese AI models like DeepSeek-V3, Qwen 3, Doubao 1.5 Pro, Kimi k2, and WuDao 3.0 have solidified their position as global competitors, excelling in areas such as natural language processing, multimodal capabilities, and coding.
📊 Competitor Analysis▸ Show
| Feature/Model | China (e.g., DeepSeek-V3, Qwen 3, GLM-5, Ernie 5.0) | Western (e.g., ChatGPT, Claude, Gemini) |
|---|---|---|
| Architecture | Predominantly Mixture-of-Experts (MoE) for efficiency, activating a fraction of parameters per query (e.g., DeepSeek-V3 uses 37B of 671B parameters). | Often dense models, though MoE is also adopted by some. |
| Parameters | DeepSeek-V3: 671 billion (MoE). Qwen 3 Max: Trillion-parameter monster. | Varies, with leading models often in the hundreds of billions to trillions of parameters. |
| Context Length | Up to 128,000 tokens (DeepSeek-V3, Qwen 3). | Comparable or varying, with continuous improvements. |
| Multimodality | Strong multimodal capabilities (text, images, video, code) (e.g., Qwen 3, Ernie 5.0). | Strong multimodal capabilities (text, images, video, code). |
| Cost Efficiency | Qwen 3: $0.38 per million tokens, noted as cheaper than models like GPT-4o. DeepSeek: Cheapest option at $0.14-0.30/M input tokens. | Generally higher pricing, though competitive options exist. |
| Key Strengths | Coding, mathematical reasoning, long-form content generation (DeepSeek-V3). Multilingual processing (Qwen 3.5 for Chinese, Japanese, Korean). Agentic coding tasks (GLM-5, Kimi K2.5). Unified multimodal architecture (Ernie 5.0). | Advanced reasoning, broad general knowledge, strong safety alignment (varies by model). |
| Benchmarks | GLM-5: 85 on BenchLM, 77.8% SWE-bench Verified. DeepSeek-V3: MMLU: 88.5; MATH-500: 90.2; Codeforces: 51.6. Ernie 5.0: Beat ChatGPT and Claude on LM Arena leaderboard. | Often top scores on various public benchmarks, though Chinese models are closing the gap. |
| Open-Source Availability | DeepSeek models are notable for being open-weight. Alibaba open-sourced QwQ-32B. GLM-5 has MIT licensing. | Varies, with some models being proprietary and others open-source. |
| Content Restrictions | All Chinese models carry hard-coded content restrictions on politically sensitive topics. | Generally fewer explicit content restrictions, but ethical guidelines and safety filters are in place. |
🛠️ Technical Deep Dive
- DeepSeek-V3: Features a Mixture-of-Experts (MoE) architecture with 671 billion parameters, dynamically activating only 37 billion parameters per input for high performance and reduced computational cost. It supports an extended context length of up to 128,000 tokens and uses multi-token prediction for faster response generation.
- Qwen 3 (Alibaba): Utilizes a Mixture-of-Experts (MoE) architecture and is trained on over 20 trillion tokens. It is a multimodal AI capable of understanding and generating text, images, and video, with an extended context length of up to 128,000 tokens.
- GLM-5 (Zhipu AI): Achieved a score of 85 on BenchLM's open-weight leaderboard and 77.8% on SWE-bench Verified for agentic coding tasks. It is trained entirely on Huawei Ascend chips and is available under an MIT license.
- Ernie 5.0 (Baidu): Features a unified multimodal architecture that processes text, images, video, and code together, ensuring consistent context, tone, and reasoning. It has reportedly surpassed models like ChatGPT and Claude on the LM Arena leaderboard.
- General Chinese AI Models: Many prominent Chinese AI models leverage Mixture-of-Experts (MoE) architectures, which can reduce inference compute by 90-97% compared to dense models of equivalent knowledge capacity.
🔮 Future ImplicationsAI analysis grounded in cited sources
Geopolitical tensions surrounding AI technology will intensify, leading to more stringent export controls and increased calls for allied cooperation.
The perceived AI leadership of China, coupled with national security concerns among US allies, will likely prompt Western nations to further restrict China's access to advanced AI and semiconductor technologies, while simultaneously strengthening their own collaborative efforts.
China's ambitious goal for domestic AI chip production will face significant hurdles, potentially delaying its self-sufficiency targets.
Despite massive state-directed investment, the target of 80% domestic sourcing for AI chips by 2028 will be challenged by the limited output of advanced semiconductor nodes from local foundries and constraints in high-bandwidth memory (HBM) supply.
Public distrust in AI, particularly concerning job displacement and information integrity, will continue to grow in Western countries, influencing regulatory debates.
Survey data indicates a sustained decline in optimism about AI's societal benefits and heightened anxiety over its impact on employment and information, especially among younger demographics in the US and UK, which will likely fuel public demand for stricter AI governance.
⏳ Timeline
2017-07
China's State Council issues the 'New Generation Artificial Intelligence Development Plan,' outlining a national strategy to become a world leader in AI by 2030.
2023-08
China implements interim regulations for generative AI services, requiring content labeling and filing information with a government registry.
2025-01
China's TC260 releases a draft 'AI Safety Standards System (V1.0),' indicating a rapid translation of governance frameworks into binding technical standards.
2025-03
DeepSeek's R1 model is released, shaking up the industry and raising questions about the US lead in AI.
2026-01
DeepSeek-V3, Qwen 3, Doubao 1.5 Pro, Kimi k2, and WuDao 3.0 are highlighted as solidified global competitors among Chinese AI models.
2026-06
Public First survey reveals global perception of China's AI capability lead but significant trust deficit.
📎 Sources (16)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
📰
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: SCMP Technology ↗