China’s Certification System for Trustworthy AI: Can You Trust the Machines?

With AI creating news, driving cars, and making employment decisions, one question stands out: Which algorithms can we place our trust in?  

The Chinese government has taken the “trustworthy AI” challenge head-on, creating a formal certification system to mark the high-risk AI technologies being developed in the country.  

In this post, we elaborate on the importance of China’s Trustworthy AI Certification System, what lessons can be drawn from it, and how the system strives to provide safer AI. If you are a digital policy expert, an AI developer, or even interested in how countries are evolving with the new era of machine intelligence, this read can be enlightening.

________________________________________

Trustworthy AI as a Global Issue

AI technology is no longer limited to being a concept in a lab; it can now be found in mobile devices, employment websites, financial institutions, and even courtrooms. With new capabilities, however, come new risks and concerns. These include:  

• Facial recognition bias  

• Deep learning models with unexplainable rationales  

• Invasion of privacy through data collection   

• Automated spread of false information  

• Job discrimination and biased credit scoring  

As tasks become more automated and critical, there’s a added loss of trust, greater ethical concern, and increased demand for transparency around systems. This is where trustable AI certification systems come into play.

________________________________________  

China’s Approach: Integrating Trust into AI Systems

MIIT (Ministry of Industry and Information Technology) and the CAICT (China Academy of Information and Communications Technology) developed the first national framework for the Trustworthiness Certification Framework for AI (TCFA) in 2022.

The primary aim of this framework focuses on ensuring AI systems are assessed on:  

1. Security

2. Equity  

3. Justifiable reasoning  

4. Privacy  

5. Responsibility 

This is more than just creating standards. It’s about rethinking the processes for building and implementing AI systems.

________________________________________  

What is the Trust AI Certification?

The Trustworthy AI Certification is a voluntary, albeit powerful designation that indicates a system meets China’s internal specifications of AI ethics.

Applied Evaluation Systems like CAICT and other designated testing labs accept applications from AI developers, be it from a private company, academic lab or government institution.

What Gets Evaluated? 

Let’s look at the primary criteria:

• Safety: Does the AI system prevent unwanted behaviors from happening? Are risks mitigated in real time applications like self-driving cars or financial systems?

• Fairness: Is the model trained on data that is diverse and free of bias? Does it provide consistent results across gender, ethnicity, or region?

• Explainability: Can the processes of the AI model be comprehended by a human? Is there a trace of the logic or the inputs which were used?

• Privacy: Is the personal information encrypted or protected? Does the system adhere to the Personal Information Protection Law (PIPL)?

• Accountability: Is there clear allocation of responsibility in the case of failure or harm? Is there a person in the loop?

Once all of these guidelines are satisfied, the system in question becomes certified with 'Trustworthy AI,' but only for a limited duration. 

________________________________________

Use Case: Baidu's Autonomous Driving Platform

Baidu's Apollo autonomous driving system gained one of the first Trustworthy AI certificates.

• The system was thoroughly tested for safety in edge-case scenarios like pedestrian movement (e.g., uncivilized walking).

• It showed no bias in urban area decisions and other drivabilities, and the system’s driving performance in varying conditions demonstrated consistent urban area decision-making without bias.

• Clarifying and justifying automated system actions is often maligned for a lack of transparency, but Baidu used system justification to explain decision-making in an unprecedented manner. 

This certification was crucial for Baidu’s regulatory approval in operating robotaxi’s in Wuhan and Chongqing, where trust and safety are paramount.

________________________________________

A Step Further: Aligning with China’s AI Governance Vision

The scope of the certification framework is in perfect alignment with China’s AI governance system that incorporates:

• The 2021 AI Ethics Guidelines noted the need for a human-centric, controllable, and trustworthy framework for AI.

• Algorithmic Recommendation Regulations (2022) placed control and disclosure requirements for AI on social media, e-commerce, and other recommendation platforms.

• Deep Synthesis Provisions (2023) governs the labeling of deepfake videos and other synthetic content.

Together these policies suggest that China is pursuing a dual aim: to emerge as the world’s preeminent AI superpower and assume responsibility for managing such power. 

________________________________________

Benefits of Certification: For Companies and the Public

For AI businesses, it’s not so much about obtaining legally recognized certification, as it is about gaining a competitive edge.

For companies:

• Gaining access to smart cities, healthcare, or autonomous mobility becomes expedient with regulatory clearance.

• Stronger brand reputation amongst an increasingly cautious public wary of algorithmic big-tech malfeasance.

• Enhanced partner and international investor trust, including domestic trust from Chinese stakeholders.

For The Public:

• Better public AI data disclosures.

• Heightened control over user interactions, particularly in financial, medical, and social applications.

• Stricter measures against unlawful surveillance or slanted algorithmic governance.

______________________________________________________________________

Challenges in Implementation

Creating a certification system for AI (without parts of it being arbitrary or vague) is anything but straightforward.

1. Measuring Explainability

Not all models are interpretable, and deep learning systems are famously inscrutable. It is difficult, if not impossible, to prove “explainability”.

2. Data Access for Audits

Evaluators require proprietary data, Tend to Eye. They need access to the training data, model’s architecture, and logs of outputs. Companies may be reluctant to share this information due to IP or trade secret risks.

3. Global Interoperability

Will it be accepted in China? What about the EU AI Act or OECD AI Principles? There are still a lot of work in progress on international cooperative systems.

___________________________________________________________________

The Global Implications: Will Others Follow?

With some refinement, China's proposal may serve as a blueprint for other developing nations or technology exporters looking for reliable structure on AI trustworthiness.

For instance:

• Southeast Asian or African countries could follow the same checklist-based structure to screen foreign AI tools utilized in public service.  

• Tech companies giving AI products to China may have to integrate trust-by-design approaches from the onset.   

In summary, China does not only control AI; actionably, this flows into global benchmarks. 

________________________________________ 

Final Thoughts: Trust In An Era of Intelligent Systems  

With adjustments in Artificial Intelligence (AI) integration into daily tasks, the central expectation shifts from being a preference to a necessity. China’s proactive creation of a framework in the form of The Trustworthy AI Certification System Trustworthy AI Certification attests to such belief.  

As a developer, AI policy strategist or changer, or as someone who relies on these technologies, there’s an undeniable reality: the world of AI is not only about its capabilities moving forward, responsibility comes into play.  

Once trust is gone, rebuilding becomes a challenge. With this set of measures, China assumes the position that for AI to be deemed powerful, it needs to be responsible.