Kimi K3 Highlights the 'Open-Washing' Problem in AI Models

When Chinese startup Moonshot's Kimi K3 model landed among the top five on industry benchmarks and outperformed expensive proprietary systems in coding tests, it generated predictable excitement. The technical achievement is legitimate. But the model's release also crystallizes a growing problem in AI development: the gap between "open-source" as a marketing claim and open-source as a verifiable standard.

According to JJ Jasser, a professor and director of data analytics at Rollins College writing in Tech Policy Press, that gap has consequences that extend far beyond semantics. When he tested K3 with questions about the 1989 Tiananmen Square protests, the model refused to provide information. When asked to build an educational app about major historical protests, it covered events from England, India, and the Arab Spring while omitting anything from China's history.

"The model did not refuse to teach; it taught confidently and incompletely, leaving the user no way to see what was missing," Jasser wrote. "Any educational tool built this way inherits how the Chinese government sees the world."

Four criteria for genuine openness

Jasser and his co-authors recently published research establishing concrete standards for what qualifies as open-source AI. Their framework requires four components to all be publicly available under licenses permitting unrestricted use, modification, and redistribution: architecture, training code, model weights, and training data.

By that measure, very few models qualify. The Allen Institute's OLMo, EleutherAI's GPT-NeoX, and LLM360's K2 meet the standard. Kimi K3 does not.

Moonshot promised to release K3's weights on July 27 under a "Modified MIT" license with added conditions that is not an Open Source Initiative-approved license. The training data and training pipeline will remain closed. At best, K3 is an "open-weight" model that users can download and run but never fully study or reconstruct.

Why it matters

The distinction between open-weight and open-source determines what researchers and educators can actually understand about a model's behavior. When a model refuses to answer questions or provides systematically incomplete information, only transparency into the training data can reveal whether events were absent from the data, removed during curation, or suppressed afterward. Releasing weights allows testing what a model avoids, but cannot expose the shaped worldview embedded in curated training data.

The stakes extend beyond individual models. At the World Artificial Intelligence Conference in Shanghai, China declared open source a "vital pathway" to inclusive AI development and announced the World Artificial Intelligence Cooperation Organization, the first intergovernmental body dedicated to AI, headquartered in Shanghai and aimed at the Global South. As Carnegie Endowment researchers have documented, this represents Beijing's pivot toward reshaping global AI governance norms.

"When 'open' becomes a geopolitical brand, the difference between open-source and open-weight stops being pedantic; it determines who can inspect the systems entire regions will build on," Jasser wrote.

Fine-tuning open weights cannot restore history that was never included in training data. Until Chinese AI companies provide transparency their government seems unlikely to allow, benchmark performance alone should not justify adoption in education or other trust-sensitive domains.

These details were first reported by JJ Jasser in Tech Policy Press.