Examining Transparency in AI Models

Jasser uses the example of Moonshot AI’s Kimi K3 model, which was promoted as open source despite releasing only model weights, rather than the full materials needed for independent scrutiny and replication.

Jasser contends that releasing weights alone does not meet established open-source standards. Without access to training data, data curation methods, post-training processes, and deployment practices, researchers and users cannot determine how a model was developed, why it behaves as it does, or whether it contains embedded biases and restrictions. He notes that this lack of transparency undermines one of the core benefits of open-source development: accountability.

Drawing on his testing of Kimi, Jasser highlights how the system could generate educational content on Western protest movements while omitting discussion of modern Chinese protest history. He argues that such examples raise important questions about how AI models are shaped and what influences their outputs. However, in the absence of transparency, users cannot identify whether those limitations stem from training data, model tuning, filtering systems, or other design decisions.

To address what he calls “open-washing,” Jasser proposes evaluating AI systems against clear criteria for openness, emphasizing that genuine open-source AI requires meaningful transparency throughout the model-development lifecycle. He concludes that policymakers, researchers, and the public should resist marketing claims that equate partial disclosure with true openness and instead demand standards that enable independent verification, scrutiny, and accountability.

Read the full article here.